WO2021082500A1 - 一种信息推荐的方法、装置、推荐服务器和存储介质 - Google Patents

一种信息推荐的方法、装置、推荐服务器和存储介质 Download PDF

Info

Publication number
WO2021082500A1
WO2021082500A1 PCT/CN2020/099662 CN2020099662W WO2021082500A1 WO 2021082500 A1 WO2021082500 A1 WO 2021082500A1 CN 2020099662 W CN2020099662 W CN 2020099662W WO 2021082500 A1 WO2021082500 A1 WO 2021082500A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
recommendation
machine learning
model parameters
threshold
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/099662
Other languages
English (en)
French (fr)
Inventor
张永池
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Bigo Technology Pte Ltd
Original Assignee
Bigo Technology Pte Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Bigo Technology Pte Ltd filed Critical Bigo Technology Pte Ltd
Priority to US17/772,925 priority Critical patent/US12614108B2/en
Priority to EP20882554.7A priority patent/EP4053716A4/en
Publication of WO2021082500A1 publication Critical patent/WO2021082500A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the embodiments of the present application relate to the field of machine learning technology, such as a method, device, recommendation server, and storage medium for information recommendation.
  • CTR click through rate
  • logistic regression algorithms are usually used to train to obtain a business prediction model that fits the user’s click behavior on recommended information as much as possible.
  • the business prediction model contains model weight coefficients for each feature dimension in the recommended information, and is used to predict the clicks of each recommended information. When passing the rate, it is necessary to query the feature value under each feature dimension in the recommended information and the corresponding model weight coefficient.
  • the embodiments of the present application provide a method, device, recommendation server, and storage medium for information recommendation, which reduce the size of the recommendation model, and improve the efficiency of information recommendation while ensuring the accuracy of recommendation.
  • an information recommendation method including:
  • the feature vector of the recommendation information is predicted by the recommendation model to obtain the click-through rate of the recommendation information, wherein the feature corresponding to the feature that is the same as the feature vector and whose frequency is less than or equal to the first threshold is removed from the recommendation model Model parameters;
  • the recommendation information is issued.
  • an information recommendation device which includes:
  • the click prediction module is set to predict the feature vector of the recommended information through the recommendation model to obtain the click-through rate of the recommended information, wherein the recommendation model removes the feature that is the same as the feature vector and the frequency is less than or equal to Model parameters corresponding to the features of the first threshold;
  • the information recommendation module is configured to issue the recommendation information according to the click-through rate.
  • an embodiment of the present application provides a recommendation server, and the recommendation server includes:
  • At least one processor At least one processor
  • the storage device is set to store at least one program
  • the at least one processor When the at least one program is executed by the at least one processor, the at least one processor implements the information recommendation method described in any embodiment of the present application.
  • an embodiment of the present application provides a model training system, the system includes: a training server cluster and a parameter server cluster, the parameter server cluster stores model parameters corresponding to different features in a preset machine learning model;
  • the training server cluster sequentially obtains the latest model parameters corresponding to different features in the machine learning model from the parameter server cluster, and processes the historical feature vector of the historical recommendation information according to the model parameters to obtain the corresponding click-through rate, and Calculating a loss coefficient of a model parameter corresponding to a feature with a frequency greater than a first threshold in the machine learning model, and returning the loss coefficient to the parameter server cluster;
  • the parameter server cluster updates the corresponding model parameters according to the loss coefficient and feeds them back to the training server cluster.
  • the model parameters corresponding to the features whose frequency is less than or equal to the first threshold are removed to obtain the remaining Model parameters, and use the remaining model parameters to generate a corresponding recommended model.
  • an embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the information described in any embodiment of the present application is realized. The recommended method.
  • FIG. 1A is a flowchart of an information recommendation method provided by an embodiment of this application.
  • FIG. 1B is a schematic diagram of the principle of the information recommendation process provided by an embodiment of this application.
  • 2A is a flowchart of training a recommendation model in a method for information recommendation provided by an embodiment of this application;
  • 2B is a schematic diagram of the principle of the recommendation model training process provided by an embodiment of the application.
  • 3A is a flowchart of training a recommendation model in another information recommendation method provided by an embodiment of this application.
  • FIG. 3B is a schematic diagram of the principle of the training process of the recommendation model in another information recommendation method provided by an embodiment of this application;
  • FIG. 5 is a schematic structural diagram of an information recommendation apparatus provided by an embodiment of this application.
  • FIG. 6 is a schematic structural diagram of a model training system provided by an embodiment of this application.
  • FIG. 7 is a schematic structural diagram of a recommendation server provided by an embodiment of the application.
  • FIG. 1A is a flowchart of a method for recommending information according to an embodiment of this application.
  • This embodiment can be applied to any background recommendation server that can provide users with a service for issuing a request for related recommendation information.
  • the technical solutions of the embodiments of the present application are applicable to the case of recommending relevant information for users.
  • the information recommendation method provided in this embodiment can be executed by the information recommendation device provided in the embodiment of the present application, and the device can be implemented in software and/or hardware, and is integrated in a recommendation server that executes this method.
  • the method may include step S110 to step S120.
  • step S110 the feature vector of the recommendation information is predicted by the recommendation model to obtain the click-through rate of the recommendation information.
  • the push information is that the push server sends a request based on the received user's role in the client, and according to the user's historical behavior data through the itemCF algorithm in related technologies, the user recommendation userCF algorithm, and the content-based algorithm or service recommendation. A large amount of recommended information corresponding to the request issued this time is found by the recommended vector-based algorithm.
  • the issuing request is when the user needs to recommend the relevant information corresponding to the current recommendation by the server, by executing the corresponding content refresh operation or searching for specific information on the client side, the current time is generated correspondingly.
  • the instruction information in the request When the specific request content is not included in the issued request, the historical behavior data of the user can be analyzed to determine the news, information, music, or short video content that the user is more interested in, so as to find the corresponding data content.
  • this request method is generally aimed at the user performing content refresh operation in the corresponding software product installed on the client, and is not sure about the specific information needed. At this time, analyze the corresponding recommendation based on historical behavior data.
  • the specific request content can be directly analyzed to find news, information, and information related to the specific content of this request. For content such as music or short videos, the related information found is used as the recommended information corresponding to this request.
  • This request method is generally aimed at the user performing the search and search operation of related content information in the corresponding software product installed on the client. At this time, the corresponding push information can be determined according to the specific search content in this request, so as to recommend matching information for the user.
  • the feature vector of the recommended information refers to the feature data in different dimensions analyzed according to the various features of the recommended information. For example, if you analyze the characteristics of a video from the three aspects of likes, reposts and comments, then The correspondingly analyzed like results, forwarding results, and comment results from the three aspects can form a three-dimensional feature vector of the video.
  • the recommendation model is a pre-trained network model that is used to quickly filter out the most matching recommendation information for users.
  • the pre-training is a feature vector corresponding to each historical recommendation information. Click behavior for training, training in related technologies will continuously update the model parameters corresponding to each feature in the feature vector preset in the network model corresponding to the recommended information. Therefore, the trained network model has the same model parameters as each feature in the feature vector corresponding to the recommended information. The model parameters corresponding to each feature in the feature vector corresponding to the recommended information, so that each feature in the feature vector can be analyzed.
  • each recommended information contains A large number of different types of features make the trained network model also have a large number of model parameters corresponding to different types of features, resulting in excessive memory usage, and it takes a long time to predict the user’s click behavior on the recommended information. Querying the model parameters corresponding to each feature in the model parameters reduces the efficiency of model calculations.
  • the feature vector of the recommended information may contain a large number of high-dimensional sparse features, such as user ID, host ID, or device ID. At this time, if such high-dimensional sparse features rarely participate in the user's click on the recommended information During the training process of the behavior, it does not affect the user's click behavior on different recommendation information. Therefore, during the model training process, as shown in FIG. 1B, this embodiment can exclude this type of information contained in the feature vector of the historical recommendation information.
  • the feature training only retains the training of the features that have a greater impact on the user's click behavior, so that only the model parameters corresponding to the features that have a greater impact on the user's click behavior are retained in the trained recommendation model.
  • the recommended model trained in this embodiment removes the model parameters corresponding to the features that are the same as the feature vector and whose frequency is less than or equal to the first threshold, and only retains the feature corresponding to the feature with a frequency greater than the first threshold. Therefore, the size of the recommended model is reduced, and the space occupied by the model parameters in the recommended model is reduced.
  • the model parameters in the recommendation model in this embodiment may be weight coefficients corresponding to different features in the recommendation model, and the weight coefficients are used to indicate the degree of influence of each feature on the user's click behavior.
  • the click-through rate of the recommended information refers to the click-through rate that the recommended information is filtered out of all recommended information and recommended to the user's client, and is clicked by the user to view, that is, the actual number of clicks on the recommended information is divided by The total amount of all recommended information is used to determine the pros and cons of each recommended information sent to the user.
  • the user generates a corresponding issuance request by performing a corresponding content refresh operation or a search and search operation for specific information on the client, and sends it to the recommendation server.
  • the recommendation server passes relevant information according to the user’s historical behavior data.
  • the similarity algorithm in the technology finds a large number of recommended information corresponding to the request issued this time, and obtains the corresponding feature vector from the feature server storing different types of features of the uploaded information according to the identification information of each recommended information, and then
  • the feature vector of each recommendation information is input into the pre-trained recommendation model of this embodiment, and the corresponding feature in the feature vector of the recommendation information is performed through the model parameters corresponding to the features retained in the recommendation model with a frequency greater than the first threshold. Analyze to predict the click-through rate of the recommended information.
  • step S120 the recommendation information is issued according to the click-through rate.
  • the user's preference for each recommended information can be clarified, and then the corresponding recommendation can be filtered out of all the recommended information according to the user's request and click-through rate Information and distributed to users.
  • the recommended information with the highest click-through rate is issued to the user.
  • the click-through rate is reliable. The previous pieces of recommendation information are issued to users.
  • the model parameters corresponding to the features that are the same as the feature in the feature vector of the recommendation information and whose frequency is less than or equal to the first threshold are removed, and only those whose frequency is greater than the first threshold are retained.
  • the model parameters corresponding to the features thereby reducing the size of the recommendation model and reducing the footprint of the model parameters in the recommendation model.
  • the model parameter corresponding to each feature in the vector reduces the time spent in the query process. Only the features in the feature vector of the recommended information corresponding to the model parameters retained in the recommended model are analyzed, while ensuring the accuracy of the recommendation, Improve the efficiency of information recommendation.
  • FIG. 2A is a flowchart of training a recommendation model in a method for information recommendation provided by an embodiment of this application
  • FIG. 2B is a schematic diagram of a training process of a recommendation model provided by an embodiment of this application.
  • This embodiment is refined on the basis of the foregoing embodiment.
  • the training of the recommendation model in this embodiment can be performed offline, and the training of the offline recommendation model and the online step of using the trained recommendation model to predict the click-through rate of each recommendation information can be implemented on different devices, so
  • the recommendation model can be trained through a pre-architected model training system. After the training is completed, the recommendation model is published to the recommendation server for online execution of the information recommendation method mentioned in any embodiment of this application .
  • the offline training process of the recommended model is mainly explained in detail.
  • this embodiment may include step S210 to step S230.
  • step S210 according to the click-through rate of the historical feature vector of the historical recommendation information in the preset machine learning model, the model parameters corresponding to the features in the machine learning model whose frequency is greater than the first threshold are sequentially updated.
  • a machine learning model Before training the recommendation model, a machine learning model will be constructed in advance, in which the initial value of the corresponding model parameter is preset for each feature under different types, and a large number of historical recommendation information under different types will be subsequently passed
  • the click-through rate prediction error of the machine learning model is continuously corrected until the click-through rate of any type of historical recommendation information can be accurately predicted.
  • the current model parameters preset in the machine learning model analyze the corresponding features in the historical feature vector, and then obtain the click-through rate of each historical recommendation information, through the historical label and click-through of the historical recommendation information Rate comparison to determine the prediction error under the current model parameters in the machine learning model, and at the same time, determine the number of times each feature in the machine learning model appears based on the features contained in the feature vector of the historical recommendation information, that is, in this embodiment
  • the frequency of the feature, and the prediction error is backpropagated in the machine learning model to correct the model parameters corresponding to the feature whose frequency is greater than the first threshold in the selected machine learning model, so as to subsequently improve the accuracy of the prediction result .
  • Model parameters can include: predicting the historical feature vector of the current historical recommendation information through a machine learning model to obtain the corresponding current click pass rate; using a preset loss function to calculate the current click pass rate corresponding to the feature whose frequency is greater than the first threshold.
  • the loss coefficient under the model parameters, and correct the corresponding model parameters in the machine learning model according to the loss coefficient continue to predict the historical feature vector of the next historical recommendation information through the machine learning model, and modify the frequency of the machine learning model to be greater than the first
  • the model parameters corresponding to the threshold feature until the machine learning model completes the training.
  • the historical feature vector of the current historical recommendation information is input into a preset machine learning model, and the corresponding feature in the historical feature vector of the current historical recommendation information is analyzed through the current model parameters, and the current historical recommendation information is predicted.
  • the current click-through rate of the historical recommendation information analyze the frequency of each feature in the machine learning model based on the features contained in the historical feature vector of the historical recommendation information, and then mark the model parameters corresponding to the features whose frequency is greater than the first threshold; at the same time, use the preset
  • the loss function determines the loss coefficient of the current click-through rate under the model parameter corresponding to the feature whose frequency is greater than the first threshold in the machine learning model, that is, the distance between the model parameter and the click-through rate that can accurately analyze the historical recommendation information.
  • the corresponding model parameters in the machine learning model are revised according to the loss coefficient; after this revision, the historical feature vector of the next historical recommendation information is continued to be obtained, and input into the revised machine learning model, which is used in the machine learning model.
  • the revised model parameters (the model parameters corresponding to the features with a frequency greater than the first threshold) continue to analyze the features corresponding to the model parameters in the historical feature vector of the next historical recommendation information, and exclude features with a frequency less than or equal to the first threshold, and get The next click-through rate of the next historical recommendation information, and continue to modify the model parameters corresponding to the features of the machine learning model with a frequency greater than the first threshold, and loop in turn until the prediction error of the click-through rate of the historical recommendation information is lower than the preset
  • the error threshold indicates that the machine learning model has completed training at this time.
  • step S220 after the machine learning model is trained, the model parameters corresponding to the features whose frequency is less than or equal to the first threshold are removed to obtain the remaining model parameters, and the remaining model parameters are used to generate a recommended model.
  • the remaining model parameters are model parameters corresponding to features in the machine learning model with a frequency greater than the first threshold.
  • the machine learning model since the features with a frequency less than or equal to the first threshold in the machine learning model hardly contribute to predicting the user's click behavior, for the efficiency of the recommendation model, in this embodiment, after the machine learning model is trained, it can be directly According to the labeling of the model parameters corresponding to the features with a frequency greater than the first threshold during the model training process, the model parameters corresponding to the features with a frequency less than or equal to the first threshold in the machine learning model are removed, and the click-through rate of the recommended information is not determined.
  • the prediction result has an impact, and the remaining model parameters (model parameters corresponding to the features with a frequency greater than the first threshold in the machine learning model) after removing the model parameters corresponding to the features with a frequency less than or equal to the first threshold are subsequently used to generate the corresponding recommendation model, so that In the recommendation model, only the model parameters corresponding to the features whose frequency is greater than the first threshold are retained. On the basis of ensuring the accuracy of the recommendation, the size of the recommendation model is reduced, and the efficiency of the recommendation model is improved.
  • the model parameters corresponding to the features in the machine learning model with a frequency less than or equal to the first threshold after sequentially updating the model parameters corresponding to the features in the machine learning model with a frequency greater than the first threshold, it may also include : Set the model parameters corresponding to the features whose frequency is less than or equal to the first threshold in the machine learning model to the preset sparse value, so that after the machine learning model is trained, the preset sparse value is removed to obtain the remaining model parameters.
  • the preset in the machine learning model can be determined according to the features contained in the historical feature vector The frequency of each feature of, at this time, the model parameter corresponding to the feature whose frequency is greater than the first threshold is corrected.
  • the model parameter corresponding to the feature whose frequency is less than or equal to the first threshold will no longer participate in the subsequent Learning and training, therefore directly set the model parameters corresponding to the features whose frequency is less than or equal to the first threshold to the preset sparse value (the preset sparse value is 0 in this embodiment), and then directly remove it after the machine learning model is trained
  • the model parameters are set as the preset sparse values, and the model parameters corresponding to the features whose frequency is greater than the first threshold are retained.
  • an improved follow the regularized logic (Followed the Regularized Leader, FTRL) algorithm is configured in the machine learning model, and the improved FTRL algorithm is used to sequentially match the frequency greater than the first threshold during the model training process.
  • the model parameters corresponding to the features are corrected.
  • the FTRL algorithm in related technologies is as follows:
  • g is the loss coefficient of the model parameter corresponding to a certain feature
  • z and n are the intermediate variables of the machine learning model in the training process
  • w is the model parameter corresponding to the feature
  • ⁇ , ⁇ , ⁇ 1 and ⁇ 2 are sparsity Qualified parameters for training.
  • the frequency of features in the machine learning model is limited, and the improved FTRL algorithm is as follows:
  • f is the frequency of a feature in the machine learning model
  • ⁇ 3 is the first threshold defined by the frequency of the feature
  • the improved FTRL algorithm still performs conventional model parameters corresponding to features in the machine learning model with a frequency greater than the first threshold. Training, and directly set the model parameter corresponding to the feature whose frequency is less than or equal to the first threshold to 0, not participating in the subsequent machine learning process, that is, in the training process of the machine learning model, according to the historical feature vector of the historical recommendation information , Determine the frequency of each feature preset in the machine learning model.
  • the model parameters corresponding to the feature Only when the feature frequency is greater than the first threshold, will the model parameters corresponding to the feature be machine-learned, otherwise the model parameters corresponding to the feature will always be set to 0; After the machine learning model is trained, the model parameters that are not 0 are used to generate the corresponding recommendation model. On the basis of ensuring the accuracy of the recommendation, the size of the recommendation model is reduced, and the efficiency of the recommendation model is improved.
  • step S230 the recommendation model is published online.
  • the recommendation model can be directly published online, that is, applied to a recommendation server that executes the information recommendation method provided by any embodiment of this application, so that the recommendation can be directly adopted online
  • the model predicts the click-through rate of each recommended information, recommends more interesting content for users, and improves the accuracy and efficiency of information recommendation.
  • the click-through rate predicted by the machine learning model on the historical recommendation information continuously updates the model parameters corresponding to the features in the machine learning model with a frequency greater than the first threshold, and after the machine learning model is trained, the frequency is removed.
  • the model parameters corresponding to the features less than or equal to the first threshold are used, and the remaining model parameters are used to generate the corresponding recommended model.
  • FIG. 3A is a flowchart of training a recommendation model in another information recommendation method provided by an embodiment of this application
  • FIG. 3B is a principle of the training process of a recommendation model in another information recommendation method provided by an embodiment of this application Schematic.
  • This embodiment is refined on the basis of the foregoing embodiment. After the machine learning model is trained, if the model scale reaches the upper limit of the memory, it will cause congestion in the model prediction. Therefore, the model parameters corresponding to the features with lower frequency need to be filtered again to ensure the efficiency of the recommended model.
  • This embodiment mainly explains in detail the secondary optimization process of the recommendation model.
  • the method may include step S310 to step S340.
  • step S310 according to the click-through rate of the historical feature vector of the historical recommendation information in the preset machine learning model, the model parameters corresponding to the features in the machine learning model whose frequency is greater than the first threshold are sequentially updated.
  • step S320 after the machine learning model is trained, the model parameters corresponding to the features whose frequency is less than or equal to the first threshold are removed to obtain the remaining model parameters.
  • the remaining model parameters Filter the model parameters corresponding to the features whose frequency is less than or equal to the second threshold, and use the filtered remaining model parameters to generate a recommended model.
  • the model parameters corresponding to the features of the machine learning model whose frequency is less than or equal to the first threshold can be removed.
  • the size of the machine learning model still reaches the upper limit of memory
  • After removing the model parameters corresponding to the features whose frequency is less than or equal to the first threshold continue to filter out the model parameters corresponding to the features whose frequency is less than or equal to the second threshold.
  • the second threshold is greater than the first threshold, release Corresponding memory space, and use the filtered remaining model parameters to generate a corresponding recommendation model, thereby reducing the size of the recommendation model.
  • step S330 after the machine learning model completes the training in the current period, the frequency of different features in the machine learning model is updated according to the preset frequency attenuation coefficient, and the model parameters corresponding to the features whose frequency is less than or equal to the first threshold are removed Among the remaining model parameters, the model parameters corresponding to the features whose filtering frequency is less than or equal to the third threshold are used to generate the recommended model in the current period of time using the filtered remaining model parameters.
  • this embodiment can train the machine learning model in time periods, obtain historical recommendation information in the current time period, and continuously modify the model parameters corresponding to the features of the machine learning model with a frequency greater than the first threshold, and complete the machine learning model After training in the current period, since a feature in the previous period may no longer appear in the current period, the frequency of different features in the machine learning model is updated according to the preset frequency attenuation coefficient.
  • the model parameters corresponding to features with a frequency less than or equal to the first threshold are first removed to obtain the remaining model Parameters, and then update the frequency of different features in the machine learning model according to the preset frequency attenuation coefficient, and then filter the model parameters corresponding to the features whose frequency is less than or equal to the third threshold in the remaining model parameters, and then use the filtered remaining model parameters Generate the recommendation model under the current period to improve the accuracy and efficiency of the recommendation model.
  • model parameters corresponding to the features whose filtering frequency is less than or equal to the third threshold in this embodiment can be determined according to the scale of the machine learning model after training. After the machine learning model completes the training scale in the current period If the upper memory limit is not reached, you do not need to filter again the model parameters corresponding to the features whose frequency is less than or equal to the third threshold, but directly use the remaining model parameters after removing the model parameters corresponding to the features whose frequency is less than or equal to the first threshold to generate the corresponding Recommended model. In this embodiment, S320 and S330 can choose to execute one of them.
  • step S340 the recommendation model is published online.
  • FIG. 4 is a flowchart of yet another information recommendation method provided by an embodiment of this application. This embodiment is refined on the basis of the foregoing embodiment. The specific information recommendation process is described in this embodiment. As shown in FIG. 4, this embodiment may include step S410 to step S430.
  • step S410 obtain the remaining model parameters in the recommendation model after removing the model parameters corresponding to the features that are the same as the feature vector and whose frequency is less than or equal to the first threshold.
  • the online recommendation model after obtaining a large amount of recommendation information according to user requests, it is necessary to predict the feature vector of each recommendation information through the online recommendation model. Therefore, first determine the online recommendation model and find the recommendation model The remaining model parameters after removing the model parameters corresponding to the features that are the same as the feature vector and the frequency is less than or equal to the first threshold, that is, the model parameters corresponding to the features with a frequency greater than the first threshold, in order to compare the features of each recommendation information The corresponding features in the vector are analyzed.
  • step S420 the click-through rate of the recommendation information is calculated according to the remaining model parameters and the features corresponding to the remaining model parameters in the feature vector.
  • the feature corresponding to the remaining model parameter is selected from the feature vector of each recommendation information.
  • the model parameter may be the weight coefficient of the feature, and a weighted summation method is adopted. Multiply each remaining model parameter and the feature vector of the recommended information with the feature corresponding to the remaining model parameter and sum them. The result of the weighted summation of the feature corresponding to each remaining model parameter in the feature vector of the recommended information is taken as the original.
  • the click-through rate of recommended information in the embodiment reduces the amount of calculation in the recommendation model and improves the efficiency of information recommendation.
  • step S430 the recommended information is sorted according to the click-through rate, and a corresponding delivery list is constructed and delivered to the user.
  • the model parameters corresponding to the features that are the same as the feature in the feature vector of the recommendation information and whose frequency is less than or equal to the first threshold are removed, and only those whose frequency is greater than the first threshold are retained.
  • the model parameters corresponding to the features thereby reducing the size of the recommendation model, and reducing the space occupied by the parameters in the recommendation model.
  • model parameters corresponding to each feature of reduce the time spent in the query process, and only analyze the features in the feature vector of the recommended information that correspond to the model parameters retained in the recommended model, which improves the accuracy of the recommendation while ensuring the accuracy of the recommendation.
  • the efficiency of information recommendation which in turn pushes more high-quality content to users, improves the effect of information push.
  • FIG. 5 is a schematic structural diagram of an information recommendation apparatus provided by an embodiment of this application. As shown in FIG. 5, the apparatus may include:
  • the click prediction module 510 is configured to predict the feature vector of the recommendation information through the recommendation model to obtain the click pass rate of the recommendation information, wherein the recommendation model removes the features that are the same as the feature vector and the frequency is less than or equal to the first threshold Model parameters corresponding to features;
  • the information recommendation module 520 is configured to issue recommendation information according to the click-through rate.
  • the model parameters corresponding to the features that are the same as the feature in the feature vector of the recommendation information and whose frequency is less than or equal to the first threshold are removed, and only those whose frequency is greater than the first threshold are retained.
  • the model parameters corresponding to the features thereby reducing the size of the recommendation model, and reducing the space occupied by the parameters in the recommendation model.
  • the above-mentioned click prediction module 510 may be configured as:
  • the click-through rate of the recommended information is calculated.
  • the aforementioned recommendation model may be generated by performing the following operations on the model parameter update module and the recommendation model generation module included in the model training system:
  • the model parameter update module is configured to sequentially update the model parameters corresponding to the features in the machine learning model whose frequency is greater than the first threshold according to the click-through rate of the historical feature vector of the historical recommendation information in the preset machine learning model;
  • the recommended model generation module is set to remove the model parameters corresponding to the features whose frequency is less than or equal to the first threshold after the machine learning model is trained to obtain the remaining model parameters, and use the latest remaining model parameters to generate the recommended model, the remaining model parameters It is the model parameter corresponding to the feature whose frequency is greater than the first threshold in the machine learning model.
  • the above model parameter update module may be set as:
  • the preset loss function is used to calculate the loss coefficient of the current click-through rate under the model parameter corresponding to the feature whose frequency is greater than the first threshold, and the corresponding model parameter in the machine learning model is corrected according to the loss coefficient, and the machine learning model continues to compare the next one
  • the historical feature vector of the historical recommendation information is predicted, and the model parameters corresponding to the features with a frequency greater than the first threshold in the machine learning model are corrected until the machine learning model is trained.
  • the aforementioned model training system may further include:
  • the sparse parameter setting module is set to set the model parameters corresponding to the features whose frequency is less than or equal to the first threshold in the machine learning model to the preset sparse value, so that after the machine learning model is trained, the model set to the preset sparse value is removed parameter.
  • the aforementioned recommendation model generation module may be set to:
  • the aforementioned recommendation model generation module may also be set to:
  • the frequency of different features in the machine learning model is updated according to the preset frequency attenuation coefficient, and the remaining model parameters after removing the model parameters corresponding to the features whose frequency is less than or equal to the first threshold
  • the model parameter corresponding to the feature whose frequency is less than or equal to the second threshold is filtered, and the remaining model parameter after filtering is used to generate the recommended model in the current period.
  • the aforementioned model training system may further include:
  • the model publishing module is set to publish recommended models online.
  • the aforementioned model parameters are weight coefficients of different features in the recommended model.
  • the above-mentioned information recommendation module 520 may be configured as:
  • the recommendation information is sorted according to the click-through rate, and the corresponding delivery list is constructed and delivered to the user.
  • the information recommendation apparatus provided in this embodiment may be applicable to the information recommendation method provided in any of the foregoing embodiments.
  • FIG. 6 is a schematic structural diagram of a model training system provided by an embodiment of this application. This embodiment is refined on the basis of the foregoing embodiment.
  • the model training system may include: a training server cluster 610 and a parameter server cluster 620.
  • the parameter server cluster 620 stores model parameters corresponding to different features in a preset machine learning model.
  • the training server cluster 610 sequentially obtains the latest model parameters corresponding to different features in the machine learning model from the parameter server cluster 620, and processes the historical feature vector of the historical recommendation information according to the model parameters to obtain the corresponding click-through rate, and The loss coefficient of the model parameter corresponding to the feature whose frequency is greater than the first threshold in the computing machine learning model is returned to the parameter server cluster 620; the parameter server cluster 620 updates the corresponding model parameter according to the loss coefficient and feeds it back to the training server cluster 610: After the machine learning model is trained, the model parameters corresponding to the features whose frequency is less than or equal to the first threshold are removed to obtain the remaining model parameters, and the remaining model parameters are used to generate a corresponding recommendation model.
  • the training server cluster 610 sets the model parameters corresponding to the features whose frequency is less than or equal to the first threshold in the machine learning model to the preset sparse value, and returns to the parameter server cluster 620; the parameter server cluster 620 sets the frequency less than or equal to the first threshold.
  • the model parameter corresponding to the threshold feature is corrected to a preset sparse value, and after the machine learning model is trained, the model parameter set to the preset sparse value is removed.
  • the parameter server cluster 620 in order to ensure the efficiency of the model, the parameter server cluster 620, after the machine learning model is trained, if the size of the machine learning model reaches the upper limit of the memory, the filtering frequency in the remaining model parameters is less than or equal to the first The model parameters corresponding to the features of the two thresholds are used, and the filtered remaining model parameters are used to generate a recommended model, and the second threshold is greater than the first threshold.
  • the parameter server cluster 620 updates the frequency of different features in the machine learning model according to the preset frequency attenuation coefficient after the machine learning model completes the training in the current period. And filter the model parameters corresponding to the features whose frequency is less than or equal to the third threshold from the remaining model parameters after removing the model parameters corresponding to the features whose frequency is less than or equal to the first threshold, and use the filtered remaining model parameters to generate recommendations in the current period model.
  • model training system may also include a recommendation server 630.
  • the parameter server cluster 620 publishes the generated recommendation model to the recommendation server 630, and the recommendation server 630 is configured to execute the information recommendation method provided by any embodiment of the present application.
  • multiple training servers are distributed in the training server cluster 610, and the training server can execute the training of the machine learning model in parallel to increase the training rate of the model; at the same time, multiple parameter servers are distributed in the parameter server cluster 620.
  • the model parameters corresponding to different features in the machine learning model are stored in fragments to reduce the storage pressure of data; and the training steps and parameter correction steps in the model training process are separated, which reduces the computational pressure of model training.
  • the recommendation model is jointly trained by the training server cluster and the parameter server cluster, thereby separating the training step and the parameter update step, and dispersing the training and model parameters to different computing nodes, reducing a single The computing pressure of the server increases the training rate of the recommended model.
  • FIG. 7 is a schematic structural diagram of a recommendation server provided by an embodiment of the application.
  • the recommendation server includes a processor 70, a storage device 71, and a communication device 72; the number of processors 70 in the recommendation server may be At least one, one processor 70 is taken as an example in FIG. 7; it is recommended that the processor 70, the storage device 71, and the communication device 72 in the server can be connected by a bus or other means. In FIG. 7, the connection by a bus is taken as an example.
  • the storage device 71 can be configured to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the information recommendation method described in any embodiment of the present application.
  • the processor 70 executes various functional applications and data processing of the recommendation server by running the software programs, instructions, and modules stored in the storage device 71, that is, realizes the above-mentioned information recommendation method.
  • the storage device 71 may mainly include a storage program area and a storage data area.
  • the storage program area may store an operating system and an application program required by at least one function; the storage data area may store data created according to the use of the terminal, and the like.
  • the storage device 71 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices.
  • the storage device 71 may include a memory remotely provided with respect to the processor 70, and these remote memories may be connected to the recommendation server through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
  • the communication device 72 may be configured to implement a network connection or a mobile data connection between the recommendation server and the client.
  • the recommendation server provided in this embodiment can be used to execute the information recommendation method provided in any of the foregoing embodiments.
  • An embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program.
  • the computer program When executed by a processor, it can implement the information recommendation method in any of the foregoing embodiments.
  • the method can include:
  • the recommendation model Predict the feature vector of the recommendation information through the recommendation model to obtain the click-through rate of the recommendation information, wherein the recommendation model removes model parameters corresponding to features that are the same as the feature vector and whose frequency is less than or equal to the first threshold;
  • the recommendation information is issued.
  • a storage medium containing computer-executable instructions provided by an embodiment of the present application is not limited to the method operations described above, and can also execute the information recommendation method provided by any embodiment of the present application. Related operations.
  • the technical solution of the present application essentially or the part that contributes to the related technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, Read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including multiple instructions to make a computer device (which can be a personal computer) , A server, or a network device, etc.) execute the method described in each embodiment of the present application.
  • a computer-readable storage medium such as a computer floppy disk, Read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种信息推荐的方法、装置、推荐服务器和存储介质。其中，该方法包括：通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，其中，所述推荐模型中去掉了与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；根据所述点击通过率，下发所述推荐信息。

Description

一种信息推荐的方法、装置、推荐服务器和存储介质
本申请要求在2019年10月31日提交中国专利局、申请号为201911053819.3的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及机器学习技术领域，例如一种信息推荐的方法、装置、推荐服务器和存储介质。
背景技术
随着人工智能交互技术的快速发展，预先训练的机器学习模型越来越广泛的应用在推荐、搜索、广告等各类业务的大数据场景下，针对用户的搜索需求，后台会获取大量相关的推荐信息，此时需要预测多条推荐信息的点击通过率(Click Through Rate，CTR)，以便根据多个推荐信息的点击通过率为用户推送最优的信息。
目前通常采用逻辑回归算法训练得到尽量拟合用户对推荐信息的点击行为的业务预测模型，该业务预测模型中包含推荐信息中每一特征维度下的模型权重系数，在预测每一推荐信息的点击通过率时，需要去查询推荐信息中每一特征维度下的特征值以及对应的模型权重系数，此时由于推荐信息中存在大量高维稀疏特征，如推荐信息中身份维度下的用户身份标识(identity，ID)、主播ID或者设备ID等，最高能够达到上亿维，使得业务预测模型在存储推荐信息中的大量高维稀疏特征时会占用大量的内存空间，导致内存不足的情况；而且对于包含大量特征的推荐信息，需要耗费较长时间去查询推荐信息中每一特征维度下的模型权重系数，这使得业务预测模型任务量过大，导致推荐信息的预测效率较低。
发明内容
本申请实施例提供了一种信息推荐的方法、装置、推荐服务器和存储介质，减小了推荐模型的大小，在保证推荐准确性的同时，提高信息推荐效率。
第一方面，本申请实施例提供了一种信息推荐的方法，包括：
通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，其中，所述推荐模型中去掉了与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
根据所述点击通过率，下发所述推荐信息。
第二方面，本申请实施例提供了一种信息推荐的装置，该装置包括：
点击预测模块，设置为通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，其中，所述推荐模型中去掉了与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
信息推荐模块，设置为根据所述点击通过率，下发所述推荐信息。
第三方面，本申请实施例提供了一种推荐服务器，该推荐服务器包括：
至少一个处理器；
存储装置，设置为存储至少一个程序；
当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现本申请任意实施例所述的信息推荐的方法。
第四方面，本申请实施例提供了一种模型训练系统，该系统包括：训练服务器集群和参数服务器集群，所述参数服务器集群中存储预设的机器学习模型中不同特征对应的模型参数；
所述训练服务器集群依次从所述参数服务器集群获取所述机器学习模型中不同特征对应的最新模型参数，根据该模型参数对历史推荐信息的历史特征向量进行处理，得到对应的点击通过率，并计算所述机器学习模型中频次大于第一阈值的特征对应的模型参数的损失系数，将所述损失系数返回给所述参数服务器集群；
所述参数服务器集群根据该损失系数更新对应的模型参数并反馈给所述训 练服务器集群，直至所述机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用所述剩余模型参数生成对应的推荐模型。
第五方面，本申请实施例提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时，实现本申请任意实施例所述的信息推荐的方法。
附图说明
图1A为本申请一实施例提供的一种信息推荐的方法的流程图；
图1B为本申请一实施例提供的信息推荐过程的原理示意图；
图2A为本申请一实施例提供的信息推荐的方法中训练推荐模型的流程图；
图2B为本申请一实施例提供的推荐模型训练过程的原理示意图；
图3A为本申请一实施例提供的另一种信息推荐的方法中训练推荐模型的流程图；
图3B为本申请一实施例提供的另一种信息推荐的方法中推荐模型的训练过程的原理示意图；
图4为本申请一实施例提供的又一种信息推荐的方法的流程图；
图5为本申请一实施例提供的一种信息推荐的装置的结构示意图；
图6为本申请一实施例提供的一种模型训练系统的原理架构图；
图7为本申请一实施例提供的一种推荐服务器的结构示意图。
具体实施方式
图1A为本申请一实施例提供的一种信息推荐的方法的流程图，本实施例可应用于任一种可以为用户提供针对相关推荐信息的下发请求服务的后台推荐服务器中。本申请实施例的技术方案适用于为用户推荐相关信息的情况中。本实施例提供的一种信息推荐的方法可以由本申请实施例提供的信息推荐的装置来 执行，该装置可以通过软件和/或硬件的方式来实现，并集成在执行本方法的推荐服务器中。
参考图1A，该方法可以包括步骤S110至步骤S120。
在步骤S110中，通过推荐模型对推荐信息的特征向量进行预测，得到推荐信息的点击通过率。
其中，推送信息是推送服务器根据接收到的用户在客户端作用的下发请求，根据用户的历史行为数据通过相关技术中的物品推荐itemCF算法、用户推荐userCF算法、内容推荐content-based算法或服务推荐vector-based算法等查找出的大量与本次下发请求对应的推荐信息。
示例性的，该下发请求为用户在需要推荐服务端在当前推荐对应的相关信息时，通过在客户端执行对应的内容刷新操作或者对具体某一信息的搜索查找操作，而对应生成本次请求中的指令信息。在下发请求中不包含具体的请求内容时，可以通过对该用户的历史行为数据进行分析，判断该用户较为感兴趣的新闻、资讯、音乐或短视频等内容，从而查找出对应的数据内容，作为本次请求对应的推荐信息，该种请求方法一般针对的是用户在客户端安装的对应软件产品中执行内容刷新操作，不确定所需要的具体信息，此时根据历史行为数据分析对应的推荐信息，从而为用户推荐优质的信息；同时，在下发请求中包含具体的请求内容时，可以直接对该具体的请求内容进行分析，从而查找与本次请求的具体内容相关联的新闻、资讯、音乐或短视频等内容，将查找到的关联信息作为本次请求对应的推荐信息，该种请求方法一般针对的是用户在客户端安装的对应软件产品中执行相关内容信息的搜索查找操作，此时可以根据本次请求中的具体搜索内容确定对应的推送信息，从而为用户推荐匹配的信息。
同时，推荐信息的特征向量是指根据推荐信息的各类特征分析出的在不同维度上的特征数据，例如，如果从点赞、转发和评论三方面对某一视频进行特征分析，则从这三方面分别对应分析出的点赞结果、转发结果和评论结果可以 组成该视频的一个三维特征向量。
此外，推荐模型为预先训练好的一种用于快速为用户筛选出最匹配的推荐信息的网络模型，该预先训练为针对每个历史推荐信息对应的特征向量，对用户每一个历史推荐信息的点击行为进行的训练，相关技术中的训练会不断更新网络模型中预设的与推荐信息对应的特征向量中每一特征相同的特征所对应的模型参数，因此所训练出的网络模型中存在与推荐信息对应的特征向量中每一特征对应的模型参数，以便对特征向量中的每个特征进行分析，此时由于在搜索、个性化推送或者广告通知等推荐场景下，每一推荐信息中包含大量不同类型下的特征，使得训练后的网络模型中也存在大量不同类型特征对应的模型参数，导致内存占用过大，而且在预测用户对推荐信息的点击行为时，需要耗费较长时间在大量模型参数中查询每一特征对应的模型参数，降低了模型运算的效率。
本实施例中考虑到推荐信息的特征向量中可能包含大量高维稀疏特征，如用户ID、主播ID或者设备ID等，此时如果该类高维稀疏特征极少参与到用户对推荐信息的点击行为的训练过程中，从而对用户对不同推荐信息的点击行为不造成影响，因此在模型训练过程中，如图1B所示，本实施例可以排除对历史推荐信息的特征向量中包含的该类特征的训练，仅保留对用户的点击行为存在较大影响的特征的训练，使得训练后的推荐模型中仅保留对用户的点击行为存在较大影响的特征对应的模型参数。本实施例中可以在模型训练过程中根据历史推荐信息的特征向量，判断网络模型中预设的每一特征出现的次数(本实施例中特征的频次)，该出现的次数表示该特征对用户的点击行为的影响程度，因此本实施例训练好的推荐模型中去掉了与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数，仅保留频次大于第一阈值的特征对应的模型参数，从而减少了推荐模型的大小，降低了推荐模型中的模型参数占用空间。示例性的，本实施例中推荐模型中的模型参数可以为该推荐模型中不同特征对应的权重系数，该权重系数用于表示每个特征对于用户的点击行为的影响 程度。
同时，推荐信息的点击通过率是指将该推荐信息从全部的推荐信息中筛选出来并推荐给用户所在客户端，并被用户点击查看的点击达到率，也就是该推荐信息的实际点击次数除以全部推荐信息的信息总量，以此确定下发给用户的每个推荐信息的优劣性。
示例性的，用户通过在客户端执行对应的内容刷新操作或者对具体某一信息的搜索查找操作，生成对应的下发请求，并发送给推荐服务器，该推荐服务器根据用户的历史行为数据通过相关技术中的相似度算法查找出大量与本次下发请求对应的推荐信息，并根据每个推荐信息的标识信息从存储有上传信息的不同类型特征的特征服务端中获取对应的特征向量，进而将每个推荐信息的特征向量输入到本实施例预先训练好的推荐模型中，通过该推荐模型中保留的频次大于第一阈值的特征对应的模型参数对推荐信息的特征向量中对应的特征进行分析，从而预测得到推荐信息的点击通过率。
在步骤S120中，根据点击通过率，下发推荐信息。
示例性的，在得到每个推荐信息的点击通过率后，可以明确用户对每个推荐信息的喜好程度，进而根据用户的下发请求和点击通过率在全部的推荐信息中筛选出对应的推荐信息，并下发给用户。示例性的，在用户当前需要获取最优的一条信息的情况下，将点击通过率最高的推荐信息下发给用户，在用户需要一次性获取多条相关信息的情况下，将点击通过率靠前的多条推荐信息下发给用户。
本实施例提供的技术方案，在推荐模型的训练过程中，去掉与推荐信息的特征向量中的特征相同且频次小于或等于第一阈值的特征对应的模型参数，仅保留频次大于第一阈值的特征对应的模型参数，从而减小了推荐模型的大小，降低了推荐模型中的模型参数的占用空间，同时在通过该推荐模型中的模型参数对推荐信息的特征向量进行预测时，无需查询特征向量中的每一特征对应的 模型参数，降低了查询过程中的耗费时长，仅对推荐信息的特征向量中与推荐模型中保留的模型参数对应的特征进行分析，在保证推荐准确性的同时，提高了信息推荐效率。
图2A为本申请一实施例提供的信息推荐的方法中训练推荐模型的流程图，图2B为本申请一实施例提供的推荐模型训练过程的原理示意图。本实施例是在上述实施例的基础上进行细化。本实施例中推荐模型的训练可以在线下执行，线下的推荐模型的训练与线上采用该训练好的推荐模型预测每个推荐信息的点击通过率的步骤可以在不同的设备上实现，因此本实施例中可以通过预先架构的模型训练系统来训练该推荐模型，在训练完成后，将该推荐模型发布到推荐服务器中，用于线上执行本申请任意实施例提及的信息推荐的方法。本实施例中主要对于推荐模型的线下训练过程进行详细的解释说明。
如图2A所示，本实施例中可以包括步骤S210至步骤S230。
在步骤S210中，根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新机器学习模型中频次大于第一阈值的特征对应的模型参数。
其中，在训练推荐模型前，会预先构建一个机器学习模型，该机器学习模型中针对不同类型下的每一特征均预先设置对应的模型参数的初始值，后续通过大量不同类型下的历史推荐信息的点击通过率的预测误差，不断对该机器学习模型中每个特征对应的模型参数进行修正，直至能够准确预测任一类型下的历史推荐信息的点击通过率。
示例性的，获取大量不同类型下的历史推荐信息，用于准确分析用户对不同类型信息的点击行为，进而判断用户的喜好；将多个历史推荐信息的历史特征向量输入到预设的机器学习模型中，由该机器学习模型中预设的当前模型参数对历史特征向量中对应的特征进行分析，进而得到每个历史推荐信息的点击通过率，通过对该历史推荐信息的历史标签和点击通过率进行比对，判断机器 学习模型中当前模型参数下的预测误差，同时根据历史推荐信息的特征向量中包含的特征，判断该机器学习模型中每个特征出现的次数，也就是本实施例中特征的频次，并将该预测误差在机器学习模型中反向传播，用于对选取出的机器学习模型中频次大于第一阈值的特征对应的模型参数进行修正，以便后续提高预测结果的准确性。
示例性的，如图2B所示，本实施例中根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新机器学习模型中频次大于第一阈值的特征对应的模型参数，可以包括：通过机器学习模型对当前历史推荐信息的历史特征向量进行预测，得到对应的当前点击通过率；采用预设损失函数计算当前点击通过率在频次大于第一阈值的特征对应的模型参数下的损失系数，并根据该损失系数修正机器学习模型中对应的模型参数，继续通过机器学习模型对下一历史推荐信息的历史特征向量进行预测，并修正机器学习模型中频次大于第一阈值的特征对应的模型参数，直至机器学习模型完成训练。
示例性的，将当前历史推荐信息的历史特征向量输入预设的机器学习模型中，通过当前的模型参数对当前历史推荐信息的历史特征向量中的对应特征进行分析，预测得到该当前历史推荐信息的当前点击通过率，并根据历史推荐信息的历史特征向量中包含的特征，分析机器学习模型中每个特征的频次，进而标记出频次大于第一阈值的特征对应的模型参数；同时采用预设损失函数确定当前点击通过率在该机器学习模型中频次大于第一阈值的特征对应的模型参数下的损失系数，也就是该模型参数距离达到能够准确分析出历史推荐信息的点击通过率的差距，进而根据该损失系数分别修正机器学习模型中对应的模型参数；在本次修正后继续获取下一历史推荐信息的历史特征向量，并输入到修正后的机器学习模型中，由该机器学习模型中修正后的模型参数(频次大于第一阈值的特征对应的模型参数)继续分析下一个历史推荐信息的历史特征向量中与该模型参数对应的特征，排除频次小于或等于第一阈值的特征，得到下一个历史推荐信息的下一点击通过率，并继续修正该机器学习模型中频次大于第一 阈值的特征对应的模型参数，依次循环，直至历史推荐信息的点击通过率的预测误差低于预设误差阈值，表示此时该机器学习模型完成训练。
在步骤S220中，在机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用剩余模型参数生成推荐模型。
其中，剩余模型参数为机器学习模型中频次大于第一阈值的特征对应的模型参数。
示例性的，由于机器学习模型中频次小于或等于第一阈值的特征对于预测用户的点击行为几乎未作出贡献，因此为了推荐模型的效率，本实施例中在机器学习模型完成训练后，可以直接根据模型训练过程中对频次大于第一阈值的特征对应的模型参数的标记情况，去掉该机器学习模型中频次小于或等于第一阈值的特征对应的模型参数，而不对推荐信息的点击通过率的预测结果造成影响，后续利用去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数(机器学习模型中频次大于第一阈值的特征对应的模型参数)生成对应的推荐模型，使得该推荐模型中仅保留频次大于第一阈值的特征对应的模型参数，在保证推荐准确性的基础上，减小了推荐模型的大小，提高了推荐模型的效率。
此外，为了准确判断出机器学习模型中频次小于或等于第一阈值的特征对应的模型参数，本实施例在依次更新机器学习模型中频次大于第一阈值的特征对应的模型参数之后，还可以包括：将机器学习模型中频次小于或等于第一阈值的特征对应的模型参数设置为预设稀疏值，以便在机器学习模型完成训练后，去掉预设稀疏值，得到剩余模型参数。
示例性的，在模型训练过程中，在将每一个历史推荐信息的历史特征向量输入到机器学习模型中进行训练的过程中，可以根据历史特征向量中包含的特征确定该机器学习模型中预设的每个特征的频次，此时对于频次大于第一阈值 的特征对应的模型参数进行修正，对于频次小于或等于第一阈值的特征对应的模型参数，由于后续该特征对应的模型参数不再参与学习训练，因此直接将频次小于或等于第一阈值的特征对应的模型参数设置为预设稀疏值(本实施例中该预设稀疏值为0)，后续在机器学习模型完成训练后，直接去掉设置为预设稀疏值的模型参数，进而保留频次大于第一阈值的特征对应的模型参数。
示例性的，本实施例中在机器学习模型中配置改进的跟随正则化逻辑(Followed the Regularized Leader，FTRL)算法，通过该改进的FTRL算法对在模型训练过程中依次对频次大于第一阈值的特征对应的模型参数进行修正。相关技术中的FTRL算法如下：
Figure PCTCN2020099662-appb-000001
其中，g为某一特征对应的模型参数的损失系数，z和n为机器学习模型在训练过程中的中间变量，w为特征对应的模型参数，α、β、λ 1和λ 2为稀疏性训练的限定参数。
本实施例中在上述FTRL算法的基础上，对机器学习模型中特征的频次进行限定，改进的FTRL算法如下：
Figure PCTCN2020099662-appb-000002
其中，f为机器学习模型中某一特征的频次，λ 3为特征的频次限定的第一阈值；改进后的FTRL算法对机器学习模型中频次大于第一阈值的特征对应的模型参数依然进行常规的训练，而直接将频次小于或等于第一阈值的特征对应的模型参数设置为0，不参与后续的机器学习过程，也就是在机器学习模型的训练过程中，根据历史推荐信息的历史特征向量，确定机器学习模型中预设的每个特征的频次，只有特征频次大于第一阈值后才会对该特征对应的模型参数进行机器学习，否则该特征对应的模型参数一直设置为0；后续在机器学习模型完成训练后，利用不为0的模型参数生成对应的推荐模型，在保证推荐准确性的基础上，减少了推荐模型的大小，提高了推荐模型的效率。
在步骤S230中，线上发布推荐模型。
示例性的，在生成对应的推荐模型后，可以直接将该推荐模型在线上发布，也就是应用到执行本申请任意实施例提供的信息推荐的方法的推荐服务器中，以便在线上直接采用该推荐模型预测每个推荐信息的点击通过率，为用户推荐更感兴趣的内容，提高信息推荐的准确性和效率。
本实施例提供的技术方案，通过机器学习模型对历史推荐信息预测的点击通过率不断更新机器学习模型中频次大于第一阈值的特征对应的模型参数，并在机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，并利用剩余模型参数生成对应的推荐模型，无需每次对频次小于或等于第一阈值的特征对应的模型参数进行修正，提高了模型训练的高效性，减少了推荐模 型的大小，同时降低了模型参数和高维稀疏特征的内存占用。
图3A为本申请一实施例提供的另一种信息推荐的方法中训练推荐模型的流程图，图3B为本申请一实施例提供的另一种信息推荐的方法中推荐模型的训练过程的原理示意图。本实施例是在上述实施例的基础上进行细化。由于在机器学习模型完成训练后，在模型规模达到内存上限的情况下，会造成模型预测的拥塞，因此需要对频次较低的特征对应的模型参数再次进行过滤，以保证推荐模型的高效性。本实施例主要对于推荐模型的二次优化过程进行详细的解释说明。
如图3A所示，该方法可以包括步骤S310至步骤S340。
在步骤S310中，根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新机器学习模型中频次大于第一阈值的特征对应的模型参数。
在步骤S320中，在机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，在机器学习模型规模达到内存上限的情况下，在剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，并利用过滤后的剩余模型参数生成推荐模型。
示例性的，在机器学习模型完成训练后，可以去掉该机器学习模型中频次小于或等于第一阈值的特征对应的模型参数，此时在该机器学习模型的规模还是达到内存上限的情况下，在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中继续过滤出频次小于或等于第二阈值的特征对应的模型参数，此时该第二阈值大于第一阈值，释放相应的内存空间，并利用过滤后的剩余模型参数生成对应的推荐模型，进而减少推荐模型的大小。
在步骤S330中，在机器学习模型完成当前时段下的训练后，根据预设的频次衰减系数更新机器学习模型中不同特征的频次，并在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中过滤频次小于或等于第三阈 值的特征对应的模型参数，利用过滤后的剩余模型参数生成当前时段下的推荐模型。
示例性的，由于在信息推荐过程中不断存在新的推荐业务，从而存在新的特征，而如果某一时段更新的推荐业务特征仅热门一段时间，后期不再对该业务特征进行参考，但机器学习模型中在热门时段已经记录该业务特征的频次大于第一阈值，在后期训练时还是会一直参考该业务特征，但该业务特征已经不再对模型训练有任何贡献，因此为了保证推荐模型的准确性，本实施例可以对机器学习模型进行分时段训练，获取当前时段下的历史推荐信息，不断对机器学习模型中频次大于第一阈值的特征对应的模型参数进行修正，在机器学习模型完成当前时段下的训练后，由于前一时段的某一特征在当前时段可能不再出现，因此根据预设的频次衰减系数更新机器学习模型中不同特征的频次，此时如果某一特征一直不出现，则该特征的频次会不断衰减，直至小于或等于第三阈值而被去掉，如图3B所示，本实施例中首先去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，然后根据预设的频次衰减系数更新机器学习模型中不同特征的频次，之后在剩余模型参数中再次过滤频次小于或等于第三阈值的特征对应的模型参数，进而利用过滤后的剩余模型参数生成当前时段下的推荐模型，提高推荐模型的准确性和高效性。
需要说明的是，本实施例中过滤频次小于或等于第三阈值的特征对应的模型参数可以根据训练后的机器学习模型的规模来判断是否执行，在机器学习模型完成当前时段下的训练后规模未达到内存上限的情况下，可以不用再次过滤频次小于或等于第三阈值的特征对应的模型参数，而直接利用去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数生成对应的推荐模型。本实施例中S320和S330可以选择执行其中的一个。
在步骤S340中，线上发布推荐模型。
本实施例提供的技术方案，在机器学习模型完成训练后，在机器学习模型 的规模达到内存上限的情况下，在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中再次过滤频次小于或等于第二阈值的特征对应的模型参数，利用过滤后的剩余模型参数生成对应的推荐模型，减少了推荐模型的大小，降低了模型参数和高维稀疏特征的内存占用；同时通过分时段训练机器学习模型，并根据频次衰减系数不断更新不同特征的频次，保证不同时段下推荐模型的准确性和高效性。图4为本申请一实施例提供的又一种信息推荐的方法的流程图。本实施例是在上述实施例的基础上进行细化。本实施例中对于具体的信息推荐过程进行说明，如图4所示，本实施例中可以包括步骤S410至步骤S430。
在步骤S410中，获取推荐模型中去掉与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数。
示例性的，在根据用户请求获取到大量推荐信息后，需要通过线上发布的推荐模型对每个推荐信息的特征向量进行预测，因此首先确定线上发布的推荐模型，并查找出该推荐模型中去掉与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数，也就是频次大于第一阈值的特征对应的模型参数，以便对每个推荐信息的特征向量中对应的特征进行分析。
在步骤S420中，根据剩余模型参数以及特征向量中与剩余模型参数对应的特征，计算推荐信息的点击通过率。
示例性的，根据推荐模型中保留的剩余模型参数，在每个推荐信息的特征向量中筛选出该剩余模型参数对应的特征，此时模型参数可以是特征的权重系数，采用加权求和的方式将每一剩余模型参数和推荐信息的特征向量中与该剩余模型参数对应的特征相乘并求和，将推荐信息的特征向量中每个剩余模型参数对应的特征加权求和的结果，作为本实施例中推荐信息的点击通过率，从而减少推荐模型中的计算量，提高信息推荐的效率。
在步骤S430中，根据点击通过率对推荐信息排序，并构建对应的下发列表 下发给用户。
示例性的，在得到多个推荐信息的点击通过率后，根据该点击通过率的高低为多个推荐信息进行排序，并根据用户的需求选择排序靠前的推荐信息构建对应的下发列表，并将该下发列表显示给用户，保证信息推荐的准确性。
本实施例提供的技术方案，在推荐模型的训练过程中，去掉与推荐信息的特征向量中的特征相同且频次小于或等于第一阈值的特征对应的模型参数，仅保留频次大于第一阈值的特征对应的模型参数，从而减小了推荐模型的大小，降低了推荐模型中的参数占用空间，同时在通过该推荐模型中的模型参数对推荐信息的特征向量进行预测时，无需查询特征向量中的每一特征对应的模型参数，降低了查询过程中的耗费时长，仅对推荐信息的特征向量中与推荐模型中保留的模型参数对应的特征进行分析，在保证推荐准确性的同时，提高了信息推荐效率，进而为用户推送更加优质的内容，改善了信息推送效果。
图5为本申请一实施例提供的一种信息推荐的装置的结构示意图，如图5所示，该装置可以包括：
点击预测模块510，设置为通过推荐模型对推荐信息的特征向量进行预测，得到推荐信息的点击通过率，其中，该推荐模型中去掉了与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
信息推荐模块520，设置为根据点击通过率，下发推荐信息。
本实施例提供的技术方案，在推荐模型的训练过程中，去掉与推荐信息的特征向量中的特征相同且频次小于或等于第一阈值的特征对应的模型参数，仅保留频次大于第一阈值的特征对应的模型参数，从而减小了推荐模型的大小，降低了推荐模型中的参数占用空间，同时在通过该推荐模型中的模型参数对推荐信息的特征向量进行预测时，无需查询特征向量中的每一特征对应的模型参数，降低了查询过程中的耗费时长，仅对推荐信息的特征向量中与推荐模型中保留的模型参数对应的特征进行分析，在保证推荐准确性的同时，提高了信息 推荐效率。
在一实施例中，上述点击预测模块510，可以设置为：
获取推荐模型中去掉与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数；
根据剩余模型参数以及特征向量中与剩余模型参数对应的特征，计算推荐信息的点击通过率。
在一实施例中，上述推荐模型可以通过模型训练系统包括的模型参数更新模块和推荐模型生成模块执行下述操作生成：
模型参数更新模块，设置为根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新机器学习模型中频次大于第一阈值的特征对应的模型参数；
推荐模型生成模块，设置为在机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用最新的剩余模型参数生成推荐模型，该剩余模型参数为机器学习模型中频次大于第一阈值的特征对应的模型参数。
在一实施例中，上述模型参数更新模块，可以设置为：
通过机器学习模型对当前历史推荐信息的历史特征向量进行预测，得到对应的当前点击通过率；
采用预设损失函数计算当前点击通过率在频次大于第一阈值的特征对应的模型参数下的损失系数，并根据该损失系数修正机器学习模型中对应的模型参数，继续通过机器学习模型对下一个历史推荐信息的历史特征向量进行预测，并修正机器学习模型中频次大于第一阈值的特征对应的模型参数，直至机器学习模型完成训练。
在一实施例中，上述模型训练系统，还可以包括：
稀疏参数设置模块，设置为将机器学习模型中频次小于或等于第一阈值的特征对应的模型参数设置为预设稀疏值，以便在机器学习模型完成训练后，去掉设置为预设稀疏值的模型参数。
在一实施例中，上述推荐模型生成模块，可以设置为：
在机器学习模型规模达到内存上限的情况下，在剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，并利用过滤后的剩余模型参数生成推荐模型，该第二阈值大于第一阈值。
在一实施例中的，上述推荐模型生成模块，还可以设置为：
在机器学习模型完成当前时段下的训练后，根据预设的频次衰减系数更新机器学习模型中不同特征的频次，并在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，利用过滤后的剩余模型参数生成当前时段下的推荐模型。
在一实施例中的，上述模型训练系统，还可以包括：
模型发布模块，设置为线上发布推荐模型。
在一实施例中，上述模型参数为推荐模型中不同特征的权重系数。
在一实施例中，上述信息推荐模块520，可以设置为：
根据点击通过率对推荐信息排序，并构建对应的下发列表下发给用户。
本实施例提供的信息推荐的装置可适用于上述任意实施例提供的信息推荐的方法。
图6为本申请一实施例提供的一种模型训练系统的原理架构图。本实施例是在上述实施例的基础上进行细化。参照图6，该模型训练系统可以包括：训练服务器集群610和参数服务器集群620，参数服务器集群620中存储预设的机器学习模型中不同特征对应的模型参数。
示例性的，训练服务器集群610依次从参数服务器集群620获取机器学习 模型中不同特征对应的最新模型参数，根据该模型参数对历史推荐信息的历史特征向量进行处理，得到对应的点击通过率，并计算机器学习模型中频次大于第一阈值的特征对应的模型参数的损失系数，将该损失系数返回给参数服务器集群620；参数服务器集群620根据该损失系数更新对应的模型参数并反馈给训练服务器集群610，直至机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用所述剩余模型参数生成对应的推荐模型。
同时，训练服务器集群610将机器学习模型中频次小于或等于第一阈值的特征对应的模型参数设置为预设稀疏值，并返回给参数服务器集群620；参数服务器集群620将频次小于或等于第一阈值的特征对应的模型参数修正为预设稀疏值，并在机器学习模型完成训练后，去掉设置为预设稀疏值的模型参数。
在一实施例中，为了保证模型的高效性，参数服务器集群620在机器学习模型完成训练后，在所述机器学习模型规模达到内存上限的情况下，在剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，并利用过滤后的剩余模型参数生成推荐模型，第二阈值大于第一阈值。
示例性的，为了保证每个时段下的推荐模型的准确性，参数服务器集群620在机器学习模型完成当前时段下的训练后，根据预设的频次衰减系数更新机器学习模型中不同特征的频次，并在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中过滤频次小于或等于第三阈值的特征对应的模型参数，利用过滤后的剩余模型参数生成当前时段下的推荐模型。
此外，模型训练系统中还可以包括推荐服务器630，参数服务器集群620将生成的推荐模型发布到推荐服务器630中，推荐服务器630设置为执行本申请任意实施例提供的信息推荐的方法。
需要说明的是，训练服务器集群610中分布式部署多个训练服务器，该训练服务器可以并行执行机器学习模型的训练，提高模型的训练速率；同时，参 数服务器集群620中分布式部署多个参数服务器，该参数服务器中分片存储机器学习模型中不同特征对应的模型参数，减少数据的存储压力；而且将模型训练过程中的训练步骤和参数修正步骤分离开来，降低了模型训练的计算压力。本实施例提供的技术方案，通过训练服务器集群和参数服务器集群共同训练推荐模型，从而将训练步骤和参数更新步骤分离开来，将训练和模型参数分散到不同的计算节点上，减小单台服务器的计算压力，提高了推荐模型的训练速率。
图7为本申请一实施例提供的一种推荐服务器的结构示意图，如图7所示，该推荐服务器包括处理器70、存储装置71和通信装置72；推荐服务器中处理器70的数量可以是至少一个，图7中以一个处理器70为例；推荐服务器中的处理器70、存储装置71和通信装置72可以通过总线或其他方式连接，图7中以通过总线连接为例。
存储装置71作为一种计算机可读存储介质，可设置为存储软件程序、计算机可执行程序以及模块，如本申请任意实施例所述的信息推荐的方法对应的程序指令/模块。处理器70通过运行存储在存储装置71中的软件程序、指令以及模块，从而执行推荐服务器的各种功能应用以及数据处理，即实现上述信息推荐的方法。
存储装置71可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的应用程序；存储数据区可存储根据终端的使用所创建的数据等。此外，存储装置71可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他非易失性固态存储器件。在一些实例中，存储装置71可包括相对于处理器70远程设置的存储器，这些远程存储器可以通过网络连接至推荐服务器。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
通信装置72可设置为实现推荐服务器与客户端之间的网络连接或者移动数据连接。
本实施例提供的一种推荐服务器可用于执行上述任意实施例提供的信息推荐的方法。
本申请一实施例还提供了一种计算机可读存储介质，计算机可读存储介质存储有计算机程序，该计算机程序被处理器执行时，可实现上述任意实施例中的信息推荐的方法。该方法可以包括：
通过推荐模型对推荐信息的特征向量进行预测，得到推荐信息的点击通过率，其中，该推荐模型中去掉了与特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
根据点击通过率，下发推荐信息。
当然,本申请实施例所提供的一种包含计算机可执行指令的存储介质，其计算机可执行指令不限于如上所述的方法操作，还可以执行本申请任意实施例所提供的信息推荐的方法中的相关操作。
通过以上关于实施方式的描述，所属领域的技术人员可以清楚地了解到，本申请可借助软件及必需的通用硬件来实现，当然也可以通过硬件实现，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对相关技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品可以存储在计算机可读存储介质中，如计算机的软盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、闪存(FLASH)、硬盘或光盘等，包括多个指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。
值得注意的是，上述信息推荐的装置的实施例中，所包括的各个单元和模块只是按照功能逻辑进行划分的，但并不局限于上述的划分，只要能够实现相应的功能即可；另外，各功能单元的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。

Claims (18)

  1. 一种信息推荐的方法，包括：
    通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，其中，所述推荐模型中去掉了与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
    根据所述点击通过率，下发所述推荐信息。
  2. 根据权利要求1所述的方法，其中，所述通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，包括：
    获取所述推荐模型中去掉与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数；
    根据所述剩余模型参数以及所述特征向量中与所述剩余模型参数对应的特征，计算所述推荐信息的点击通过率。
  3. 根据权利要求1所述的方法，其中，所述推荐模型通过模型训练系统执行下述操作生成：
    根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新所述机器学习模型中频次大于第一阈值的特征对应的模型参数；
    在所述机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用所述剩余模型参数生成所述推荐模型，所述剩余模型参数为所述机器学习模型中频次大于第一阈值的特征对应的模型参数。
  4. 根据权利要求3所述的方法，其中，所述根据历史推荐信息的历史特征向量在预设的机器学习模型中的点击通过率，依次更新所述机器学习模型中频次大于第一阈值的特征对应的模型参数，包括：
    通过所述机器学习模型对当前历史推荐信息的历史特征向量进行预测，得到对应的当前点击通过率；
    采用预设损失函数计算所述当前点击通过率在频次大于第一阈值的特征对应的模型参数下的损失系数，并根据所述损失系数修正所述机器学习模型中对 应的模型参数，继续通过所述机器学习模型对下一个历史推荐信息的历史特征向量进行预测，并修正所述机器学习模型中频次大于第一阈值的特征对应的模型参数，直至所述机器学习模型完成训练。
  5. 根据权利要求3所述的方法，在依次更新所述机器学习模型中频次大于第一阈值的特征对应的模型参数之后，还包括：
    将所述机器学习模型中频次小于或等于所述第一阈值的特征对应的模型参数设置为预设稀疏值，以便在所述机器学习模型完成训练后，去掉设置为预设稀疏值的模型参数。
  6. 根据权利要求3所述的方法，其中，所述利用所述剩余模型参数生成所述推荐模型，包括：
    在所述机器学习模型规模达到内存上限的情况下，在所述剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，并利用过滤后的剩余模型参数生成所述推荐模型，所述第二阈值大于所述第一阈值。
  7. 根据权利要求3所述的方法，其中，所述在所述机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用所述剩余模型参数生成所述推荐模型，包括：
    在所述机器学习模型完成当前时段下的训练后，根据预设的频次衰减系数更新所述机器学习模型中不同特征的频次，并在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中过滤频次小于或等于第三阈值的特征对应的模型参数，利用过滤后的剩余模型参数生成当前时段下的推荐模型。
  8. 根据权利要求3-7任一项所述的方法，在利用所述剩余模型参数生成所述推荐模型之后，还包括：
    线上发布所述推荐模型。
  9. 根据权利要求1-7任一项所述的方法，其中，所述模型参数为所述推荐模型中不同特征的权重系数。
  10. 根据权利要求1所述的方法，其中，根据所述点击通过率，下发所述 推荐信息，包括：
    根据所述点击通过率对所述推荐信息排序，并构建对应的下发列表下发给用户。
  11. 一种信息推荐的装置，包括：
    点击预测模块，设置为通过推荐模型对推荐信息的特征向量进行预测，得到所述推荐信息的点击通过率，其中，所述推荐模型中去掉了与所述特征向量的特征相同且频次小于或等于第一阈值的特征对应的模型参数；
    信息推荐模块，设置为根据所述点击通过率，下发所述推荐信息。
  12. 一种推荐服务器，包括：
    至少一个处理器；
    存储装置，设置为存储至少一个程序；
    当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现如权利要求1-10中任一项所述的信息推荐的方法。
  13. 一种模型训练系统，包括：训练服务器集群和参数服务器集群，所述参数服务器集群中存储预设的机器学习模型中不同特征对应的模型参数；
    所述训练服务器集群依次从所述参数服务器集群获取所述机器学习模型中不同特征对应的最新模型参数，根据该模型参数对历史推荐信息的历史特征向量进行处理，得到对应的点击通过率，并计算所述机器学习模型中频次大于第一阈值的特征对应的模型参数的损失系数，将所述损失系数返回给所述参数服务器集群；
    所述参数服务器集群根据该损失系数更新对应的模型参数并反馈给所述训练服务器集群，直至所述机器学习模型完成训练后，去掉频次小于或等于第一阈值的特征对应的模型参数，得到剩余模型参数，并利用所述剩余模型参数生成对应的推荐模型。
  14. 根据权利要求13所述的系统，其中，所述训练服务器集群将所述机器学习模型中频次小于或等于所述第一阈值的特征对应的模型参数设置为预设稀 疏值，并返回给所述参数服务器集群；
    所述参数服务器集群将频次小于或等于所述第一阈值的特征对应的模型参数修正为所述预设稀疏值，并在所述机器学习模型完成训练后，去掉设置为预设稀疏值的模型参数。
  15. 根据权利要求13所述的系统，其中，所述参数服务器集群在所述机器学习模型完成训练后，在所述机器学习模型规模达到内存上限的情况下，在所述剩余模型参数中过滤频次小于或等于第二阈值的特征对应的模型参数，并利用过滤后的剩余模型参数生成所述推荐模型，所述第二阈值大于所述第一阈值。
  16. 根据权利要求13所述的系统，其中，所述参数服务器集群在所述机器学习模型完成当前时段下的训练后，根据预设的频次衰减系数更新所述机器学习模型中不同特征的频次，并在去掉频次小于或等于第一阈值的特征对应的模型参数后的剩余模型参数中过滤频次小于或等于第三阈值的特征对应的模型参数，利用过滤后的剩余模型参数生成当前时段下的推荐模型。
  17. 根据权利要求13-16任一项所述的系统，还包括权利要求12所述的推荐服务器，所述参数服务器集群将生成的所述推荐模型发布到所述推荐服务器中。
  18. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时，实现如权利要求1-10中任一项所述的信息推荐的方法。
PCT/CN2020/099662 2019-10-31 2020-07-01 一种信息推荐的方法、装置、推荐服务器和存储介质 Ceased WO2021082500A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US17/772,925 US12614108B2 (en) 2019-10-31 2020-07-01 Method for recommending information, recommendation server, and storage medium
EP20882554.7A EP4053716A4 (en) 2019-10-31 2020-07-01 INFORMATION RECOMMENDATION METHOD, DEVICE, RECOMMENDATION SERVER AND STORAGE MEDIUM

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911053819.3A CN110825966B (zh) 2019-10-31 2019-10-31 一种信息推荐的方法、装置、推荐服务器和存储介质
CN201911053819.3 2019-10-31

Publications (1)

Publication Number Publication Date
WO2021082500A1 true WO2021082500A1 (zh) 2021-05-06

Family

ID=69551843

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/099662 Ceased WO2021082500A1 (zh) 2019-10-31 2020-07-01 一种信息推荐的方法、装置、推荐服务器和存储介质

Country Status (3)

Country Link
EP (1) EP4053716A4 (zh)
CN (1) CN110825966B (zh)
WO (1) WO2021082500A1 (zh)

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110825966B (zh) * 2019-10-31 2022-03-04 广州市百果园信息技术有限公司 一种信息推荐的方法、装置、推荐服务器和存储介质
CN111478963B (zh) * 2020-04-07 2022-05-20 北京奇艺世纪科技有限公司 消息推送方法、装置、电子设备及计算机可读存储介质
CN113495715A (zh) * 2020-04-08 2021-10-12 北京意锐新创科技有限公司 适用于支付设备管控平台的语音下发方法和装置
CN111970335B (zh) * 2020-07-30 2021-09-07 腾讯科技（深圳）有限公司 一种信息推荐的方法、装置及存储介质
CN114443671B (zh) * 2020-11-04 2025-06-03 腾讯科技（深圳）有限公司 推荐模型的更新方法、装置、计算机设备和存储介质
CN112437349B (zh) * 2020-11-10 2022-09-23 杭州时趣信息技术有限公司 一种视频流推荐方法及相关装置
CN112340609B (zh) * 2020-11-25 2023-06-16 广州三叠纪元智能科技有限公司 参数信息的配置方法、电箱、服务器和存储介质
CN112925926B (zh) * 2021-01-28 2022-04-22 北京达佳互联信息技术有限公司 多媒体推荐模型的训练方法、装置、服务器以及存储介质
CN113794900B (zh) * 2021-08-31 2023-04-07 北京达佳互联信息技术有限公司 视频处理方法和装置
CN114463041B (zh) * 2021-12-31 2026-04-21 深圳市东信时代信息技术有限公司 智能预警方法、装置、设备及存储介质
CN117271462A (zh) * 2023-08-25 2023-12-22 微梦创科网络科技（中国）有限公司 一种数据存储方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107203891A (zh) * 2016-03-17 2017-09-26 阿里巴巴集团控股有限公司 一种自动多阀值特征过滤方法及装置
CN108965951A (zh) * 2018-07-12 2018-12-07 北京奇艺世纪科技有限公司 广告的播放方法及装置
US20190012387A1 (en) * 2017-07-10 2019-01-10 Facebook, Inc. Selecting content for presentation to a user of a social networking system based on a topic associated with a group of which the user is a member
CN109299994A (zh) * 2018-07-27 2019-02-01 北京三快在线科技有限公司 推荐方法、装置、设备及可读存储介质
CN110825966A (zh) * 2019-10-31 2020-02-21 广州市百果园信息技术有限公司 一种信息推荐的方法、装置、推荐服务器和存储介质

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2431488A (en) * 2005-10-11 2007-04-25 Hewlett Packard Development Co Data transfer device
CN103310003A (zh) * 2013-06-28 2013-09-18 华东师范大学 一种基于点击日志的新广告点击率预测方法及系统
CN105160548A (zh) * 2015-08-20 2015-12-16 北京奇虎科技有限公司 对广告点击率进行预测的方法及装置
CN107045511B (zh) * 2016-02-05 2021-03-02 阿里巴巴集团控股有限公司 一种目标特征数据的挖掘方法和装置
CN107122327B (zh) * 2016-02-25 2021-06-29 阿里巴巴集团控股有限公司 一种利用训练数据训练模型的方法和训练系统
CN107944589A (zh) * 2016-10-12 2018-04-20 北京奇虎科技有限公司 广告点击率的预测方法和预测装置
CN108960264A (zh) * 2017-05-19 2018-12-07 华为技术有限公司 分类模型的训练方法及装置
CN107515909B (zh) * 2017-08-11 2020-05-19 深圳市云网拜特科技有限公司 一种视频推荐方法及系统
CN108062573A (zh) * 2017-12-29 2018-05-22 广东欧珀移动通信有限公司 模型训练方法及装置
CN110321422B (zh) * 2018-03-28 2023-04-14 腾讯科技（深圳）有限公司 在线训练模型的方法、推送方法、装置以及设备
CN109582903B (zh) * 2018-11-29 2023-10-20 广州市百果园信息技术有限公司 一种信息展示的方法、装置、设备和存储介质
CN109992710B (zh) * 2019-02-13 2021-10-15 网易传媒科技（北京）有限公司 点击率预估方法、系统、介质和计算设备
CN110378434B (zh) * 2019-07-24 2025-01-10 腾讯科技（深圳）有限公司 点击率预测模型的训练方法、推荐方法、装置及电子设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107203891A (zh) * 2016-03-17 2017-09-26 阿里巴巴集团控股有限公司 一种自动多阀值特征过滤方法及装置
US20190012387A1 (en) * 2017-07-10 2019-01-10 Facebook, Inc. Selecting content for presentation to a user of a social networking system based on a topic associated with a group of which the user is a member
CN108965951A (zh) * 2018-07-12 2018-12-07 北京奇艺世纪科技有限公司 广告的播放方法及装置
CN109299994A (zh) * 2018-07-27 2019-02-01 北京三快在线科技有限公司 推荐方法、装置、设备及可读存储介质
CN110825966A (zh) * 2019-10-31 2020-02-21 广州市百果园信息技术有限公司 一种信息推荐的方法、装置、推荐服务器和存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4053716A4 *

Also Published As

Publication number Publication date
EP4053716A1 (en) 2022-09-07
CN110825966A (zh) 2020-02-21
EP4053716A4 (en) 2022-12-21
CN110825966B (zh) 2022-03-04
US20220405641A1 (en) 2022-12-22

Similar Documents

Publication Publication Date Title
WO2021082500A1 (zh) 一种信息推荐的方法、装置、推荐服务器和存储介质
CN110489644B (zh) 信息推送方法、装置、计算机可读存储介质和计算机设备
EP3353670A1 (en) Regularized model adaptation for in-session recommendations
US10348550B2 (en) Method and system for processing network media information
CN108431833A (zh) 端到端深度协作过滤
CN112883265A (zh) 信息推荐方法、装置、服务器及计算机可读存储介质
US9864951B1 (en) Randomized latent feature learning
JP2018526710A (ja) 情報推薦方法および情報推薦装置
WO2015188699A1 (zh) 推荐项目的方法和装置
CN110633796B (zh) 模型更新方法、装置、电子设备及存储介质
JP7549668B2 (ja) パターンベースの分類
WO2016107354A1 (zh) 提供用户个性化资源消息推送的方法和装置
US10996944B2 (en) Automated software selection using matrix factorization
CN116823410B (zh) 数据处理方法、对象处理方法、推荐方法及计算设备
CN119831684B (zh) 物品推荐方法、装置、存储介质及电子设备
CN115391589A (zh) 内容召回模型的训练方法、装置、电子设备及存储介质
CN113672793A (zh) 一种信息召回方法、装置、电子设备及存储介质
CN105260458A (zh) 一种用于显示装置的视频推荐方法及显示装置
CN115203577B (zh) 对象推荐方法、对象推荐模型的训练方法及装置
CN116756387A (zh) 一种基于时间感知的图神经网络会话推荐方法及系统
CN114637926A (zh) 内容推荐方法、装置、电子设备及存储介质
RU2785764C1 (ru) Способ рекомендации информации, устройство, сервер рекомендаций и запоминающее устройство
US12614108B2 (en) Method for recommending information, recommendation server, and storage medium
CN112258297A (zh) 推送物品的描述信息的方法、装置和计算机可读存储介质
CN108304370B (zh) 数据更新方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20882554

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2020882554

Country of ref document: EP

Effective date: 20220531