CN103473291B

CN103473291B - Personalized service recommendation system and method based on latent semantic probability models

Info

Publication number: CN103473291B
Application number: CN201310392446.9A
Authority: CN
Inventors: 彭启民; 胡堰; 胡晓惠
Original assignee: Institute of Software of CAS
Current assignee: Institute of Software of CAS
Priority date: 2013-09-02
Filing date: 2013-09-02
Publication date: 2017-01-18
Anticipated expiration: 2033-09-02
Also published as: CN103473291A

Abstract

The invention relates to a personalized service recommendation system and method based on a hidden semantic probability model, which belongs to the technical field of service computing and includes the following steps: determining a QoS index system for evaluating the performance of a series of similarly functioning services; establishing a user, The implicit semantic probability model between user index preference and service context; collect the index preference information provided by different users in the system when they use services with different functions in different service contexts through interaction with the system, and store them in the historical experience database; Use the collected data to train the parameters of the latent semantic probability model; when the user is not familiar with a certain service situation, but needs to call a specific function service in this service situation, use the trained latent semantic probability model to evaluate the user's index preference Prediction: According to the predicted user's personalized QoS index preference, the candidate services are comprehensively screened, and the service that is closest to the user's needs is selected from them, so as to realize personalized service recommendation.

Description

A Personalized Service Recommendation System and Method Based on Latent Semantic Probability Model

技术领域technical field

本发明属于服务计算技术领域，具体涉及一种基于隐语义概率模型的个性化服务推荐系统及方法。The invention belongs to the technical field of service computing, and in particular relates to a personalized service recommendation system and method based on a hidden semantic probability model.

背景技术Background technique

随着互联网技术的飞速发展，服务计算技术得到了广泛应用，Web服务就是这样一种分布运行于Internet之上、支持不同平台之间互操作的松耦合软件系统，它主要通过“发布-查找-绑定”的模式允许服务使用者和提供者之间形成松散的绑定关系，这为服务的使用奠定了基础。但web服务的使用者和提供者相分离的特性，增加了服务使用者理解服务的难度，同时随着Internet上运行的web服务数量不断增多，服务使用者需要从众多功能相似的服务中选出最符合自身需求的一个或一组服务，这对大多数缺乏专业知识的服务使用者来说无疑是一项繁重的任务，所以发展有效的服务推荐技术是服务选择的必然需求。With the rapid development of Internet technology, service computing technology has been widely used. Web service is such a loosely coupled software system that runs on the Internet and supports interoperability between different platforms. The "binding" pattern allows for a loosely bound relationship between service consumers and providers, which lays the foundation for service usage. However, the separation of users and providers of web services increases the difficulty for service users to understand services. At the same time, as the number of web services running on the Internet continues to increase, service users need to choose from many services with similar functions. It is undoubtedly a heavy task for most service users who lack professional knowledge to find a service or a group of services that best meet their own needs, so the development of effective service recommendation technology is an inevitable requirement for service selection.

经过对现有技术的检索发现，中国专利申请号200710162463.8，记载了一种自适应服务推荐设备，该装置主要包括：语义分析装置、服务选择装置、服务推荐装置；语义分析装置用于对用户的查询进行语义上的分析；服务选择装置用于找出与语义分析后的查询对应的选择的服务，并根据选择的服务更新服务相关数据库；最后服务推荐装置用于利用获取的选择的服务来查找服务相关数据库以向用户推荐相关服务。After searching the prior art, it is found that Chinese patent application number 200710162463.8 records an adaptive service recommendation device, which mainly includes: a semantic analysis device, a service selection device, and a service recommendation device; the semantic analysis device is used for the user's Perform semantic analysis on the query; the service selection device is used to find out the selected service corresponding to the query after the semantic analysis, and update the service-related database according to the selected service; finally, the service recommendation device is used to use the obtained selected service to search Serving related databases to recommend related services to users.

进一步检索发现，中国专利申请号200910236492.3，记载了一种个性化服务推荐系统和方法，该方法主要包括：用户信息收集器监控在终端进行的各种操作信息，并进行预处理后存入用户信息数据库，若用户信息数据库方法更新，则启动用户行为分析器进行分析；用户行为分析器扫描用户信息数据库，提取新的用户信息并计入资源信息数据库，计算新的推荐策略并记入推荐策略数据库；上下文感知处理器感知用户的当前上下文，输出当前的上下文描述信息，启动个性化推荐处理器；个性化推荐处理器接收来自上下文感知处理器的消息后，获得当前上下文信息，通过检索推荐策略数据库，获得匹配的左右推荐策略，并按照左右推荐策略通过检索资源信息数据库，匹配合适的资源信息，实时生成个性化推荐服务。Further search found that Chinese patent application number 200910236492.3 records a personalized service recommendation system and method, which mainly includes: the user information collector monitors various operation information performed on the terminal, and stores user information after preprocessing Database, if the user information database method is updated, start the user behavior analyzer for analysis; the user behavior analyzer scans the user information database, extracts new user information and counts it into the resource information database, calculates a new recommendation strategy and records it into the recommendation strategy database ;The context-aware processor perceives the user's current context, outputs the current context description information, and starts the personalized recommendation processor; after receiving the message from the context-aware processor, the personalized recommendation processor obtains the current context information, and retrieves the recommendation strategy database , obtain the matching left and right recommendation strategies, and match the appropriate resource information by searching the resource information database according to the left and right recommendation strategies, and generate personalized recommendation services in real time.

进一步检索发现，中国专利申请号20121014234.2记录了一种基于上下文感知和用户偏好的空间信息服务匹配方法，在用户的服务请求和候选空间信息服务的功能性匹配和非功能性匹配的基础上，考虑上下文敏感的用户偏好，计算出用户服务需求与智能空间中的各候选空间信息服务的匹配度，然后依据匹配度将候选空间信息服务推荐给用户；虽然用户可能具有固定的或者重复的偏好，但这些偏好不是在任何时候都相关，该发明基于上下文感知和用户偏好的空间信息服务匹配方法从上下文的角度对用户偏好进行精简，剔除与用户无关的用户偏好；该发明从功能匹配和非功能匹配两个方面进行匹配度计算，提高空间信息服务匹配的准确率。Further search found that Chinese Patent Application No. 20121014234.2 records a spatial information service matching method based on context awareness and user preference. On the basis of functional matching and non-functional matching between user service requests and candidate spatial information services, consider Context-sensitive user preferences, calculate the matching degree of user service needs and each candidate spatial information service in the smart space, and then recommend the candidate spatial information service to the user according to the matching degree; although the user may have fixed or repeated preferences, but These preferences are not relevant at any time, the invention based on context awareness and user preference spatial information service matching method streamlines user preferences from the perspective of context, and eliminates user preferences irrelevant to users; the invention combines functional matching and non-functional matching Two aspects are used to calculate the matching degree to improve the accuracy of spatial information service matching.

进一步检索发现，中国专利申请号201210253884.2记录了一种用于Web服务推荐的个性化搜索方法，包括以下步骤：步骤1，预处理WSDL文档：通过去除停用词和提取词干两个预处理步骤，形成词袋；步骤2，抽取用户兴趣：使用改进的TF-IDF公式计算词袋中的每一个词的权重，并乘以该词的时间衰减因子，得到新的权重；选择权重由大至小前k个词作为用户的兴趣词，以及每个词的对应权重，组成k维的用户兴趣向量；步骤3，计算兴趣相似度：设定相似度阈值，超过阈值的用户入选为目标用户的邻居用户；步骤4，排序服务检索结果，根据邻居用户的相似度及其选择服务的次数计算服务的推荐预测值，并将检索结果按照推荐预测值降序排列，从而得到个性化搜索结果。Further search found that Chinese Patent Application No. 201210253884.2 records a personalized search method for Web service recommendation, including the following steps: Step 1, preprocessing WSDL documents: through two preprocessing steps of removing stop words and extracting word stems , forming a bag of words; step 2, extracting user interests: use the improved TF-IDF formula to calculate the weight of each word in the bag of words, and multiply it by the time decay factor of the word to get a new weight; select the weight from large to The first k words are used as the user's interest words, and the corresponding weights of each word form a k-dimensional user interest vector; step 3, calculate interest similarity: set a similarity threshold, and users exceeding the threshold are selected as target users Neighbor users; step 4, sorting service retrieval results, calculating service recommendation prediction values based on the similarity of neighbor users and the number of service selections, and sorting the retrieval results in descending order of recommendation prediction values to obtain personalized search results.

上述的方法都涉及了个性化的服务推荐，但很少针对Web服务的非功能属性筛选服务以满足用户的个性化需求，服务的非功能属性往往是服务性能的主要体现，所以如何利用服务的非功能属性客观地评价服务的性能同时又能满足不同用户的个性化需求是服务推荐需要解决的问题。The above methods all involve personalized service recommendation, but few services are screened for the non-functional attributes of web services to meet the individual needs of users. The non-functional attributes of services are often the main manifestation of service performance, so how to use the service Non-functional attributes can objectively evaluate the performance of services while meeting the individual needs of different users is a problem to be solved for service recommendation.

发明内容Contents of the invention

本发明技术解决问题：针对现有技术存在的不足，提供一种基于隐语义概率模型的个性化服务推荐系统及方法，旨在为用户在不熟悉的服务情境下提供指标偏好预测，从而完成基于多维QoS指标的服务综合排序，为用户提供个性化的服务推荐结果。The technology of the present invention solves the problem: Aiming at the deficiencies of the existing technology, a personalized service recommendation system and method based on the hidden semantic probability model is provided, aiming at providing index preference prediction for users in unfamiliar service situations, so as to complete the system based on The comprehensive service ranking of multi-dimensional QoS indicators provides users with personalized service recommendation results.

本发明的技术解决方案：一种个性化服务推荐中基于隐语义概率模型的用户指标偏好预测方法，具体步骤如下：The technical solution of the present invention: a user index preference prediction method based on a hidden semantic probability model in personalized service recommendation, the specific steps are as follows:

步骤1、确定评价服务性能优劣的服务QoS指标体系Step 1. Determine the service QoS index system for evaluating service performance

所述的服务QoS指标体系是指整个Web服务系统统一采用的用于评价一系列功能相似服务性能优劣所用QoS指标的集合，不同的系统可以根据需要选取适当的QoS指标组成自己的QoS指标体系用于评价服务的性能优劣；The service QoS indicator system refers to the collection of QoS indicators used by the entire Web service system to evaluate the performance of a series of similar services. Different systems can select appropriate QoS indicators to form their own QoS indicator systems. To evaluate the performance of the service;

步骤2、建立用户、用户指标偏好以及服务情境三者之间的隐语义概率模型Step 2. Establish an implicit semantic probability model among users, user index preferences, and service contexts

所述的用户指标偏好是指用户对步骤1中所述服务QoS指标体系中各个QoS指标的偏好程度，对每个QoS指标的偏好程度值介于0到1之间，并且对各个QoS指标的偏好值总和为1；所述的服务情境是指在何种场景下使用何种功能的服务，每一个服务情境e用三元组(w₁,w₂,w₃)表示，其中w₁表示服务完成的与业务无关的基本功能，如视频功能、导航功能等，w₂表示用户使用该功能的服务完成的具体业务活动，如学术会议、军事导航等，w₃表示用户调用服务的终端设备，如手机、PC机等；所述的用户、用户指标偏好、服务情境三者之间的隐语义概率模型是指单个用户以不同的概率依赖不同的用户隐类而存在，单则服务情境以不同的概率依赖不同的服务情境隐类存在，指标偏好同时以不同的概率同时依赖不同的用户隐类和服务情境隐类存在的概率模型，模型的图像化表示如7所示。The user index preference refers to the user's degree of preference for each QoS index in the service QoS index system described in step 1, the preference value for each QoS index is between 0 and 1, and the preference for each QoS index The sum of preference values is 1; the service context refers to the service with which function is used in which scenario, and each service context e is represented by a triplet (w ₁ , w ₂ , w ₃ ), where w ₁ represents The basic functions irrelevant to the business completed by the service, such as video function, navigation function, etc., w ₂ represents the specific business activities completed by the user using the service of this function, such as academic conferences, military navigation, etc., w ₃ represents the terminal equipment that the user calls the service , such as mobile phones, PCs, etc.; the implicit semantic probability model between users, user index preferences, and service contexts means that a single user exists with different probabilities depending on different user hidden categories, and a single service context is based on Different probabilities depend on the existence of different service context hidden classes, and indicator preferences depend on the probability models of different user hidden classes and service context hidden classes at the same time with different probabilities. The graphical representation of the model is shown in Figure 7.

所述的用户隐类是指非人为事先确定的而是从历史经验数据中学习得到的用户聚簇；所述的服务情境隐类是指非人为事先确定的而是从历史经验数据中学习得到的服务情境聚簇；The user hidden class refers to user clusters that are not determined in advance by humans but learned from historical experience data; the hidden class of service situations refers to user clusters that are not determined in advance by humans but learned from historical experience data clustering of service scenarios;

步骤3、收集不同用户在不同的服务情境下使用不同功能服务时自主提供的指标偏好信息，作为历史经验数据存入数据库，为训练步骤2中建立的隐语义概率模型的参数作准备，存储格式为（用户，服务情境、指标偏好）三元组；Step 3. Collect the index preference information provided by different users when using different functional services in different service scenarios, and store them in the database as historical experience data to prepare for training the parameters of the latent semantic probability model established in step 2. The storage format It is a triplet of (user, service context, indicator preference);

步骤4、用EM算法及已收集的历史经验数据训练隐语义概率模型的参数Step 4. Use the EM algorithm and the collected historical experience data to train the parameters of the implicit semantic probability model

所述的隐语义概率模型的参数是指所有用户隐类{U₁,U₂,...,U_I}的先验概率P(U_i)(1≤i≤I)、服务情境隐类{E₁,E₂,...,E_J}先验概率P(E_j)(1≤j≤J)、给定一个用户隐类U_i的情况下单个用户u出现的条件概率P(u|U_i)(1≤i≤I)、给定一个服务情境隐类E_j的情况下单则服务情境e出现的条件概率P(e|E_j)(1≤j≤J)、给定用户隐类U_i和服务情境隐类E_j的情况下用户指标偏好向量r出现的概率P(r|U_i,E_j)(1≤i≤I,1≤j≤J)；其中I,J分别表示用户隐类和服务情境隐类的总个数，i,j分别表示用户隐类和服务情境隐类的编号；The parameters of the hidden semantic probability model refer to the prior probability P(U _i )(1≤i≤I) of all user hidden classes {U ₁ , U ₂ ,..., U _I }, service situation hidden class {E ₁ ,E ₂ ,...,E _J } prior probability P(E _j )( _1≤j≤J ), conditional probability P( u|U _i )(1≤i≤I), the conditional probability P(e|E _j )(1≤j≤J) of a single service situation e in the case of a given service situation hidden class E _j , given The probability of user index preference vector r appearing in the case of given user hidden class U _i and service situation hidden class E _j P(r|U _i , E _j )(1≤i≤I, 1≤j≤J); where I , J represent the total number of user hidden classes and service situation hidden classes respectively, i, j represent the numbers of user hidden classes and service situation hidden classes respectively;

步骤5、获得用户需要服务推荐的请求，包括个人信息、服务推荐所依赖的服务情境；Step 5. Obtain the user's request for service recommendation, including personal information and the service context on which the service recommendation depends;

步骤6、用已训练的隐语义概率模型预测指定用户在特定服务情境下的未知指标偏好；Step 6. Use the trained latent semantic probability model to predict the unknown index preference of the specified user in a specific service situation;

步骤7、根据预测出的用户个性化的QoS指标偏好，对候选服务进行综合筛选，从而选出最贴近该用户需求的服务，将推荐结果返回给用户。Step 7. According to the predicted user's personalized QoS index preference, comprehensively screen the candidate services, so as to select the service that is closest to the user's needs, and return the recommendation result to the user.

所述步骤3中收集历史经验数据的过程如下：The process of collecting historical experience data in the step 3 is as follows:

（1）用户在其熟悉的服务情境下调用熟悉的web服务时可以直接与系统交互，使用层次分析法给出自己给出对各个指标偏好的两两比较结果，建立比较矩阵；(1) Users can directly interact with the system when invoking familiar web services in their familiar service contexts, use the AHP to give their own pairwise comparison results for each index preference, and establish a comparison matrix;

$C C = = [\begin{matrix} {c c}_{1111},, & {c c}_{1212},, & . . . . . .,, & {c c}_{11 n no} \\ {c c}_{21 twenty one},, & {c c}_{22 twenty two},, & . . . . . .,, & {c c}_{33 n no} \\ . . & . . & . . \\ . . & . . & . . \\ . . & . . & . . \\ {c c}_{n no 11},, & {c c}_{n no 22},, & . . . . . .,, & {c c}_{nn nn} \end{matrix}],,$

（2）验证比较矩阵的一致性是否在可接受的范围内，若可接受，进入（3），否者返回（1）；(2) Verify whether the consistency of the comparison matrix is within the acceptable range, if acceptable, go to (3), otherwise return to (1);

（3）计算比较矩阵的最大特征值对应的特征向量，将该特征向量归一化后所得到的就是用户对各个QoS指标的偏好向量W_QoS，将（用户、服务情境、指标偏好）三元组存入数据库。(3) Calculate the eigenvector corresponding to the largest eigenvalue of the comparison matrix. After normalizing the eigenvector, the user’s preference vector W _QoS for each QoS indicator is obtained, and the (user, service situation, indicator preference) ternary Groups are stored in the database.

所述步骤4中隐语义概率模型的参数训练过程如下：The parameter training process of hidden semantic probability model in described step 4 is as follows:

（1）从数据库中取出所有的用户、服务情境、指标偏好三元组作为训练数据，是用户在其熟悉的服务情境下利用层次分析法给出的个性化指标偏好权值，每个三元组(u,e,r)表示用户u在服务情境e下对服务的各个QoS指标的偏好权重为r，r是一个K维向量(r₁,r₂,...,r_K)，为简单起见，假设权重向量r的每一维相互独立并且服从正态分布，则联合概率P(u,e,r)可用下面两个公式表示，(1) Take all triples of user, service context, and index preference from the database as training data, which is the personalized index preference weight given by the user in the familiar service context using the AHP. Each triple The group (u,e,r) means that the user u's preference weight for each QoS indicator of the service in the service situation e is r, and r is a K-dimensional vector (r ₁ ,r ₂ ,...,r _K ), which is For simplicity, assuming that each dimension of the weight vector r is independent of each other and obeys a normal distribution, the joint probability P(u,e,r) can be expressed by the following two formulas,

$P P ((u u,, e e,, r r)) = = \underset{i i}{Σ Σ} \underset{j j}{Σ Σ} P P (({U u}_{i i})) P P (({E E.}_{j j})) P P ((u u | | {U u}_{i i})) P P ((e e | | {E E.}_{j j})) P P ((r r | | {U u}_{i i},, {E E.}_{j j})),,$

其中P(r|U_i,E_j)为K维正态分布，i，j分别表示用户隐类和服务情境隐类编号；Among them, P(r|U _i , E _j ) is a K-dimensional normal distribution, and i, j respectively represent the number of user hidden class and service situation hidden class;

（2）E步计算隐类的联合后验概率，(2) Step E calculates the joint posterior probability of hidden classes,

$P P (({U u}_{p p},, {E E.}_{q q} | | u u,, e e,, r r)) = = \frac{{[[P P (({U u}_{p p})) P P (({E E.}_{q q})) P P ((u u | | {U u}_{p p})) P P ((e e | | {E E.}_{q q})) P P ((r r | | {U u}_{p p},, {E E.}_{q q}))]]}^{b b}}{\underset{i i}{Σ Σ} \underset{j j}{Σ Σ} {[[P P (({U u}_{i i})) P P (({E E.}_{j j})) P P ((u u | | {U u}_{i i})) P P ((e e | | {E E.}_{j j})) P P ((r r | | {U u}_{p p},, {E E.}_{q q}))]]}^{b b}} . .$

其中b是一个介于0到1之间的模拟退火参数，P(r|U_p,E_q)为K维正态分布，p,q表示用户隐类和服务情境隐类编号；Where b is a simulated annealing parameter between 0 and 1, P(r|U _p , E _q ) is a K-dimensional normal distribution, and p, q represent the number of user hidden classes and service situation hidden classes;

（3）M步使用E步计算得到的后验概率重新估算模型参数，包括每一个用户隐类和每一个服务情境隐类的先验概率P(U_p)和P(E_q)，每一个用户在给定不同用户隐类时出现的条件概率P(u|U_p)，每一则服务情境在给定不同服务情境隐类时出现的条件概率P(e|E_q)，同时给定不同的用户情境隐类和服务情境隐类时K个指标偏好的正态分布均值和方差 ${σ^{2}}_{k, U_{p}, E_{q}} :$ (3) Step M uses the posterior probability calculated in step E to re-estimate the model parameters, including the prior probability P(U _p ) and P(E _q ) of each user hidden class and each service situation hidden class, each The conditional probability P(u|U _p ) of the user when different user implicit classes are given, the conditional probability P(e|E _q ) of each service situation when the different service situation implicit classes are given, and given The normal distribution mean value of K indicator preferences when different user contextual implicit classes and service contextual implicit classes and variance ${σ^{2}}_{k, u_{p}, {E.}_{q}} :$

$P P (({U u}_{p p})) = = \frac{\underset{l l}{Σ Σ} \underset{j j}{Σ Σ} P P (({U u}_{p p},, {E E.}_{j j} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{L L},,$ $P P (({E E.}_{q q})) = = \frac{\underset{l l}{Σ Σ} \underset{i i}{Σ Σ} P P (({U u}_{i i},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{L L},,$

$P P ((u u | | {U u}_{p p})) = = \frac{\underset{l l : : {u u}_{((l l))} = = u u}{Σ Σ} \underset{j j}{Σ Σ} P P (({U u}_{p p},, {E E.}_{j j} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{L L * * P P (({U u}_{p p}))},,$ $P P ((e e | | {E E.}_{q q})) = = \frac{\underset{l l : : {e e}_{((l l))} = = e e}{Σ Σ} \underset{i i}{Σ Σ} P P (({U u}_{i i},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{L L * * P P (({E E.}_{q q}))},,$

${μ μ}_{k k,, {U u}_{p p},, {E E.}_{q q}} = = \frac{\underset{l l}{Σ Σ} {r r}_{k k ((l l))} P P (({U u}_{p p},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{\underset{l l}{Σ Σ} P P (({U u}_{p p},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))},,$ ${σ σ}^{22}_{k k,, {U u}_{p p},, {E E.}_{q q}} = = \frac{\underset{l l}{Σ Σ} {(({r r}_{k k ((l l))} - - {μ μ}_{k k,, {U u}_{p p},, {E E.}_{q q}}))}^{22} P P (({U u}_{p p},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))}{\underset{l l}{Σ Σ} P P (({U u}_{p p},, {E E.}_{q q} | | {u u}_{((l l))},, {e e}_{((l l))},, {r r}_{((l l))}))} . .$

其中l表示历史训练样本的编号，i,p表示用户隐类的标号，j,q表示服务情境隐类的编号，u,e表示单个用户和单则服务情境，U,E表示用户隐类和服务情境隐类，k表示QoS指标的编号。Among them, l represents the number of historical training samples, i, p represent the label of user hidden class, j, q represent the number of service situation hidden class, u, e represent a single user and single service situation, U, E represent user hidden class and Implicit class of service context, k represents the number of the QoS indicator.

（4）检查模型参数是否收敛，若收敛，结束并保存模型参数，若不收敛，返回（2）执行。(4) Check whether the model parameters are converged, if converged, end and save the model parameters, if not converged, return to (2) to execute.

所述步骤6中利用已训练的隐语义概率模型预测用户指标偏好的实现方式如下：In the step 6, the implementation of using the trained latent semantic probability model to predict user index preference is as follows:

（1）用户u^t登陆服务推荐系统，提供需要进行服务推荐的服务情境e^t；(1) The user u ^t logs in to the service recommendation system and provides the service situation e ^t that requires service recommendation;

（2）推荐系统获取用户的个人数据及其提供的服务情境信息，用下面的公式对该用户在该服务情境下的各个QoS指标偏好进行独立预测(2) The recommendation system obtains the user's personal data and the service context information it provides, and uses the following formula to independently predict the user's preference for each QoS indicator in the service context

${R R}_{k k} (({e e}^{t t},, {u u}^{t t})) = = {&Integral; &Integral;}_{00}^{11} {r r}_{k k} \frac{P P (({e e}^{t t},, {u u}^{t t},, {r r}_{k k}))}{{&Integral; &Integral;}_{00}^{11} P P (({e e}^{t t},, {u u}^{t t},, {r r}_{k k})) d d {r r}_{k k}} d d {r r}_{k k}$

其中in

$P P (({e e}^{t t},, {u u}^{t t},, {r r}_{k k})) = = \underset{i i,, j j}{Σ Σ} P P (({U u}_{i i})) P P (({E E.}_{j j})) P P (({u u}^{t t} | | {U u}_{i i})) P P (({e e}^{t t} | | {E E.}_{j j})) P P (({r r}_{k k} | | {U u}_{i i},, {E E.}_{j j}))$

$P P (({r r}_{k k} | | {U u}_{i i},, {E E.}_{j j})) = = \frac{11}{\sqrt{22 π π} {σ σ}_{k k,, {U u}_{i i},, {E E.}_{j j}}} exp exp [[- - \frac{{(({r r}_{k k} - - {μ μ}_{k k,, {U u}_{i i},, {E E.}_{j j}}))}^{22}}{22 {σ σ}^{22}_{k k,, {U u}_{i i},, {E E.}_{j j}}}]] . .$

最后预测得到用户u^t在该服务情境e^t下对各个QoS指标的偏好权重为 $(R_{1} (e^{t}, u^{t}), . . ., R_{K} (e^{t}, u^{t})),$ 简单记作 $({\hat{r}}_{1}, {\hat{r}}_{2}, . . ., {\hat{r}}_{K}) .$ Finally, it is predicted that the preference weight of user u ^t for each QoS indicator in the service situation e ^t is $(R_{1} (e^{t}, u^{t}), . . ., R_{K} (e^{t}, u^{t})),$ simply write $({\hat{r}}_{1}, {\hat{r}}_{2}, . . ., {\hat{r}}_{K}) .$

所述步骤7中根据预测出的用户个性化的QoS指标偏好，对候选服务进行筛选的过程如下：用户u^t在服务情境e^t下已得到多个功能相似的候选服务，假设每个服务的各QoS指标上的性能可有其他方法得到，记为(q₁,q₂,...,q_K)，那么每一个服务的总得分采用表示，最后根据各个服务的总得分给出服务的排序，作为服务推荐的依据。In the step 7, according to the predicted user's personalized QoS index preference, the process of screening the candidate services is as follows: the user u ^t has obtained multiple candidate services with similar functions in the service context e ^t , assuming that each service's The performance of each QoS index can be obtained by other methods, recorded as (q ₁ ,q ₂ ,...,q _K ), then the total score of each service is Finally, according to the total score of each service, the ranking of services is given, which is used as the basis for service recommendation.

本发明与现有技术相比的优点在于：本发明提供的技术方案建立了一种有关用户、用户指标偏好、服务情境之间的概率依赖模型，用历史数据对模型训练以后，可以利用该模型方便地预测出特定用户在陌生服务情境下的指标偏好向量，最后结合服务的非功能属性和已经预测的用户指标偏好对候选服务进行综合评价以实现个性化服务推荐。已有方法一般只简单地将候选服务的各维度QoS值进行加权平均，据此给出候选服务的综合排序，权值一般是先确定的，显然无法体现用户在服务调用过程中的个性化需求。本专利提出的基于隐语义概率模型的个性化服务推荐方法可以以一种自动的方式预测出用户在特定情境下的个性化指标偏好，即使在用户自己也无法明确表达自己的偏好信息的时候。同时本发明的模型训练可以离线进行，所以提高了在线服务推荐的效率。Compared with the prior art, the present invention has the advantages that: the technical solution provided by the present invention establishes a probability dependence model among users, user index preferences, and service situations. After the model is trained with historical data, the model can be used It is convenient to predict the indicator preference vector of a specific user in an unfamiliar service situation, and finally combine the non-functional attributes of the service and the predicted user indicator preference to comprehensively evaluate the candidate services to realize personalized service recommendation. Existing methods generally simply weight and average the QoS values of each dimension of candidate services, and then give a comprehensive ranking of candidate services. The weights are generally determined first, which obviously cannot reflect the individual needs of users in the process of service invocation. . The personalized service recommendation method based on the latent semantic probability model proposed in this patent can predict the user's personalized index preference in a specific situation in an automatic way, even when the user cannot clearly express his preference information. At the same time, the model training of the present invention can be performed offline, so the efficiency of online service recommendation is improved.

附图说明Description of drawings

图1是本发明系统的功能框图；Fig. 1 is the functional block diagram of the system of the present invention;

图2为图1中历史信息收集模块的实现流程图；Fig. 2 is the realization flowchart of historical information collection module in Fig. 1;

图3为图1中隐语义概率模型参数训练模块的实现流程图；Fig. 3 is the realization flowchart of implicit semantic probability model parameter training module in Fig. 1;

图4为图1中服务推荐请求模块的实现流程图；Fig. 4 is the implementation flowchart of the service recommendation request module in Fig. 1;

图5为图1中个性化指标偏好预测模块的实现流程图；Fig. 5 is the implementation flowchart of the personalized index preference prediction module in Fig. 1;

图6为图1中个性化服务推荐模块的实现流程图；Fig. 6 is the implementation flowchart of the personalized service recommendation module in Fig. 1;

图7为隐语义概率模型的图像化表示。Figure 7 is a graphical representation of the latent semantic probability model.

具体实施方式detailed description

如图1所示，本发明一种基于隐语义概率模型的个性化服务推荐系统及方法由历史信息收集模块、隐语义概率模型参数训练模块、服务推荐请求模块、个性化指标偏好预测模块、个性化服务推荐模块组成。As shown in Figure 1, a personalized service recommendation system and method based on a hidden semantic probability model in the present invention consists of a historical information collection module, a hidden semantic probability model parameter training module, a service recommendation request module, a personalized index preference prediction module, a personality Composed of personalized service recommendation modules.

整个实现过程如下：The whole implementation process is as follows:

所述的服务QoS指标体系是指整个Web服务系统统一采用的用于评价一系列功能相似服务性能优劣所用QoS指标的集合，不同的系统可以根据需要选取适当的QoS指标组成自己的QoS指标体系用于评价服务的性能优劣；The service QoS indicator system refers to the collection of QoS indicators used by the entire Web service system to evaluate the performance of a series of similar services. Different systems can select appropriate QoS indicators to form their own QoS indicator systems. Used to evaluate the performance of the service;

所述的用户指标偏好是指用户对步骤1中所述服务QoS指标体系中各个QoS指标的偏好程度，对每个QoS指标的偏好程度值介于0到1之间，并且对各个QoS指标的偏好值总和为1；所述的服务情境是指在何种场景下使用何种功能的服务，每一个服务情境e用三元组(w₁,w₂,w₃)表示，其中w₁表示服务完成的与业务无关的基本功能，如视频功能、导航功能等，w₂表示用户使用该功能的服务完成的具体业务活动，如学术会议、军事导航等，w₃表示用户调用服务的终端设备，如手机、PC机等；所述的用户、用户指标偏好、服务情境三者之间的隐语义概率模型是指单个用户以不同的概率依赖不同的用户隐类而存在，单则服务情境以不同的概率依赖不同的服务情境隐类存在，指标偏好同时以不同的概率同时依赖不同的用户隐类和服务情境隐类存在的概率模型，模型的图像化表示如图7所示。图中U代表用户隐类，E代表服务情境隐类，u、e分别代表单个用户和打个服务情境，r代表指标偏好向量，图7显示了这几个要素之间的概率依赖关系，用户u、服务情境e、指标偏好向量r对用户隐类U和服务情境隐类E的依赖度的大小以条件概率P(u|U)、P(e|E)、P(r|U,E)的大小来衡量。The user index preference refers to the user's degree of preference for each QoS index in the service QoS index system described in step 1, the preference value for each QoS index is between 0 and 1, and the preference for each QoS index The sum of preference values is 1; the service context refers to the service with which function is used in which scenario, and each service context e is represented by a triplet (w ₁ , w ₂ , w ₃ ), where w ₁ represents The basic functions irrelevant to the business completed by the service, such as video function, navigation function, etc., w ₂ represents the specific business activities completed by the user using the service of this function, such as academic conferences, military navigation, etc., w ₃ represents the terminal equipment that the user calls the service , such as mobile phones, PCs, etc.; the implicit semantic probability model between users, user index preferences, and service contexts means that a single user exists with different probabilities depending on different user hidden categories, and a single service context is based on Different probabilities depend on the existence of different service context hidden classes, and indicator preferences depend on the probability models of different user hidden classes and service context hidden classes at the same time with different probabilities. The graphical representation of the model is shown in Figure 7. In the figure, U represents the user hidden class, E represents the service situation hidden class, u and e represent a single user and a service situation respectively, and r represents the index preference vector. Figure 7 shows the probability dependence relationship between these elements. User The dependence of u, service context e, and index preference vector r on user implicit class U and service context implicit class E is determined by the conditional probability P(u|U), P(e|E), P(r|U,E ) to measure the size.

所述的隐语义概率模型的参数是指所有用户隐类{U₁,U₂,...,U_I}的先验概率P(U_i)(1≤i≤I)、服务情境隐类{E₁,E₂,...,E_J}先验概率P(E_j)(1≤j≤J)、给定一个用户隐类U_i的情况下单个用户u出现的条件概率P(u|U_i)(1≤i≤I)、给定一个服务情境隐类E_j的情况下单则服务情境e出现的条件概率P(e|E_j)(1≤j≤J)、给定用户隐类U_i和服务情境隐类E_j的情况下用户指标偏好向量r出现的概率P(r|U_i,E_j)(1≤i≤I,1≤j≤J)；i,j分别代表什么,其中I，J分别表示用户隐类和服务情境隐类的总个数，i，j分别表示用户隐类和服务情境隐类的编号；The parameters of the hidden semantic probability model refer to the prior probability P(U _i )(1≤i≤I) of all user hidden classes {U ₁ , U ₂ ,..., U _I }, service situation hidden class {E ₁ ,E ₂ ,...,E _J } prior probability P(E _j )( _1≤j≤J ), conditional probability P( u|U _i )(1≤i≤I), the conditional probability P(e|E _j )(1≤j≤J) of a single service situation e in the case of a given service situation hidden class E _j , given The probability of user index preference vector r appearing in the case of given user hidden class U _i and service situation hidden class E _j P(r|U _i , E _j )(1≤i≤I, 1≤j≤J); i, What do j stand for respectively, where I and J respectively represent the total number of user hidden classes and service situation hidden classes, and i and j represent the numbers of user hidden classes and service situation hidden classes respectively;

上述各模块的具体实现过程如下：The specific implementation process of the above modules is as follows:

1.历史信息收集模块1. Historical information collection module

实现过程如图2所示：收集用户的个人信息，包括姓名、职业、年收入、兴趣爱好等，为每个用户分配一个编号；收集用户提供指标偏好所处的服务情境信息，服务情境表示为三元组（服务基本功能，目标活动，终端设备），比如（导航功能，日常出行，手机）或者（语音功能，学术会议，个人电脑）等等。假设实际确定的各个QoS维度为：可靠性、响应时间、可用性、吞吐量、价格，那么用户在该服务情境下给出的比较矩阵就为5*5的矩阵C，计算其最大特征值对应的特征向量并归一化，即为指标偏好，将用户，服务情境，指标偏好三元组存入数据库。The implementation process is shown in Figure 2: collect personal information of users, including name, occupation, annual income, hobbies, etc., and assign a number to each user; collect service context information where users provide index preferences, and the service context is expressed as The triplet (basic service function, target activity, terminal device), such as (navigation function, daily travel, mobile phone) or (voice function, academic conference, personal computer) and so on. Assuming that the actual determined QoS dimensions are: reliability, response time, availability, throughput, and price, then the comparison matrix given by the user in this service scenario is a 5*5 matrix C, and the corresponding maximum eigenvalue is calculated The eigenvectors are normalized, which is the indicator preference, and the user, service situation, and indicator preference triplet are stored in the database.

2、隐语义概率模型参数训练模块2. Latent semantic probability model parameter training module

实现过程如图3所示：将数据库中的历史训练数据取出，然后人工设置用户隐类和服务隐类的个数，假设都为5个，以及隐语义概率模型的初始参数：5个用户隐类的初始先验概率，可均设为1/5；5个服务情境隐类的概率，可均设为1/5；每一个用户在给定不同用户隐类的条件概率初值；每一则服务情境在给定不同服务情境隐类的条件概率初值；每一个指标偏好在给定不同用户隐类和服务情境隐类时的正态分布均值和方差的初值，然后用EM算法结合训练数据迭代训练模型的参数直至收敛并将参数保存。The implementation process is shown in Figure 3: take out the historical training data in the database, and then manually set the number of user hidden classes and service hidden classes, assuming that they are both 5, and the initial parameters of the hidden semantic probability model: 5 user hidden classes The initial prior probability of each class can be set to 1/5; the probabilities of the five service context hidden classes can all be set to 1/5; each user is given the initial value of the conditional probability of different user hidden classes; each Then the initial value of the conditional probability of the service situation given different service situation hidden classes; the initial value of the normal distribution mean and variance of each index preference when given different user hidden classes and service situation hidden classes, and then combined with EM algorithm The training data iterates the parameters of the training model until convergence and saves the parameters.

3.服务推荐请求模块3. Service recommendation request module

实现过程如图4所示，用户登录服务推荐系统，假设该用户已经是系统的注册用户，并且有特定的登录名标识，记为u^t，他提供了需要进行服务推荐的服务情境记为e^t，具体含义是（汇率转换，网络购物，个人电脑），该模块会记录下u^t和e^t，提交给个性化指标偏好预测模块和个性化服务推荐模块。The implementation process is shown in Figure 4. The user logs in to the service recommendation system. Assume that the user is already a registered user of the system and has a specific login name, denoted as u ^t . ^t , the specific meaning is (exchange rate conversion, online shopping, personal computer), this module will record u ^t and e ^t , and submit them to the personalized index preference prediction module and the personalized service recommendation module.

4.个性化指标偏好预测模块，实现过程如图5所示。4. Personalized index preference prediction module, the implementation process is shown in Figure 5.

该模块需要获取已训练好的模型参数和用户提供的服务推荐请求信息，基于上面提出的指标偏好预测方法预测出用户指标偏好，假设预测的指标偏好为r=(0.3,0.1,0.2,0.3,0.1)。This module needs to obtain the trained model parameters and the service recommendation request information provided by the user, and predict the user's index preference based on the index preference prediction method proposed above, assuming that the predicted index preference is r=(0.3,0.1,0.2,0.3, 0.1).

5.个性化服务推荐模块5. Personalized service recommendation module

实现过程如图6所示，假设满足汇率转换的服务共有3个s1，s2，s3，它们在五个QoS维度上的归一化值可以由其他方式获得，假设分别记为：q1=(0.5,0.5,0.6,0.7,0.6),q2=(0.8,0.4,0.7,0.2,0.6)，q3=(0.4,0.8,0.2,0.6,0.4)，那么三个候选服务的得分分别为：The implementation process is shown in Figure 6. Assume that there are three s1, s2, and s3 services that satisfy exchange rate conversion, and their normalized values on the five QoS dimensions can be obtained in other ways, assuming that they are respectively recorded as: q1=(0.5 ,0.5,0.6,0.7,0.6),q2=(0.8,0.4,0.7,0.2,0.6),q3=(0.4,0.8,0.2,0.6,0.4), then the scores of the three candidate services are:

g1=r·q1=(0.3,0.1,0.2,0.3,0.1)·(0.5,0.5,0.6,0.7,0.6)=0.59；g1=r·q1=(0.3,0.1,0.2,0.3,0.1)·(0.5,0.5,0.6,0.7,0.6)=0.59;

g2=r·q2=(0.3,0.1,0.2,0.3,0.1)·(0.8,0.4,0.7,0.2,0.6)=0.54；g2=r·q2=(0.3,0.1,0.2,0.3,0.1)·(0.8,0.4,0.7,0.2,0.6)=0.54;

g3=r·q3=(0.3,0.1,0.2,0.3,0.1)·(0.4,0.8,0.2,0.6,0.4)=0.46；g3=r·q3=(0.3,0.1,0.2,0.3,0.1)·(0.4,0.8,0.2,0.6,0.4)=0.46;

所以可得g1>g2>g3,所以给出了服务排序列表为s1，s2，s3，返回给用户供用户选择。Therefore, g1>g2>g3 can be obtained, so the service sorting list is given as s1, s2, s3, and returned to the user for the user to choose.

本发明未详细阐述部分属于本领域公知技术。Parts not described in detail in the present invention belong to the well-known technology in the art.

以上所述，仅为本发明部分具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本领域的人员在本发明揭露的技术范围内，可轻易想到的变化或替换，都应涵盖在本发明的保护范围之内。The above are only some specific implementations of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A personalized service recommendation system based on hidden semantic probability model, characterized in that it comprises:

Historical information collection module: collect the index preference comparison matrix independently provided by different users through the AHP in different service scenarios, check that the matrix meets the consistency constraints, calculate the eigenvector corresponding to the largest eigenvalue of the matrix, and obtain Find out the indicator preferences of different users in different service scenarios, and store users, service scenarios, and indicator preferences in the database in the form of triples, and provide historical training data for the implicit semantic probability model parameter training module;

Implicit Semantic Probability Model Parameter Training Module: Obtain triplets of users, service situations, and index preferences collected by the historical information collection module from the database as training samples, and use EM algorithm to train Iteratively train model parameters on the data until the model parameters converge, save the obtained training parameters, and provide them to the personalized index preference prediction module;

User service recommendation request module: This module obtains the personal data information provided by the user when logging in and the service situation information that needs to be recommended for service input after the user logs in, and finally provides these two pieces of information to the personalized index preference prediction module;

Personalized index preference prediction module: Obtain the user's personal information and service context information provided by the user service recommendation request module, and combine the trained model parameters provided by the implicit semantic probability model parameter training module to predict the user's index in the service context preferences, and provide the predicted index preferences to the personalized service recommendation module;

Personalized service recommendation module: Receive user indicator preferences from the personalized indicator preference module, and service context information from the user service recommendation request module, and obtain all services with similar functions applicable to the service context, as well as QoS indicators of each dimension of the service Value, dot product with the user's personalized index preference, give the comprehensive ranking of the service, return it to the user as a recommendation list, and the user chooses the item he is interested in according to the recommendation result to pay attention to.

2. the personalized service recommendation system based on hidden semantic probability model according to claim 1, is characterized in that: described hidden semantic probability model parameter training module realization process:

(1) Obtain the user, service situation, and index preference triplet stored in the historical information collection module from the database as historical training data;

(2) Manually set the number of user hidden classes and service situation hidden classes and the initial value of the model parameters, including the prior probability of user hidden classes and service situation hidden classes, and the conditional probability of a single user when different user hidden classes are given , the conditional probability of a single service situation when given different service situation hidden classes, and the normal distribution mean and variance of each index preference when given different user hidden classes and different service situation hidden classes;

(3) Using the EM algorithm, combined with historical training data, iteratively trains the model parameters until convergence, saves the model parameters, and provides them to the personalized index preference prediction module.

3. A personalized service recommendation method based on hidden semantic probability model, characterized in that the implementation steps are as follows:

Step 1. Determine the service QoS index system for evaluating service performance

The service QoS indicator system refers to the collection of QoS indicators used by the entire Web service system to evaluate the performance of a series of similar services. Different systems can select appropriate QoS indicators to form their own QoS indicator systems. Used to evaluate the performance of the service;

Step 2. Establish an implicit semantic probability model among users, user index preferences, and service contexts

The user index preference refers to the user's degree of preference for each QoS index in the service QoS index system described in step 1, the preference value for each QoS index is between 0 and 1, and the preference for each QoS index The sum of preference values is 1; the service context refers to the service with which function is used in which scenario, and each service context e is represented by a triplet (w ₁ , w ₂ , w ₃ ), where w ₁ represents Basic functions irrelevant to business completed by the service, including video function and navigation function, w ₂ indicates the specific business activities completed by the user using the service of this function, w ₃ indicates the terminal device that the user calls the service;

The implicit semantic probability model between the user, user index preference, and service context means that a single user depends on different user hidden classes with different probabilities, and a single service context depends on different service contexts with different probabilities. Hidden classes exist, and indicator preferences depend on different probabilities at the same time. The probability model of the existence of different user hidden classes and service situation hidden classes. The user hidden classes are not determined in advance by humans but learned from historical experience data. The user clustering; the service context implicit category refers to the service context clustering that is not determined in advance by humans but learned from historical experience data;

Step 3. Collect the index preference information provided by different users when using different functional services in different service scenarios, and store them in the database as historical experience data to prepare for training the parameters of the latent semantic probability model established in step 2. The storage format It is a triplet of user, service context, and indicator preference;

Step 4. Use the EM algorithm and the collected historical experience data to train the parameters of the implicit semantic probability model

The parameters of the hidden semantic probability model refer to all user hidden classes The prior probability of service context implicit class Priori probability Given a user hidden class The conditional probability of a single user u appearing in the case Given a service context implicit class The conditional probability of occurrence of a single service situation e in the case of Given user implicit class and service context implicit class The probability that the user index preference vector r appears in the case Among them, I and J respectively represent the total number of user hidden classes and service situation hidden classes, and i and j respectively represent the numbers of user hidden classes and service situation hidden classes;

Step 5. Obtain the user's request for service recommendation, including personal information and the service context on which the service recommendation depends;

Step 6. Use the trained latent semantic probability model to predict the unknown index preference of the specified user in a specific service situation;

Step 7. According to the predicted user's personalized QoS index preference, comprehensively screen the candidate services, so as to select the service that is closest to the user's needs, and return the recommendation result to the user.

4. The personalized service recommendation method based on hidden semantic probability model according to claim 3, characterized in that: the process of collecting historical experience data in the step 3 is as follows:

(1) When the user calls the familiar web service in the familiar service situation, he directly interacts with the system, uses the AHP to give the pairwise comparison results of his preferences for each indicator, and establishes a comparison matrix;

C C = = [\begin{matrix} {c c}_{1111},, & {c c}_{1212},, & ... ...,, & {c c}_{11 n no} \\ {c c}_{21 twenty one},, & {c c}_{22 twenty two},, & ... ...,, & {c c}_{33 n no} \\ . . & . . & . . \\ . . & . . & . . \\ . . & . . & . . \\ {c c}_{n no 11},, & {c c}_{n no 22},, & ... ...,, & {c c}_{n no n no} \end{matrix}],,

(2) Verify whether the consistency of the comparison matrix is within an acceptable range, if acceptable, enter step (3), otherwise return to step (1);

(3) Calculate the eigenvector corresponding to the largest eigenvalue of the comparison matrix. After normalizing the eigenvector, the user’s preference vector W _QoS for each QoS indicator is obtained. Store the user, service situation, and indicator preference triplet into the database.

5. the method for recommending personalized services based on the hidden semantic probability model according to claim 3, characterized in that: the parameter training process of the hidden semantic probability model in the step 4 is as follows,

(1) Take out all triples of user, service context, and index preference from the database as training data, which is the personalized index preference weight given by the user in the familiar service context using the AHP. Each triple The group (u,e,r) means that the user u's preference weight for each QoS indicator of the service in the service situation e is r, and r is a K-dimensional vector (r ₁ ,r ₂ ,...,r _K ), which is For simplicity, assuming that each dimension of the weight vector r is independent of each other and obeys a normal distribution, the joint probability P(u,e,r) is expressed by the following formula,

in It is a K-dimensional normal distribution, i and j respectively represent the number of hidden users and service situations, and K is the number of indicators in the QoS indicator system;

(2) Step E calculates the joint posterior probability of the hidden class,

where b is a simulated annealing parameter between 0 and 1, Satisfy the K-dimensional normal distribution, i, p represent the hidden class number of the user, j, q represent the hidden class number of the service situation;

(3) Step M uses the joint posterior probability of user hidden class and service situation hidden class calculated in E step to re-estimate model parameters, including the prior probability of each user hidden class and each service situation hidden class and The conditional probability of each user appearing given different user hidden classes The conditional probability of each service situation when given different service situation hidden classes The mean value of the normal distribution of K indicator preferences when different user contextual implicit classes and service contextual implicit classes are given at the same time and variance

Among them, l represents the number of historical training samples, L represents the total number of training samples, i, p represents the label of user hidden class, j, q represents the number of service situation hidden class, u, e represent a single user and a single service situation, Respectively represent user hidden class and service situation hidden class, k represents the number of QoS index;

(4) Check whether the model parameters are converged, if converged, end and save the model parameters, if not converged, return to (2) for execution.

6. The personalized service recommendation method based on the hidden semantic probability model according to claim 5, characterized in that: in step 6, the implementation of using the trained hidden semantic probability model to predict user index preference is as follows:

(1) The user u ^t logs in to the service recommendation system, and provides the service situation e ^t that requires service recommendation;

(2) The recommendation system obtains the user's personal data and the service context information it provides, and uses the following formula to independently predict each QoS index preference of the user in the service context:

{R R}_{k k} (({e e}^{t t},, {u u}^{t t})) = = {&Integral; &Integral;}_{00}^{11} {r r}_{k k} \frac{P P (({e e}^{t t},, {u u}^{t t},, {r r}_{k k}))}{{&Integral; &Integral;}_{00}^{11} P P (({e e}^{t t},, {u u}^{t t},, {r r}_{k k})) {dr dr}_{k k}} {dr dr}_{k k};;

in:

Among them, u ^t and e ^t respectively represent the users who are preferred by the index to be predicted and the service context they are in, k represents the number of the QoS index, R _k (e ^t , ^u ^t ) represents the user u ^t ’s preference for the first Predicted values for k indicator preferences, Represent user hidden class and service situation hidden class respectively, i represents the label of user hidden class, j represents the number of service situation hidden class, and Respectively represent the implicit class of given user context and service context implicit class The normal distribution mean of the k-th index preference and variance Finally, it is predicted that user u ^t ’s preference weight for each QoS indicator in the service situation e ^t is (R ₁ (e ^t ,u ^t ),...,R _K (e ^t ,u ^t )), denoted as

7. The personalized service recommendation method based on the hidden semantic probability model according to claim 6, characterized in that: in step 7, according to the user's personalized QoS index preference predicted, the process of screening candidate services is as follows: u ^t has obtained multiple candidate services with similar functions in the service context e ^t , assuming that the performance of each QoS index of each service can be obtained by other methods, recorded as (q ₁ ,q ₂ ,...,q _K ), the total score of each service is Finally, according to the total score of each service, the ranking of services is given, which is used as the basis for service recommendation.