JPH0922405A - Non-linear model generation method - Google Patents

Abstract

(57) [Abstract] [Purpose] The objective is to build a non-linear model from a large amount of data using a client and a server, and to shorten the construction period of a non-linear model from a large amount of data. [Structure] A non-linear model generation method for creating a non-linear model from original data, which is data having two or more items and two or more records, and performing a feature preservation compression process (102) for compressing the original data. , The compressed data is created by the server on the network, the data transfer process (103) of transferring the created compressed data from the server to the client is performed, and the compressed data is used to be predetermined by the client. The model generation parameters for creating the nonlinear model (104 to 107),
A model generation process (109) of generating a non-linear model from the original data is performed by the server using the adjusted model generation parameter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of using a client-server system to generate a non-linear model from a collection of information represented by numerical values or symbols stored in a database held by a server.

[0002]

2. Description of the Related Art The amount of data to be stored is increasing with the increase in speed of computers and the increase in capacity of disks. Also, with the progress of networking, database utilization technology from clients is also advancing.

The information that a user can directly obtain from a large amount of data is limited. By setting conditions and extracting only the data of interest, or by constructing a model such as a neuro model and utilizing it for prediction and diagnosis, information buried in the data can be acquired. The following three are known as typical methods for constructing a nonlinear model from a large amount of data.

(1) Method using a neuro model A method for constructing a neuro model is described in many documents. For example, there is an explanation in the literature of learning control of an inverted pendulum by neural network driven fuzzy inference by Isao Hayashi et al. On pages 183 to 188 of the 5th Fuzzy System Symposium Proceedings. (Hereinafter, referred to as Prior Art 1) (2) Method of using rule extraction algorithm A rule extraction system (RI: Rule Induction System) is a method of extracting a category data represented by a symbol by using a predetermined rule extraction algorithm. Extract the characteristics of the data,
It is a system that expresses in a rule format.

For a typical rule extraction algorithm ID3, see the Tiga Publishing Comp.
Learning by JR Quinlan, on any of Machine Learning, pages 463-482.
Efficient Classification Process and Their Applic
See ation to Chess End Games for more details.
(Hereinafter, referred to as Prior Art 2) (3) Method of using statistical model A statistical model such as a multiple regression model is widely known in general and will not be described in detail here. (Hereinafter, Prior Art 3
In addition, as a database utilization technology considering the network, for example, Nikkei Computer July 12, 1993, entitled "Direct connection of server RDB-basic data to spreadsheet software"
It is described on pages 67 to 75 of the Japanese issue (hereinafter referred to as prior art 4). In this paper, a server holding data and a client used by a user are connected by a network. When a search request is issued from a client using spreadsheet software, SQL (Structured Query Langu
age) and other search commands are sent to the server, the server executes the search, and the search results are transferred to the client. The client formats and displays the search results in spreadsheet format.

[0006]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
The conventional nonlinear model generation methods described in 2 and 3 relate to a model generation algorithm, and are not conscious of the client and the server as described in Related Art 4. Therefore, the data must be on the device that performs the process. Or, since it has to be collected on the device that executes the processing, there is a problem that the network load is applied.

Since a large amount of data exceeding 100,000 cases is not assumed, it takes a considerable amount of time to generate a model when a large amount of data is handled, because trial and error is required for parameter setting. Occurs.

The database technology described in the above-mentioned prior art 4 utilizes a network to search a large amount of data from a client at high speed, but the processing to be executed is limited to the search, such as rule generation. The model generation process of is not assumed.

In view of the problems of the above prior arts 1, 2, 3, and 4, the first object of the present invention is to construct a non-linear model generation method using a client and a server.

A second object is to construct a non-linear model generation method capable of rapidly constructing a non-linear model from a large amount of data of over 100,000 cases.

[0011]

In order to achieve the above object,
In the non-linear model generation method of the present invention, the compressed data (compressed data) is transferred from the server to the client, and the compressed data is used to adjust the model generation parameter at the client. Using the model generation process to generate a model from the original data on the server and the model characteristic information transfer process to transfer the model characteristic information indicating the input / output relationship of the generated model including the compressed data to the client, the client transfer Model characteristic information display processing for displaying the obtained information is provided.

According to a preferred aspect of the present invention, in the data compression processing, the distance between records is measured, and if the distance is smaller than a preset threshold value, the feature preservation compression processing represented by one point is performed. The feature preservation compression parameter designating process for designating a compression parameter includes a process of selecting an item to be used for feature preservation compression, a process of setting the threshold value, and a feature preservation compression feature applied according to the threshold value. Check the total number of records of saved compressed data,
It has a process of resetting or understanding the threshold value.

The feature preservation compression process is characterized in that the number of records represented by the data selected by the compression process is set as the weight of the record.

The process of setting the model generation parameter includes an item selection process of selecting an item used for model generation, an item generation process of generating a new item, and a categorization process of converting numerical data into category data. , And a process of setting a unique parameter for each type of model to be generated.

More specifically, in the categorization processing, for each of the items, processing for displaying a histogram of the original data approximately from compressed data and operation of a pointer displayed overlaid on the displayed histogram are operated. Input processing for dividing the domain of the item at a predetermined position by using the pointer, dividing the domain into several sections, assigning a category to each section, and including the numerical data in the numerical data. And a process of converting into a category corresponding to the section.

The model characteristic information display process is performed by graphically displaying the data subjected to the characteristic preservation and compression in the server and the generated model, and designating the data by using the pointer displayed on the graphic display. ,
It is characterized in that it is provided with processing for displaying the values of items other than the items displayed graphically of the data.

[0017]

In order to achieve the above-mentioned first object, the following processing is executed.

First, the data stored in the server is compressed according to the parameters set by the client (processes 101 and 102 (FIG. 1)). Here, as a compression method, a feature-preserving compression process in which a distance between data is measured and when the distance is smaller than a preset threshold value is collectively represented by one point can be used. When the feature save compression save process is used, the process of selecting an item used for the feature save compression (process 202 (FIG. 2)) can speed up the compression without lowering the compression accuracy. Further, by checking the number of records of the compressed data at the client (process 204 (FIG. 1)), unnecessary compression process can be prevented.

Then, the data transfer process transfers the compressed data from the server to the client (process 10).
3 (Fig. 1)).

The client uses the transferred data to perform a preliminary experiment for determining parameters for model generation (processes 104, 105, 106, 107 (FIG. 1)).

After the parameters are determined, the parameters are transferred to the server (process 108 (FIG. 108), and the server
Model generation processing is performed using the transferred parameters (processing 109).

The model characteristic information representing the input / output relationship of the generated model, which includes the characteristic-preserved and compressed data (feature-preserving compression processing 111), is transferred to the client (processing 112 (FIG. 1)).

At the client, the transferred model characteristic information is graphically displayed, the model generated by the user is determined, and if the model is not satisfied, the model generation parameter is reset (process 115 (FIG. 1)) and model generation is performed. repeat.

By the above processing, it is possible to generate a non-linear model using a client and a server.

In order to achieve the above-mentioned second object, the following processing is executed.

When a model is generated from a large amount of data, 1
Since the time required to generate a model once increases in proportion to the amount of data, it takes an enormous amount of time to set parameters by trial and error using all large amounts of data. Therefore, in the present invention, data of a size that allows trial and error is extracted for parameter setting by compression processing, and model generation is repeated (processing 104, 105, 106).

As a result, the time required for parameter setting can be greatly reduced.

Further, when the feature preservation compression is used as the compression process (process 102 (FIG. 1)), data holding the features of the original data before compression can be obtained as preliminary experiment data, so that the parameter setting with high accuracy can be performed. Can be expected.

Further, when the input / output relationship of the model is graphically displayed, the compressed data is transferred and displayed, so that the time required for the data display can be greatly reduced.

From the above processing, according to the present invention, the model development period from a large amount of data is significantly shortened,
It becomes practical and feasible.

[0031]

[First Embodiment] First, an outline of a feature preservation compression method will be described.
Characteristic preservation compression technology is Y. Linde, A. Buzo and RMGray:
"An Algorithm for Vector Quantization", IEEE Tran
As described in s.Commun., COM-28,1, pp.84-95 (Jan.1980), when compressing a database while preserving features, consider records as multidimensional vectors and This can be achieved by dividing the space into predetermined regions, setting the initial value of the quantized representative vector for each region, and converging using the LBG algorithm.

Embodiments of the present invention will be described below with reference to the drawings. In this embodiment, a rule model is used as the non-linear model.

FIG. 1 is a flowchart of the nonlinear model generation method which is the object of the present invention. Using this method, one or more inputs and a unique output are selected from the items of data for which the non-linear model is to be generated, and a non-linear model representing the input / output relationship between them is generated.

In FIG. 1, 101 is a characteristic preservation compression parameter designation process, 102 is a characteristic preservation compression process, 103 is a data transfer process, 104 is a model generation parameter setting process, 105 is a model generation process, 106 is a model characteristic information display process, 107 is a model determination process, 108 is a model generation parameter transfer process, 109 is a model generation process, 11
0 is a feature preservation compression parameter designation process, 111 is a feature preservation compression process, 112 is a model characteristic information transfer process, 113
Is model characteristic information display processing, 114 is model determination processing,
Reference numeral 115 is a model generation parameter setting process.

In the present embodiment, a part of the data held by the server is compressed by feature preservation compression, the compressed data is transferred to the client, the model generation parameters are determined, and the determined parameters are transferred to the server. Then, the server generates a model using all the data, transfers the generated model and the newly compressed data to the client, and the client evaluates the model.

Each process of FIG. 1 will be described in detail.

The details of the feature preservation compression parameter designation processing 101 are shown in FIG. As shown in FIG. 2, the feature preservation compression parameter designation process 101 includes an item reading process 201, a feature preservation compression use item designation process 202, a feature preservation compression threshold value designation process 203, and a record determination process 204.

In the item reading process 201, the client instructs the server to extract all the item names in the data.
The feature save compression use item designation process 202 selects an item to be used for feature save compression from all item names and transfers it to the server. The characteristic preservation compression threshold value specifying process 203 sets the threshold value of the characteristic preservation compression at the client and transfers it to the server. In FIG. 2, the feature preservation compression use parameter name transferred from the client to the server and the feature preservation compression threshold value are the feature preservation compression parameter information in FIG. If the user is satisfied with the number of records after compression, the record determination processing 204 instructs the server to transfer the data, and if not satisfied, the processing returns to the feature preservation compression threshold value designation processing 203.

In the model generation parameter setting process 104, parameters for model generation are set. When generating a rule model, b) select items used for model generation c) generate new items (ex. Moving average) d) convert numerical data to category data (rule model only) e) number of generated rules f) Rule generation method g) Set the generality parameter.

In the model generation processing 105, the compressed data and the model generation parameter are input, and the rule model and the feature-stored compressed data used for generating the rule model are output.

In the model characteristic information display processing 106, the compressed data and the rule model are displayed graphically. Also, the rule sentence is displayed.

In step 107, it is determined whether the model generated by the user is satisfied based on the rule statement, the versatility of the rule, the accuracy of the rule, and the graphic display of the rule and data. If satisfied, model generation parameter transfer process 1
If not satisfied, the procedure returns to model generation parameter setting processing 104.

In the model generation parameter transfer process 108, the model generation parameter set in the model generation parameter setting process 104 is transferred from the client to the server.

The model generation processing 109 generates a rule model from the original data of the server according to the set parameters and outputs the generated rule model.

The model characteristic information transfer processing 112 transfers the model characteristic information representing the characteristic of the model and the characteristic-preserved compressed data which has been subjected to the characteristic-preserving compression to the client.

The operation of the above non-linear model automatic generation method will be described using the problem of selecting a direct mail delivery area of a department store. Every year, a department store sends direct mail recommending a "membership card" to residents in Tokyo.
Average response rate is not bad. Therefore, using the past data, characteristics of regions with good response rates and regions with poor response rates are extracted in a rule format, which is useful for planning future direct mail delivery strategies.

In the format of FIG. 3, tens of thousands of pieces of data representing the characteristics of each business area independently defined by the department store are stored in the host machine. Since this host is also used for other tasks, the sales person who creates the model accesses this host from the client machine and extracts the characteristics of the areas with good or bad response to direct mail in the rule format.

First, the item used for the feature preservation compression is selected. Items not used for model generation need not be used for feature preservation compression. Here, I decided to omit the item children (the number of children) from the experience of the person in charge. Then, the threshold for feature preservation compression is set. The threshold value is a real number of 0 or more and 1 or less. When calculating the distance between data, all items are normalized to a value of 0 or more and 1 or less. As for the category data, by default, the distance is 0 when they match and the distance 1 when they do not match. If the user grasps the order relation between categories, the position of each category may be defined within the range of 0 or more and 1 or less. When the threshold value is 0, the data is regarded as the data having the same content and compressed only when the values of all the usage items match, and when the threshold value is 1, all the data are determined as the data having the same content.

When the threshold value is input, the server performs the feature preservation compression and transfers the number of data after compression to the client. In this case, a threshold value of 0.2
Then, I was able to reduce it to about 5,000, so the person in charge was satisfied,
{symbol 148 \ f "Times NewRoman" |} OK {symbol 148 \ f
Enter "Times New Roman" |}. Then, about 5,000 pieces of compressed data are transferred from the server to the client. The number of records represented by each record is added to the compressed data as a weight. The data after the feature preservation compression is shown in FIG.

At the client, pre-processing for rule model generation is performed. First is the generation of new items used for model generation. For example, define a new item, which is the average income obtained by dividing annual income by dependents + 1. Next, the items used for model generation are selected. Here, all items are selected. Next, all numerical data are converted into category data. The division number is 3 for input items, 2 for output items, and small, medium, and large for input items, and low and high for output variables. FIG. 5 shows an example of the histogram. A rule model is generated by setting the number of generated rules and the measure of generality.
The user changes the input item, the number of divisions, the division position, the number of generation rules, and the measure of generality until the user is satisfied with the generated model. The rules satisfied by the user are shown in FIG. This is the preliminary experiment. In the preliminary experiment, trial and error are repeated as described above, but if the number of data is 5,000 and there are about 20 items, even a general personal computer takes several seconds to generate a new item and 1 to generate a rule. Since it only takes 2 minutes, trial and error can be sufficiently repeated.

When the generated model is satisfied, the set input / output items, the categorization division position and category name, the number of generation rules, and the model generation parameter information of the measure of generality are transferred to the host computer. The host computer executes rule generation on the original data before compression.
The generated model is transferred to the client as model characteristic information in which the rule statement information shown in FIG. 7 and the categorization information shown in FIG. 8 are combined.

When the rules are displayed graphically, they can be displayed simultaneously with the data.

When data is displayed on a three-dimensional graphic, even if all 100,000 pieces of data are displayed at the same time, the data cannot be identified. Therefore, feature preservation compression is performed to reduce the apparent number of records. Similar to the preliminary experiment, first input the item name to be used for feature storage compression from the client.
The default is all the item names used in the model.
Here, the default items are used. Next, the threshold for feature preservation compression is input. In order to identify on the graphic display, it is compressed to about 1000 records. The compressed record and rule model information are transferred to the client as model characteristic information. If desired by the user, it is possible to transfer only the information of the rule model. The client displays the transferred model characteristic information. FIG. 9 shows an example in which the rule model and the data are displayed simultaneously. The transferred rule sentence information and categorization information are at most 10 kbytes. Compressed data is not graphic information either, so if there are about 1000 records and about 20 items,
At most 10 Mbytes. The advantage of transferring the data before the graphic display is that the number of bytes to be transferred is reduced and the display axis is changed when the graphic is displayed.
With the change of the display angle, it is not necessary to go to the server for each reference, and the display speed can be increased.

If the generated rule is not satisfied, the parameters are reset and the rule generation is repeated. The repetition here uses all data (about 100,000 records), so it takes several tens of minutes for a general workstation.
Since we have completed preliminary experiments, we can generate a satisfactory rule in a few times.

[0055]

According to the nonlinear model generation method of the present invention,
It is possible to set parameters at high speed with the generation of the nonlinear model. This preliminary experiment uses not the data sampled at random but the data subjected to feature preservation compression, so that the preliminary experiment with higher accuracy can be performed as compared with the random sample. In addition, by conducting preliminary experiments on the client, the load on the server and network can be reduced, and the server can be effectively used for other tasks. Also,
When the generated model is displayed, not the image data of the model but the structural data of the model is transferred, so that it is not necessary to access the server when changing the viewpoint or converting the display coordinates. As a result, these functions can be realized at high speed. The image data of the multi-input non-linear model displays one surface of the model in detail and cannot represent all the information at the same time.

Further, when displaying the data used for model generation, not the whole data but the data subjected to the feature preservation compression is transferred, so that the transfer data amount can be reduced and the load on the host and the network can be reduced. When displaying a large amount of data, there is a limit to the number of records that can be identified on the screen. If the data is drawn as a circle with a certain area, the data within a certain range will overlap each other and will be indistinguishable. In the case of data that has been subjected to the feature preservation compression, the number of records is apparently reduced while the features of the data are retained, so that the amount of transferred data is reduced and a visual effect can be obtained.

According to the present invention, by allowing the user to input the items used for the feature storage compression, only the items used by the user for the model generation are designated as the items used for the feature storage compression. , It can speed up without any loss of quality of feature preservation compression. Further, by checking the total number of records of the feature-preserving compressed data after performing the feature-preserving compression by the set threshold value and making it possible to re-set, the useless feature-preserving compression process can be prevented.

According to the present invention, a weighting item is newly set in the feature-preserved compressed data, and the number of records represented by the data is set as the value of the weighting item in each record. It is also possible to approximately reproduce the data before performing the feature preservation compression within the range of and use it for model generation.

According to the present invention, the histogram of the original data can be reproduced and displayed from the feature-preserved compressed data within the range of the threshold error of the feature-preserved compression. Further, the position of the displayed histogram can be designated by the pointer displayed in an overlapping manner.

[Brief description of drawings]

FIG. 1 is a flowchart of a nonlinear model generation method according to an embodiment of the present invention.

FIG. 2 is a flowchart of a feature preservation compression parameter designation process according to the first embodiment.

FIG. 3 is a diagram showing an example of original data in the embodiment.

FIG. 4 is a diagram illustrating an example of feature-preserving compressed data according to an embodiment.

FIG. 5 is a diagram showing an example of a histogram in categorization processing in the embodiment.

FIG. 6 is an example showing an example in which a non-linear model generated in a preliminary experiment in an example is described by a rule statement.

FIG. 7 is a diagram showing an example of rule statement information according to the embodiment.

FIG. 8 is a diagram showing an example of categorization information in the embodiment.

FIG. 9 is a diagram showing an example of a graphic display of rules and data in the embodiment.

[Description of sign]

101 ... Feature save compression parameter designation process, 102 ... Feature save compression process, 103 ... Data transfer process, 104 ... Model generation parameter setting process, 105 ... Model generation process, 106 ... Model characteristic information display process, 107 ... Model determination process, 108 ... Model generation parameter transfer processing, 1
09 ... Model generation processing, 110 ... Feature preservation compression parameter designation processing, 111 ... Feature preservation compression processing, 112 ... Model characteristic information transfer processing, 113 ... Model characteristic information display processing, 114 ... Model determination processing, 115 ... Model generation parameter setting processing

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yori Takahashi 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture Hitachi Ltd. Software Development Division

Claims (4)

[Claims] 1. A non-linear model generation method for creating a non-linear model from original data, which is data having two or more items and two or more records, wherein the original data is compressed to obtain compressed data on a network. A model for creating a non-linear model created by a server, transferring the created compressed data from the server to a client on the network, and using the compressed data to create the non-linear model predetermined by the client A non-linear model generation method comprising adjusting a generation parameter, and using the adjusted model generation parameter, generating a non-linear model from the original data in the server. 2. The non-linear model generation method according to claim 1, wherein model characteristic information representing an input / output relationship of the generated non-linear model is transferred to the client, and the model characteristic information transferred by the client is displayed. A non-linear model generation method characterized by: 3. The non-linear model generation method according to claim 1, wherein the distance between the records is measured, and if the measured distance is smaller than a preset threshold value, a predetermined one point is represented. A method for generating a non-linear model, characterized in that the original data is compressed by the following method. 4. The non-linear model generation method according to claim 1, wherein the number of records represented by the original data is set as a weight of the record, and the original data is compressed based on the set weight. Non-linear model generation method characterized by.