JP3974763B2 - Prediction analysis apparatus and prediction analysis program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention provides a function for presenting how to predict unknown data using known data accumulated in the past and to change the prediction result to a desired result. Predictive analysis device to be realized and predictive analysis program used for realizing the predictive analysis deviceAndRelated.
[0002]
In recent years, with the development of computers and the Internet, various and large amounts of information including remote locations can be easily obtained. In addition, storage of the obtained information has been facilitated by increasing the density and price of the storage device.
[0003]
For example, in a point-of-sale (POS) system in the distribution business, it is possible to collect sales contents of retail stores in various parts of the country in a head office computer or the like, and it is accumulated every moment as a relationship between time and sold products. In addition to this, the conditions of various manufacturing equipment in the manufacturing industry and the yield data of the manufactured products, personal credit card usage status in the financial industry, personal data and usage status information of insurance users in the insurance business, etc. There are many fields where a large amount of information is accumulated.
[0004]
There is an increasing demand for automatically and efficiently extracting valuable information such as causal relationships and rules inherent in such a large amount of accumulated data to be useful for business.
[0005]
[Prior art]
In the past, it has been possible to predict the results of unknown data using statistical processing, AI, neural networks, etc., using the accumulated known data.
[0006]
However, in actual application situations, it is desired not only to simply predict unknown data, but also to indicate what to do next based on the prediction results.
[0007]
For example, in the case of the manufacturing industry, if a product manufactured from the condition data of various manufacturing equipment is predicted to be defective, how to change the manufacturing condition to see if the manufactured product changes to good is important. Also, in the problem of investigating risks from information on insurance users and usage status in the insurance industry, for example, a person who is predicted to be "dangerous" is predicted to be closest to that person's situation and "safe" It is important to seek such conditions.
[0008]
However, the conventional prediction apparatus does not provide such information, but only measures the sensitivity to the attribute result by sensitivity analysis.
[0009]
[Problems to be solved by the invention]
Thus, the conventional prediction apparatus does not provide information describing how and how much unknown data should be changed in order to change the prediction result to a desired prediction value.
[0010]
From now on, according to the prior art, when a prediction result for unknown data is obtained, if the prediction result is not what is desired, the user can change the prediction result to a desired prediction value by: There is a problem that it is not possible to know what to do.
[0011]
The present invention has been made in view of such circumstances, and when predicting unknown data using known data accumulated in the past, in order to change the prediction result to a desired result, The purpose is to provide a new predictive analysis technology that realizes the function of presenting what to do.
[0012]
[Means for Solving the Problems]
FIG. 1 shows a schematic configuration of the present invention.
[0013]
In the figure, reference numeral 1 denotes a predictive analysis apparatus having the present invention, which is used for predicting unknown data using previously stored known data, and for changing the prediction result to a desired result. Is a process that presents what to do.
[0014]
The prediction analysis apparatus 1 of the present invention includes a known data storage unit 10, a prediction function generation unit 11, a prediction unit 12, and an analysis unit 13 in order to realize this process.
[0015]
The known data storage unit 10 stores known data describing a correspondence relationship between an attribute value and a result value associated with the attribute value.
[0016]
  The prediction function generation unit 11 uses the known data stored in the known data storage unit 10 to determine a decision tree (predicting the result value of unknown data by following the conditional branch of each node) as shown in FIG. , Rules as shown in FIG. 3 (predict the result value of unknown data by matching each rule)soGenerate the prediction function to be built.
[0017]
The prediction unit 12 includes a prediction function generated by the prediction function generation unit 11, and predicts a result value associated with an attribute value possessed by unknown data.
[0018]
The analysis unit 13 obtains analysis information indicating how to change the attribute value of the unknown data in order to change the prediction result of the prediction unit 12 to a desired prediction value, and presents it. .
[0019]
  This analysis unit 13 uses interactive processing in order to obtain effective analysis information using important attributes, or to expand the scope of application to real problems by removing attributes that cannot be changed from the analysis target. Set the attributes to be analyzed according to theEvery timeTherefore, a processing attribute setting unit 14 that automatically sets an attribute to be analyzed may be provided.
[0020]
In addition, the analysis unit 13 obtains effective analysis information using a meaningful attribute value search range, or excludes an unchangeable attribute value search range from an analysis target, thereby expanding the application range to a real problem. In order to widen, a search range setting unit 15 that sets a search range of attributes to be analyzed may be provided.
[0021]
Here, the function of the prediction analysis apparatus 1 of the present invention is specifically realized by a program, and this program can be stored in an appropriate recording medium such as a semiconductor memory readable by a computer. .
[0028]
  In the prediction analysis apparatus 1 of the present invention configured as described above, when the prediction result of the unknown data is obtained by the prediction unit 12 and the prediction result is not what the user desires, the analysis unit 13 By using an algorithm (decision tree or rule) for deriving the result value constructed from the known data, and changing the algorithm so that the unknown data will have the desired predicted value as the result value, One or a plurality of known data that has a predicted value as a result value and that follows the same predicted route as the unknown data predicted route that will realize the desired predicted value is specified and presented to the user as analysis information .
  That is, in the prediction analysis apparatus 1 of the present invention configured as described above, when a decision tree is used, first, a correspondence relationship between an attribute value and a result value associated therewith is described.Accumulated and stored in a known data database (known data storage unit 10)By following a decision tree made up of branching rules for attribute values created based on known data,Input from terminalPredict the result value associated with the attribute value of unknown data. Next, by changing the attribute description of some branch rules of the decision tree, the predicted result valueInput from terminalA new decision tree for changing to a desired predicted value is determined, and by tracing the new decision tree, known data having the desired predicted value as a result value is extracted from the known data, and the Output information about the extracted known data.
  At this time, the attribute to be analyzed may be set, and the branch rule attribute description of the decision tree may be changed with the branch rule describing the set attribute as the change target.
  Further, in the prediction analysis apparatus 1 of the present invention configured as described above, when using a rule set, first, a correspondence relationship between an attribute value and a result value associated therewith is described.Accumulated and stored in a known data database (known data storage unit 10)By matching against a rule set made up of rule descriptions for attribute values created based on known data,Input from terminalPredict the result value associated with the attribute value of unknown data. Next, by changing some rule descriptions of the rule set, the predicted result valueInput from terminalBy determining a new rule set for changing to the desired predicted value and collating with the new rule set, the known data having the desired predicted value as a result value is extracted from the known data, Information about the extracted known data is output.
  At this time, an attribute to be analyzed may be set, and a rule description describing the set attribute may be changed, and a part of the rule description of the rule set may be changed.
[0029]
When performing this process, the analysis unit 13 may also output the certainty factor indicated by the same predicted route.
[0030]
Then, when specifying a plurality of known data, one known data is selected by selecting one known data from the known data, or by applying an average value or majority of attribute values of the known data. May be processed so as to determine one known data to be presented.
[0031]
In this way, according to the prediction analysis apparatus 1 of the present invention, the user can obtain a prediction result when the prediction unit 12 obtains a prediction result of unknown data, and the prediction result is not what he / she desires. By having the predicted value as a result value and being able to obtain known data that follows the same predicted route as the unknown data predicted route where the desired predicted value will be realized, the predicted result can be changed to the desired predicted value. Therefore, it is possible to know how to change the attribute value of unknown data.
[0032]
As described above, according to the present invention, when a prediction result of unknown data is obtained and the prediction result is not what he / she desires, the user can convert the prediction result to a desired prediction value. In order to change, it becomes possible to know how to change the attribute value of unknown data.
[0033]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, the present invention will be described in detail according to embodiments.
[0034]
FIG. 4 illustrates an example of an embodiment of the prediction analysis apparatus 1 of the present invention.
[0035]
As shown in this figure, the prediction analysis apparatus 1 of the present invention includes a terminal 2, a known data database 100, a prediction function generation mechanism 200, a prediction mechanism 300 composed of software and hardware, and the present invention. And an analysis program 400 for performing characteristic processing.
[0036]
Here, the analysis program 400 and the prediction mechanism 300 configured by software can be stored in an appropriate recording medium such as a semiconductor memory readable by a computer.
[0037]
The known data database 100 stores and stores known data. This known data is the accumulated past case data, the conditions at that time and the accompanying results are known, and are the basis for the prediction.
[0038]
FIG. 5A illustrates an example of known data describing what kind of manufacturing conditions the non-defective product is manufactured and what kind of manufacturing condition the defective product is manufactured. In addition to the binary or multi-value discrete values such as “OK” / “NG”, the result may also indicate continuous values included in a certain region such as 0 to 1. is there.
[0039]
The predictive analysis apparatus 1 according to the present invention receives unknown data from the terminal 2 as shown in FIG. 5B when the known data database 100 stores and stores such known data and the result value is unknown. When it is input, the result value (result field value) of the unknown data is predicted, and when it is different from what the user wants, how is the attribute value (condition field value) of the unknown data? Processing for generating analysis information indicating whether or not to change the information and outputting it to the terminal 2 is performed.
[0040]
  The prediction mechanism 300 executes a process for predicting the result value of the unknown data, and is constructed by the prediction function generation mechanism 200. For example, the prediction tree shown in FIG. 2 or the decision tree shown in FIG. LouIf the technology related to the present invention is further realized,Y = ΣA_i× X_i(X_i: Attribute value, A_i: Coefficient, Y: result value) "or neural networkAnd MThis is realized by BR (Memory-Based Reasoning).
[0041]
If MBR which implement | achieves this prediction mechanism 300 is demonstrated, in MBR, as shown in FIG. 6, while considering the multidimensional space which the attribute which known data and unknown data have, the known data is statistically processed. The degree of influence of each attribute is calculated, and the degree of similarity between unknown data and each known data in the multidimensional space is calculated in consideration of this degree of influence.
[0042]
Then, according to this similarity, a plurality of known data similar to unknown data is extracted in the order of high similarity, and the certainty factor indicated by the result value of those known data (the certainty factor of the OK result value is such a value, Processing for predicting the result value indicating the highest certainty among the unknown data as a result value of unknown data I do.
[0043]
Various methods have been proposed for calculating the degree of influence, the method of calculating the degree of similarity, and the method of calculating the certainty factor used at this time. For example, Japanese Patent Laid-Open No. 2000-155681 disclosed by the present applicant It is possible to use the calculation method described.
[0044]
If a neural network that realizes the prediction mechanism 300 is described, the neural network calculates a product-sum value of an input value and a weight value associated therewith as shown in FIG. The basic unit to be converted is configured by being connected in a form such as a hierarchical network.
[0045]
The neural network configured in this way uses known data as learning data, and according to the back-propagation method as shown in FIGS. As a result of learning, when an attribute value of unknown data is input, a process of predicting a result value of the unknown data is performed.
[0046]
Here, in the back-propagation method, as shown in FIG. 9 (c), when the weight or threshold value is changed, the weight or threshold value vibrates during learning by adding the previous change as an inertia term. May be to prevent. Further, in FIG. 10, the back propagation method is specifically described by taking a case where the basic unit performs function conversion using a sigmoid function.
[0047]
As a modified example of this back-propagation method, there is a method of performing learning by adding a growth side suppression term S to a weight or threshold update formula as shown in FIG. According to this growth-side restrained back-propagation method, growth of a large weight is promoted, and growth of a small weight is suppressed and degenerates. As a result, as shown in FIG. Only the connection survives, and a structured neural network can be constructed.
[0048]
Here, in the structured neural network shown in FIG. 12, the internal connection having a weight indicating a sufficiently small absolute value is not displayed. The solid line represents a weight having a positive value, the broken line represents a weight having a negative value, and the thickness of the line represents the absolute value of the weight.
[0049]
Since the weights learned by the normal back-propagation method are combined in a complex manner, it is difficult to understand what kind of law is learned. On the other hand, since the weight learned by this growth-side restrained back-propagation method is simplified, it is possible to read what kind of law has been learned.
[0050]
FIG. 13 illustrates a processing flow for the overall processing executed by the analysis program 400.
[0051]
When the analysis program 400 is started, as shown in the processing flow of FIG. 13, first, in step 1, first, the known data is read from the known data database 100, and in step 2, unknown data to be analyzed is read. Input from the terminal 2 or the like.
[0052]
Subsequently, in step 3, the input unknown data is designated as a prediction target, and the prediction mechanism 300 is called to predict a result value of the unknown data and present the prediction result to the terminal 2.
[0053]
That is, when the prediction mechanism 300 is constructed with a decision tree as shown in FIG. 2, the result of the unknown data is obtained by following the conditional branch of each node of the decision tree according to the attribute value of the unknown data. Predict the value.
[0054]
Further, when the prediction mechanism 300 is constructed with the rules as shown in FIG. 3, the result value of the unknown data is predicted by collating with each rule according to the attribute value of the unknown data. .
[0055]
In addition, the prediction mechanism 300 displays “Y = ΣA_i× X_i(X_i: Attribute value, A_i:: coefficient, Y: result value) ”, the result value of the unknown data is predicted by substituting the attribute value of the unknown data into the computer expression.
[0056]
Further, when the prediction mechanism 300 is constructed by MBR as described with reference to FIG. 6, the similarity between unknown data and each known data is calculated in consideration of the influence of each attribute. By extracting a plurality of known data similar to the unknown data and specifying the largest certainty factor of the result values of the known data, the result value of the unknown data is predicted.
[0057]
When the prediction mechanism 300 is constructed by the neural network shown in FIG. 7 or FIG. 12, an attribute value of unknown data is input to the neural network, and an output value output in response to the input As a result, the result value of unknown data is predicted.
[0058]
In this way, the result value of unknown data is predicted in step 3 and displayed on the terminal 2, and then, in step 4, the predicted value is changed by the user in response to the display and analyzed. It is determined whether or not a request for whether or not to be issued has been issued. If the request has not been issued, the process is terminated, and if it has been issued, the process proceeds to step 5 to obtain a predicted value desired by the user. input.
[0059]
For example, when the result value indicates “OK” or “NG” and the prediction result “NG” is presented, the user knows how to make the prediction result “OK”. If it is desired, “OK” is input, so that “OK” is input.
[0060]
Subsequently, in step 6, it is determined whether or not there is a change attribute setting instruction from the user, and when it is determined that there is a change attribute setting instruction, the process proceeds to step 7 to change the change object (processing object). To determine that all the attributes are uniformly treated as change targets (processing targets) when it is determined that there is no change attribute setting instruction The processing of step 7 is omitted.
[0061]
As will be described later, the analysis program 400 calculates the change amount of the attribute value necessary for changing the result value of the unknown data to the predicted value desired by the user, or converts the result value into the unknown data having the desired predicted value. By searching for similar known data, a process for suggesting to the user how to change the attribute value of the unknown data is performed.
[0062]
In executing this processing, the analysis program 400 treats only attributes for which a change attribute setting instruction has been received from the user as a change target (processing target), so that an attribute that cannot be changed in practice is excluded from the calculation target. By removing the inferior attribute from the target of processing, processing is performed to expand the scope of application to real problems. In this step 7, the attribute to be changed (processing target) is set.
[0063]
In this setting process, as shown in FIG. 14A, the attribute “temperature 1” is treated as a change target (process target), and the attribute “temperature 2” is not treated as a change target (process target). Although setting is performed, this setting process can be automatically performed without depending on the dialog using the MBR or the structured neural network in addition to the dialog with the user.
[0064]
That is, when automatically setting using MBR, the degree of influence of each attribute obtained by MBR (calculated by statistical processing of known data and indicating how much influence is exerted on the prediction target) is used. For example, an attribute having an influence level equal to or less than 1/100 of the maximum influence level is set so as not to be changed (processed) because it does not contribute to prediction.
[0065]
The statistical methods for calculating the degree of influence include PCF (Per-Category Feature importance) method, CCF (Cross-Category Feature importance) method, MIC (Mutual Information Content) method, and ACF (Averaged Category Feature) method. And a technique described in Japanese Patent Application Laid-Open No. 2000-155681 disclosed by the present applicant.
[0066]
In the case of automatic setting using a structured neural network as shown in FIG. 12, an input unit that forms the foremost layer (provided in association with an attribute, and an attribute value is assigned to the basic unit of the next layer) The sum of the absolute values of the weights of the unit that distributes) is calculated. For example, the attribute input by an input unit having a total value less than or equal to 1/100 of the maximum total value does not contribute to prediction, and therefore is subject to change. This is done by setting so that it is not (processing target).
[0067]
In this way, when the change attribute setting is executed in step 7, subsequently, in step 8, it is determined whether or not there is an instruction to set the attribute search range from the user, and there is a search range setting instruction. When judging this, the process proceeds to step 9 to set the search range, and when judging that there is no instruction to set the search range, in order not to impose a search range restriction on all attributes, The process of step 9 is omitted.
[0068]
As will be described later, the analysis program 400 calculates the change amount of the attribute value necessary for changing the result value of the unknown data to the predicted value desired by the user, or converts the result value into the unknown data having the desired predicted value. By searching for similar known data, a process for suggesting to the user how to change the attribute value of the unknown data is performed.
[0069]
In executing this processing, the analysis program 400 changes the attribute or sets it as a processing target attribute on the condition that the search range is set by the user on the condition that the user enters the search range. We are processing to expand the scope of application. In this step 9, the attribute search range is set.
[0070]
In this setting process, as shown in FIG. 14B, the upper and lower limits are set such that the maximum value of the search range of the attribute “temperature 1” is “20.0” and the minimum value is “10.0”. Since the attribute “temperature 2” is not set as a change target, the search range is not set, and when the attribute value is a category value, a possible category value is set, For search results, for example:
f = ε−Σ (u_i-X_i)²
ε: Maximum allowable value i: Attribute number
u_i: Unknown data x_i: Known data
If the evaluation value f is 0 or more, the search is continued, and if it is 0 or less, the search is terminated. Process.
[0071]
In this way, when the setting of the attribute search range is executed in step 9, subsequently, in step 10, the prediction result desired by the user can be obtained by changing how the attribute value of the unknown data is changed. Predictive analysis is performed as to whether or not this is the case, and in subsequent step 11, the analysis result of the predictive analysis is output to the terminal 2, and the process is terminated.
[0072]
  15 is executed in step 10 of the processing flow of FIG.Related to the present inventionOf predictive analytics processingOne caseIs illustrated.
[0073]
When the analysis program 400 enters the prediction analysis processing shown in the processing flow of FIG. 15, first, in step 100, the known data read from the known data database 100, the input unknown data to be analyzed, and the prediction mechanism. The predicted value of unknown data obtained by 300 (the predicted value of the result value) and the desired predicted value input in a form of changing the predicted value are prepared.
[0074]
Subsequently, in step 101, the maximum storage number l is set, and an array D for storing data corresponding to the maximum storage number l._lAnd in step 102, 0 is set to the variable j.
[0075]
Subsequently, in step 103, the value of the variable j is incremented by one. In the following step 104, it is determined whether or not the value of the variable j exceeds the number N of known data. Then, the process proceeds to step 105, where the known data indicated by the l record numbers stored is output to the terminal 2, and the process is terminated.
[0076]
At this time, instead of outputting a list of the known data, the latest known data included in the known data is output as a representative, or a prescribed arithmetic processing (average in the case of numeric attributes) is performed on the known data. In some cases, the representative known data obtained by calculating a value and applying a majority decision in the case of a category attribute is output.
[0077]
On the other hand, when it is determined in step 104 that the value of the variable j does not exceed the number N of known data, the process proceeds to step 106 where the result value is specified by referring to the known data indicated by the record number j. In the following step 107, it is determined whether or not the identified result value matches the desired predicted value input in a manner of changing the predicted value of unknown data.
[0078]
When it is determined by this determination processing that the result value of the known data indicated by the record number j does not match the desired predicted value, the process returns to step 103 to process the next known data. On the other hand, when it is determined that they match, the process proceeds to step 108 and the search range set in the attribute value of the known data indicated by the record number j (set in step 9 of the processing flow in FIG. 13) is set. Determine if there is something that doesn't fit.
[0079]
When it is determined by this determination processing that there is an attribute value of the known data indicated by the record number j that does not fall within the search range, the process returns to step 103 to process the next known data. On the other hand, when it is determined that there is nothing that does not fall within the search range, the routine proceeds to step 109, where the attribute value x of the known data indicated by the record number j is stored._iAnd attribute value u of unknown data_iFor example, the distance d between
d = [Σ (u_i-X_i)²]^1/2
Calculate according to
[0080]
In calculating the distance d, when setting the attribute to be processed and the attribute not to be processed in step 7 of the processing flow of FIG. 13, the distance d is calculated while excluding the attribute not to be processed. To process. Further, when the attribute is a numerical attribute, for example, the distance d is calculated by partitioning it and converting it into a category value, or as known in Japanese Patent Laid-Open No. 2000-155681 disclosed by the present applicant, Processing is performed so that the distance d is calculated using a distance definition using the standard deviation obtained from the data as the denominator and the difference between the attribute values as the numerator.
[0081]
Subsequently, in step 110, the calculated distance d is changed to the array D._lWhen it is determined whether or not the distance d is smaller than any of the distances d stored in, and when it is determined that the distance d is not smaller than any of the distances d, that is, when it is determined that the calculated distance d is greater, Return to step 103 to process the data.
[0082]
On the other hand, array D_lWhen it is determined that there is a distance d that is greater than the calculated distance d among the distances d stored in, the processing proceeds to step 111 and the array D_lDiscarding the largest one of the distances d stored in, and newly calculating the distance d to the array D_lIn addition, the storage record number associated with the discarded distance d is discarded, the record number j associated with the newly calculated distance d is stored, the storage record number is updated, and the process returns to step 103. .
[0083]
In this way, when the analysis program 400 executes the prediction analysis processing shown in the processing flow of FIG. 15, the analysis program 400 has a desired prediction value as a result value and l pieces of attribute values similar to the attribute values of unknown data. The known data is identified and processed so as to be output as an analysis result.
[0084]
In the processing flow of FIG. 15, the attribute value x of the known data indicated by the record number j is considered without considering the importance of the attribute._iAnd attribute value u of unknown data_iAs shown in the processing flow of FIG. 16, the distance d between each of the attributes is calculated by the MBR._iConsidering the shape
d = [Σw_i(U_i-X_i)²]^1/2
You may make it calculate according to. Where w_iAs described above, it is also possible to use the network weight of each attribute obtained by the structured neural network.
[0085]
  FIG. 17 shows a prediction analysis process executed in step 10 of the process flow of FIG.The fruit ofAn example embodiment is illustrated. Here, in this prediction analysis processing, it is assumed that the prediction mechanism 300 is constructed by a decision tree as shown in FIG.
[0086]
When the analysis program 400 enters the prediction analysis process shown in the processing flow of FIG. 17, first, in step 100, the known data read from the known data database 100, the input unknown data to be analyzed, and the prediction mechanism. The predicted value of unknown data obtained by 300 (the predicted value of the result value) and the desired predicted value input in a form of changing the predicted value are prepared.
[0087]
Subsequently, in step 101, the path of the decision tree that obtained the predicted value of unknown data is referred to. Subsequently, in step 102, it is determined whether or not the node of the path can be traced up one by one. When it is determined that the path cannot be traced, the process is terminated. When it is determined that the trace can be traced, the process proceeds to step 103. Then, the node is moved up by one in the path of the decision tree that obtained the predicted value of the unknown data.
[0088]
Subsequently, in step 104, it is determined whether or not the traced node describes a branch rule relating to the attribute to be changed (set in step 7 of the processing flow in FIG. 13). When it is determined that the branching rule related to the attribute of is not described, the process returns to step 102 to further follow the node by one.
[0089]
On the other hand, when it is determined in step 104 that the traced node describes the branch rule relating to the attribute to be changed, the process proceeds to step 105 so that the unknown data follows another decision tree path. Then, change the attribute description of the branch rule to another one.
[0090]
Subsequently, in step 106, when it is determined whether or not the changed attribute description enters the set search range (set in step 9 of the processing flow of FIG. 13) and not, Further, the process returns to step 102 to go up one node. On the other hand, when it is determined to enter the search range, the process proceeds to step 107, and the result value of unknown data is predicted by following the decision tree from the node.
[0091]
Subsequently, at step 108, it is determined whether or not the result value of the unknown data becomes the desired predicted value. When it is determined that it does not become the desired predicted value, the process goes to step 102 to further follow the node up by one. Return.
[0092]
On the other hand, when it is determined that the desired predicted value is obtained, the process proceeds to step 109, where the known data having the desired predicted value as a result value is extracted from the known data that follows the decision tree path leading to the desired predicted value, In the subsequent step 110, a list of the extracted known data is output, the latest known data included in the extracted known data is output as a representative, or a prescribed arithmetic processing (numerical attribute of the numerical attribute) is output. In this case, an average value is calculated, and in the case of a category attribute, representative known data obtained by performing a calculation process such as taking a majority vote is output.
[0093]
At the time of this output, the desired predicted value from the result values of the known data entering the leaf of the changed decision tree (although the result values are almost the same but not necessarily the same result value due to the nature of the decision tree) The certainty factor defined by the ratio of the number of the desired predicted values and the number of all the result values that enter the leaf may be obtained and output.
[0094]
In this way, when the analysis program 400 follows the processing flow of FIG. 17, as shown in FIG. 18, the analysis program 400 starts from the leaf where the unknown data is predicted, proceeds to the node immediately above, and corresponds to that node. Proceed to another leaf by changing the attribute and make a prediction. If the desired prediction value is not obtained, specify the path to obtain the desired prediction value by repeating the process of proceeding to the upper node again. Then, known data having a desired predicted value that follows the identified path as a result value is obtained and processed so as to be output as an analysis result.
[0095]
FIG. 19 illustrates another embodiment of the prediction analysis process executed in step 10 of the process flow of FIG. Here, in this prediction analysis processing, it is assumed that the prediction mechanism 300 is constructed with the rules as shown in FIG.
[0096]
When the analysis program 400 enters the prediction analysis process shown in the processing flow of FIG. 19, first, in step 100, the known data read from the known data database 100, the input unknown data to be analyzed, and the prediction mechanism. The predicted value of unknown data obtained by 300 (the predicted value of the result value) and the desired predicted value input in a form of changing the predicted value are prepared.
[0097]
Subsequently, in step 101, the rule that obtained the predicted value of the unknown data is referred to. Subsequently, in step 102, it is determined whether or not the rule can be changed. When it is determined that the rule cannot be changed, the process is ended. When it is determined that the rule can be changed, the process proceeds to step 103. One rule part to be changed is selected from the rules that have obtained predicted values of unknown data.
[0098]
Subsequently, in step 104, it is determined whether or not the selected rule part describes a rule relating to the attribute to be changed (set in step 7 of the processing flow in FIG. 13). When it is determined that the rule regarding is not described, the process returns to step 102 to change another rule part.
[0099]
On the other hand, when it is determined in step 104 that the selected rule portion describes a rule relating to the attribute to be changed, the process proceeds to step 105 so that unknown data will follow another rule path. Change the attribute description of the rule part to another one.
[0100]
Subsequently, in step 106, when it is determined whether or not the changed attribute description enters the set search range (set in step 9 of the processing flow of FIG. 13) and not, Return to step 102 to change another rule location. On the other hand, when it is determined to enter the search range, the process proceeds to step 107, and the result value of unknown data is predicted by following the changed rule.
[0101]
Subsequently, at step 108, it is determined whether or not the result value of the unknown data becomes the desired predicted value, and when it is determined that it does not become the desired predicted value, the process goes to step 102 to change another rule part. Return.
[0102]
On the other hand, when it is determined that the desired predicted value is obtained, the process proceeds to step 109, where the known data having the desired predicted value as a result value is extracted from the known data that follows the rule path that reaches the desired predicted value, and continues. In step 110, a list of the extracted known data is output, the latest known data included in the extracted known data is output as a representative, or the specified arithmetic processing (in the case of a numerical attribute) In the case of category attribute, representative known data obtained by performing an arithmetic process such as calculating an average value and taking a majority vote in the case of a category attribute is output.
[0103]
In this output, it becomes a desired predicted value from the result values of known data that enter the conclusion of the changed rule (although the result values are almost the same but not necessarily the same result value due to the nature of the rules). Sometimes, a certainty factor defined by the ratio between the number of items that become the desired predicted value and the number of all result values that enter the conclusion is obtained and output.
[0104]
In this way, when the analysis program 400 follows the processing flow of FIG. 19, as shown in FIG. 20, for example, the analysis program 400 changes the attribute value corresponding to the rule location starting from the rightmost rule location. If the predicted value is not the desired predicted value, the rule that obtains the desired predicted value is identified by repeating the process of proceeding to the next rule location, and the desired predicted value that follows the specified rule Is obtained as a result value and processed so as to be output as an analysis result.
[0105]
  21 is executed in step 10 of the processing flow of FIG.Related to the present inventionPredictive analytics processingExampleIs illustrated. Here, in this prediction analysis processing, it is assumed that the prediction mechanism 300 is constructed by a neural network as shown in FIG.
[0106]
When the analysis program 400 enters the prediction analysis process shown in the process flow of FIG. 21, first, in step 100, the known data read from the known data database 100, the input unknown data to be analyzed, and the prediction mechanism. The predicted value of unknown data obtained by 300 (the predicted value of the result value) and the desired predicted value input in a form of changing the predicted value are prepared.
[0107]
Subsequently, in step 101, the attribute value of the unknown data is input to the neural network that constructs the prediction mechanism 300, and the learning process of the back propagation method according to the Iterative Inversion method is executed.
[0108]
In the normal back-propagation learning process, as shown in FIGS. 8 to 10, the weight is learned by back-propagating the error between the prediction value output from the neural network and the desired prediction value on the neural network. On the other hand, in the back propagation method according to the Iterative Inversion method, as shown in FIG. 22, in the learned neural network, the error between the predicted value output from the neural network and the desired predicted value is represented on the neural network. Is processed to learn the input value (in this case, the attribute value of unknown data).
[0109]
From now on, when the back-propagation method according to the Iterative Inversion method is executed, the attribute value of the unknown data that realizes the desired predicted value can be obtained.
[0110]
Subsequently, in step 102, it is determined whether or not the attribute to be changed (set in step 7 of the processing flow in FIG. 13) is set, and if the attribute to be changed is set, In step 103, the attribute value of only the attribute to be changed is changed according to the attribute value change amount obtained by the Iterative Inversion method. On the other hand, if the attribute to be changed is not set, the process proceeds to step 104, and the attribute values of all attributes of the unknown data are changed according to the attribute value change amount obtained by the Iterative Inversion method.
[0111]
Subsequently, in step 105, it is determined whether or not the changed attribute value falls within the set search range (set in step 9 of the processing flow in FIG. 13). When it is determined to enter, the process proceeds to step 106 to obtain a predicted value output from the neural network in response to the input of unknown data whose attribute value has been changed.
[0112]
Subsequently, in step 107, it is determined whether or not the acquired predicted value is the desired predicted value. If the predicted value is not the desired predicted value, the process returns to step 101 to advance learning by the Iterative Inversion method, If it is the desired predicted value, the process proceeds to step 108, and the attribute value of the unknown data changed by the Iterative Inversion method is output. At this time, instead of outputting the attribute value itself of the unknown data, the change amount between the original attribute value of the unknown data and the attribute value obtained by learning may be output.
[0113]
In this way, when the analysis program 400 executes the prediction analysis process shown in the process flow of FIG. 21, the attribute value of the unknown data that will realize the desired prediction value and the amount of change to the attribute value are determined. It is calculated and processed to output it as an analysis result.
[0114]
In the processing flow of FIG. 21, it is assumed that the prediction mechanism 300 is constructed by a neural network as shown in FIG._i× X_i(X_i: Attribute value, A_i:, Coefficient, Y: result value) ", the unknown data whose desired predicted value will be realized using an inverse calculation method as shown in FIG. The attribute value possessed by and the amount of change to the attribute value are calculated.
[0115]
(Supplementary Note 1) A prediction unit that predicts a result value associated with an attribute value of unknown data using known data that describes a correspondence relationship between the attribute value and a result value associated with the attribute value; and In order to change the prediction result to a desired prediction value, the analysis information indicating how to change the attribute value of the unknown data is obtained and provided with an analysis unit that presents the analysis information A predictive analyzer.
[0116]
(Supplementary note 2) In the prediction analysis apparatus described in supplementary note 1, the analysis unit has a desired prediction value as a result value as the analysis information and one attribute value similar to the attribute value of unknown data Alternatively, a predictive analyzer characterized by processing so as to specify a plurality of known data.
[0117]
(Supplementary note 3) In the prediction analysis apparatus described in Supplementary note 2, the analysis unit performs processing so as to identify known data similar to unknown data while considering the importance indicated by each attribute. Prediction analysis device.
[0118]
(Supplementary Note 4) In the predictive analysis apparatus described in Supplementary Note 3, the analysis unit performs processing so that each attribute obtained by the MBR method has an influence on a result value as the importance. Prediction analysis device.
[0119]
(Supplementary note 5) Predictive analysis apparatus described in supplementary note 3, wherein the analysis unit performs processing so as to use a network weight obtained by learning of a structured neural network as the importance. apparatus.
[0120]
(Supplementary Note 6) In the prediction analysis apparatus described in any one of Supplementary Notes 2 to 5, the analysis unit selects a plurality of known data, and selects one known data from the known data. Alternatively, a predictive analyzer characterized by processing to determine one known data to be presented by calculating one known data from the known data according to a prescribed operation.
[0121]
(Supplementary note 7) In the prediction analysis apparatus described in supplementary note 1, the analysis unit may use, as the analysis information, an attribute value of unknown data that realizes a desired prediction value or a change to the attribute value. A predictive analyzer characterized by processing to calculate a quantity.
[0122]
(Supplementary note 8) In the predictive analysis apparatus described in supplementary note 7, the analysis unit uses a data conversion function for calculating a result value constructed from known data, and the unknown without changing the data conversion function. By learning in the form of back-propagating the error between the result value of the data and the desired predicted value, the attribute value of the unknown data that will achieve the desired predicted value, or the amount of change to the attribute value A predictive analyzer characterized by processing to calculate.
[0123]
(Supplementary note 9) In the prediction analysis apparatus described in supplementary note 1, the analysis unit has a desired prediction value as a result value as the analysis information, and a prediction route of unknown data that realizes the desired prediction value A prediction analysis apparatus characterized by processing to identify one or a plurality of known data according to the same prediction route.
[0124]
(Additional remark 10) The prediction analysis apparatus described in Additional remark 9 WHEREIN: The said analysis part is processed so that it may output also about the reliability which the said same prediction route shows, The prediction analysis apparatus characterized by the above-mentioned.
[0125]
(Additional remark 11) In the prediction analysis apparatus described in Additional remark 9 or 10, the said analysis part uses the algorithm for derivation | leading-out the result value constructed | assembled by known data, and unknown data has a desired prediction value as a result value A prediction analysis apparatus characterized in that processing is performed so as to specify known data according to the same prediction route while changing the algorithm so as to be different.
[0126]
(Supplementary Note 12) In the predictive analysis apparatus according to any one of Supplementary Notes 9 to 11, the analysis unit selects a plurality of known data, and selects one known data from the known data. Alternatively, a predictive analyzer characterized by processing to determine one known data to be presented by calculating one known data from the known data according to a prescribed operation.
[0127]
(Supplementary note 13) In the predictive analysis apparatus according to any one of supplementary notes 1 to 12, the analysis unit sets an attribute to be analyzed in obtaining the analysis information, and sets the analysis target attribute that has been set A predictive analysis apparatus characterized by processing so as to obtain the analysis information using the attribute value.
[0128]
(Additional remark 14) The prediction analysis apparatus described in additional remark 13 WHEREIN: According to dialogue processing, the said analysis part is processed so that the attribute used as analysis object may be set, The prediction analysis apparatus characterized by the above-mentioned.
[0129]
(Supplementary Note 15) In the prediction analysis apparatus described in Supplementary Note 13, the analysis unit performs processing so as to set an attribute to be analyzed based on the importance indicated by each attribute calculated from known data. , Predictive analysis device characterized.
[0130]
(Supplementary note 16) In the prediction analysis apparatus described in supplementary note 15, the analysis unit performs processing so that each attribute obtained by the MBR method has an influence on a result value as the importance. Prediction analysis device.
[0131]
(Supplementary note 17) Prediction analysis apparatus according to supplementary note 15, wherein the analysis unit performs processing so as to use a network weight obtained by learning of a structured neural network as the importance. Analysis equipment.
[0132]
(Supplementary note 18) In the predictive analysis apparatus described in any one of supplementary notes 1 to 17, the analysis unit sets a search range for attribute values and obtains the set search range in obtaining the analysis information. A predictive analyzer characterized in that processing is performed so as to obtain the analysis information while being entered.
[0133]
(Supplementary note 19) A process of predicting a result value associated with an attribute value possessed by unknown data using known data describing a correspondence relationship between the attribute value and a result value associated with the attribute value, and hope for the prediction result A prediction analysis program for obtaining analysis information indicating how to change an attribute value of unknown data in order to change the attribute value to a predicted value, and causing the computer to execute a process of presenting the analysis information.
[0134]
(Supplementary note 20) A process for predicting a result value associated with an attribute value possessed by unknown data using known data describing a correspondence relationship between the attribute value and a result value associated with the attribute value, and hope for the prediction result In order to change the attribute value to the predicted value, the analysis information indicating how to change the attribute value of the unknown data was obtained, and the program for causing the computer to execute the process of presenting it was recorded Prediction analysis program recording medium.
[0135]
【The invention's effect】
  As described above, according to the present invention, when a prediction result of unknown data is obtained and the prediction result is not what he desires, the user canValue as a result value and known data that follows the same prediction route as the unknown data prediction route that will realize the desired prediction value.As a result, it is possible to know how to change the attribute value of unknown data in order to change the prediction result to a desired predicted value.
[Brief description of the drawings]
FIG. 1 is a schematic configuration diagram of the present invention.
FIG. 2 is an explanatory diagram of a decision tree.
FIG. 3 is an explanatory diagram of rules.
FIG. 4 is an example of an embodiment of the present invention.
FIG. 5 is an explanatory diagram of known data / unknown data.
FIG. 6 is an explanatory diagram of MBR.
FIG. 7 is an explanatory diagram of a neural network.
FIG. 8 is an explanatory diagram of a back propagation method.
FIG. 9 is an explanatory diagram of a back propagation method.
FIG. 10 is an explanatory diagram of a back propagation method.
FIG. 11 is an explanatory diagram of a growth-side restrained back propagation method.
FIG. 12 is an explanatory diagram of a structured neural network.
FIG. 13 is a processing flow for overall processing executed by an analysis program.
FIG. 14 is an explanatory diagram of change attribute setting processing;
FIG. 15 is an example of a predictive analysis process related to the present invention.
FIG. 16 is an example of a prediction analysis process related to the present invention.
FIG. 17 is an example of a predictive analysis process.
FIG. 18 is an explanatory diagram of a prediction analysis process.
FIG. 19 is another example of a predictive analysis process.
FIG. 20 is an explanatory diagram of a prediction analysis process.
FIG. 21Related to the present inventionOf predictive analytics processingOne caseIt is.
FIG. 22 is an explanatory diagram of the Iterative Inversion method.
FIG. 23 is an explanatory diagram of an inverse calculation method.
[Explanation of symbols]
  1 Predictive analyzer
  10 Known data storage
  11 Prediction function generator
  12 Predictor
  13 Analysis Department
  14 Processing attribute setting part
  15 Search range setting section

Claims (6)

By following a decision tree composed of branching rules for attribute values created based on known data stored and stored in a known data database that describes the correspondence between attribute values and associated result values A prediction unit that predicts a result value associated with an attribute value of unknown data input from a terminal ;
By changing the attribute description of some branch rules of the decision tree, a new decision tree for changing the predicted result value of the prediction unit to the desired predicted value input from the terminal is determined. An analysis unit that extracts the known data having the desired predicted value as a result value from the known data by tracing the new decision tree, and outputs information about the extracted known data; To prepare,
A characteristic predictive analyzer. The prediction analysis apparatus according to claim 1,
The analysis unit sets an attribute to be analyzed, changes the branch rule attribute description of the decision tree, with the branch rule describing the set attribute as a change target,
A characteristic predictive analyzer. Attribute values and created known data database that describes correspondence between the result values associated therewith based on the accumulated stored known data, collating rule set to consist of rules description of the attribute values And a prediction unit that predicts a result value associated with the attribute value of unknown data input from the terminal ,
By changing a part of the rule description of the rule set, a new rule set for changing the predicted result value of the prediction unit to the desired predicted value input from the terminal is determined, and the new An analysis unit that extracts the known data having the desired predicted value as a result value from the known data by collating with a set of rules and outputs information about the extracted known data The
A characteristic predictive analyzer. In the prediction analysis device according to claim 3,
The analysis unit sets an attribute to be analyzed, changes the rule description included in the rule set, with the rule description describing the set attribute as a change target,
A characteristic predictive analyzer. A prediction analysis program used for realizing a prediction analysis device,
Computer
By following a decision tree composed of branching rules for attribute values created based on known data stored and stored in a known data database that describes the correspondence between attribute values and associated result values A prediction unit that predicts a result value associated with an attribute value of unknown data input from a terminal ;
By changing the attribute description of some branch rules of the decision tree, a new decision tree for changing the predicted result value of the prediction unit to the desired predicted value input from the terminal is determined. The analysis unit that extracts the known data having the desired predicted value as a result value from the known data by tracing the new decision tree and outputs information on the extracted known data Predictive analysis program. A prediction analysis program used for realizing a prediction analysis device,
Computer
Attribute values and created known data database that describes correspondence between the result values associated therewith based on the accumulated stored known data, collating rule set to consist of rules description of the attribute values And a prediction unit that predicts a result value associated with the attribute value of unknown data input from the terminal ,
By changing a part of the rule description of the rule set, a new rule set for changing the predicted result value of the prediction unit to the desired predicted value input from the terminal is determined, and the new In order to function as an analysis unit that extracts the known data having the desired predicted value as a result value from the known data and outputs information about the extracted known data by collating with a known rule set Prediction analysis program.

Images