JP4957127B2 - Apparatus, method, and program for extracting partial relationship between attributes - Google Patents

Description

  The present invention relates to an apparatus, a method, and a program for extracting a partial relationship between attributes in order to perform determination of a model as a basis of inference in a certain inference / simulation system using known learning data.

  In this specification, “learning data” means data for constructing a certain model (or model expression) that expresses the characteristics using the learning data. Hereinafter, it is assumed that “data” is “learning data” unless otherwise specified in the text.

  In a system that estimates a result without executing a process that actually corresponds, such as a prediction / inference system or an approximate value calculation system, a model that is a prerequisite for that is required.

  In many cases (except when the laws of nature and physics can be used), this model generation requires a lot of study effort. For this reason, a future that occurs under the same conditions using known information in recent years A technology for automatically generating a model for estimating the information of the above has been established.

  Since the performance of prediction / inference is model-dependent, this model requires a simple grasp of known information. In view of the calculation load in the prediction process and the recognizability when the model is presented to the user, it is desirable to be as simple as possible.

  For this reason, the function which reflects the characteristic which known information contains and produces a simpler model is needed.

  There are two problems here. One is how the model generation mechanism recognizes given known information, and the other is what model is generated using the recognized information.

<Problems in model generation>
For example, a Bayesian Network learning function exists as a technique for extracting the relationship between fields using known information and modeling the data space using this relationship. This learning function is a very general function, and these are used to determine the existence of a relationship between fields based on a certain evaluation criterion (MDL, AIC, BIC, etc.) and to build a model.

  In this specification, regarding the “presence / absence of relationship”, “relevant” means that when the value of a certain field (group) is uniquely (or within a certain range), another field (group) ) Attribute values (or probabilities of taking those values) are limited (divergence from the average). For example, when there is a gender field, an educational field, and an undergraduate field (for university graduates), the ratio of junior high school, high school and university graduates to all boys and the ratio of middle school, high school and university graduates to all girls are not so different. In some cases, it is expressed that the relationship between gender and educational background is thin (or small), and when there is a bias such as a low percentage of female graduates in engineering, there is a strong relationship between gender and faculty of origin. It expresses.

  Further, in this specification, “relationship strength” is an index that expresses the strength of the relationship. In some cases, it can be simply calculated from the bias of attribute values between two fields, or “model evaluation value (field "When there is a relationship between group A and field group B)-model evaluation value (when there is no relationship between field group A and field group B)" may be used as the relationship strength between field group A and field group B.

  Further, in this specification, “field” means one attribute constituting data. For example, a vertical line in FIG. 5 (a certain survey item such as sex, age, etc.) is shown. Further, the “subset of attribute values” means a part of attribute values that a field can take. Also, “subset constituent element” means each attribute value belonging to the subset.

  In such a model, generating a brute force relationship between fields makes a causal relation expression (eg, probability table, probability expression function) huge or complicated, and makes the model complicated. In the modeling using the evaluation criteria exemplified in (1), it is not adopted, and therefore, fields having small evaluation values are expressed as irrelevant.

  However, there are cases in which there is a strong relationship in the case of a specific value even if the relationship is weak when viewed as the entire corresponding field. For example, the number of samples is small, but when the corresponding field takes the value, the value of the other field is determined with high probability. When such a relationship is “no relationship”, it causes a situation in which useful information is not reflected in the model in the inference process.

<Information recognition problems>
In addition, when there are a wide variety of field values in the model construction, there is a function to combine them when taking continuous values.
However, the process of “combining” in a known system means one of the following two.
1) Collect the ones with low appearance frequency as “others” into one value.
2) When values have permutations or are continuous, adjacent values are processed together so that the number of attributes is less than the allowable range.

  For this reason, the characteristics of individual attribute values to be processed together are ignored and may disappear before model construction.

  In this specification, “attribute value” means each value of field of a single record. In addition to discrete values such as “male” and “female”, it may mean continuous values such as “height”.

<Example of problem occurrence>
In the case where the influence of the above two problems is particularly large, the number of attributes is extremely large and there are infrequent attributes, and the relationship with other fields in the case of taking this infrequent attribute value is the respective attribute value. An example of the operation of the conventional function is shown below using the model generation process in this case.

<Data overview>
Take the example of handling the country of origin in the process of creating an annual income model from a personal information database in the USA. The data shown in FIG. 5 lists age, sex, educational background, origin, family relationship, annual income, and other various personal information in each of a plurality of fields. As a matter of course, as shown in FIG. 6, it is composed of the majority of people from USA and a small number of people from other countries. Based on this, consider creating an annual income expression model.

  In addition, in FIG. 6, the color classification is an educational background distribution (red: graduated from elementary school, green: graduated from middle school, blue: graduated from high school, purple: university or higher), and has 40 countries. As can be seen from the above, the proportion of each educational background varies.

  Note that “data” constituting “learning data” indicates a certain information group, and takes a form as shown in FIG. 5 (personal information DB in U.S.A), for example. In the figure, all cells (single record, single field) are filled, but they may be missing. In addition, since information having a plurality of meanings is integrated into a single field, it may be easier to generate a model later if field division is performed for each semantic unit in advance.

<Modeling>
Since the model including the brute force of each item is very complicated, only the important relationship is extracted and a relationship graph having the network structure shown in FIG. 7 is created.

  For the relationship extraction, various relationship strength evaluation formulas (known examples include MDL, AIC, etc.). In many cases, these evaluation criteria determine the strength of the relationship by comprehensively judging the relationship between items and the complexity of the model (interrelation between items as a whole information). That is, the relationship strength treats the complexity of the model when there is a corresponding relationship as a penalty.

<When processing the raw information>
Therefore, an item with a high degree of freedom represented by the country of origin (item with a large number of attributes) is difficult to use as a cause of other items. As for the origin, the majority is USA, and as a matter of course, the average of various items of USA personnel is close to the overall average, so the relationship between origin and others is expressed lightly as shown in FIG.

<Use of existing data summary function>
Here, in order to confirm whether the weakness of this relationship is unique to the data or due to the excessive number of attributes, it is considered to integrate country names. The technique presumed to be provided by the most systems is to collect the ones that appear infrequently.

  In this case, it will be divided into U.S.A with absolute quantity and others. Alternatively, when the user designates a category that is explicitly meaningful, for example, there is a method of classifying the category into five categories based on the name of the continent in which the country exists.

  However, the above summary method does not consider the influence on the inference process.

  For example, considering the relationship between the country name and educational background as an example of data characteristics, the normal summary function, as shown in Figure 9, is a relatively high-educated USA (or North American) graduate and other low-educated graduates It will be recognized as a person.

  However, in Asia and Europe, as shown in FIG. 10, it can be seen that there is a bias in educational background depending on the country of origin.

  For example, if Taiwan and Vietnam are grouped together outside the USA or Asia, it will be clear at a glance that the relationship between educational background and country of origin will be lost.

  Therefore, when it is necessary to predict educational background or some information that depends on educational background, if the above method is used, the generated model is used even though the relevant relationship exists in the learning data. However, this relationship could not be used in the prediction process, which caused a problem that the prediction performance deteriorated.

  It is an object of the present invention to reflect the relationship held in the original data in the model while avoiding the complexity of the model by collecting items that have similar relationships with other field groups. This realizes the provision of a high-performance inference model.

  In order to solve the above-described problem, the present invention provides an apparatus for extracting a partial relationship between attributes in a model creation apparatus that forms a predetermined model based on attribute values in a plurality of fields included in each of a plurality of data and the appearance frequency of the attribute values. An overall relationship strength calculating means for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for the predetermined attribute and its frequency distribution, and about the predetermined attribute Based on a subset of attribute values appearing in the field group, a partial relation strength calculating means for calculating the strength of a partial relation between fields to which the attribute value belongs, and a partial relation strength calculating means and an inter-field relationship setting means for determining the relationship between the fields based on the strength of the partial relationship between the fields.

  Further, in the partial relation extraction device between attributes of the present invention, the inter-field relation setting means, when the strength of the partial relation between fields calculated by the partial relation strength calculation means is equal to or greater than a predetermined threshold, As a relationship between fields, the relationship strength between the subsets is used as an evaluation point between fields, or the presence or absence of a relationship between subsets is determined.

  Further, the present invention causes a computer to execute a method for extracting a partial relationship between attributes in a model creating apparatus that constitutes a predetermined model based on attribute values in a plurality of fields included in each of a plurality of data and the appearance frequency of the attribute values. A program for extracting a partial relationship between attributes, a total relationship strength calculation step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution And a partial relationship strength calculating step for calculating the strength of the partial relationship between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute, and the partial relationship A field relationship setting step for determining the relationship between the fields based on the strength of the partial relationship between the fields calculated in the strength calculation step.

  Moreover, in the partial relation extraction program between attributes of the present invention, the inter-field relation setting step is performed when the strength of the partial relation between fields calculated by the partial relation strength calculation step is equal to or greater than a predetermined threshold. As a relationship between fields, the strength of the relationship between the subsets is used as an evaluation point between fields, or the presence or absence of a relationship between subsets is determined.

  Further, the present invention is a method for extracting a partial relationship between attributes in a model creating apparatus that constitutes a predetermined model based on attribute values in a plurality of fields included in each of a plurality of data and the appearance frequency of the attribute values. An overall relationship strength calculating step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for the attribute of the attribute and its frequency distribution; and in the field group for the predetermined attribute A partial relation strength calculating step for calculating the strength of a partial relation between fields to which the attribute value belongs based on a subset of attribute values appearing in the field, and a partial relation between fields calculated by the partial relation strength calculating step And a field relationship setting step for determining the relationship between the fields based on the strength of the field.

  Note that in the embodiment of the present invention, some model (or based on the attribute value of each field appearing in given data and its appearance frequency (or occurrence probability distribution: frequency distribution) In the process of constructing model formula parameters and approximate solution derivation functions (models below, implying all of them), the learning data has multiple fields, and it is required that the strength of the relationship between fields be ordered or evaluated (For example, the following may be considered as the cause of the request. Because of the large number of fields, it is difficult to consider the relationship between all the fields in the processing. For this reason, the related field group is based on some criteria. There is a problem in considering the effects of other fields on one field in an equivalent manner, and the effects of each field may be ordered or scored. When necessary, pay attention to the relationship between fields in the learning data, and the strength of the relationship calculated based on all attribute values appearing in the field group and their frequency distribution is low (the relationship is sparse) Even in such cases, the relationship calculated using the “subset” of attribute values appearing on the field (or its frequency distribution, bias when compared to the entire average value, contribution to prediction performance) is constant. If the threshold (fixed in advance, automatically calculated from the overall relationship strength or the number of attribute types, specified by the user, etc.) is stronger than the corresponding field, the relationship strength between the “subsets” is evaluated between the fields or A partial relationship extraction apparatus, a method thereof, and a program thereof, which employ the presence / absence of a relationship between “subsets”, are described.

  As described above, according to the present invention, by extracting the relationship partially established between fields in the learning data and adopting it in the model, it is small when measured as an influence on the entire model. However, when viewed as a specific field relationship, it is possible to use a large relationship for the inference processing, thereby improving the inference processing performance.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(Outline of the embodiment)
FIG. 1 is a block diagram showing the concept of an embodiment of the present invention.
The inference model creation apparatus in FIG. 1 includes a learning data input mechanism (DB) 1, a model learning mechanism 2, an evaluation criterion calculation mechanism (corresponding to the whole relationship strength calculation means) 3, an evaluation criterion calculation mechanism (partial relationship strength calculation). 4) Correspondence strength similarity calculation mechanism (subset component similarity calculation mechanism: constituting first and second means) 5, partial relationship calculation target / condition designation mechanism (first, second) 2, which constitutes the third designating means) 6, and the model created by this apparatus is provided to the inference mechanism 7.

  In this apparatus, the learning data input mechanism 1, the model learning mechanism 2, and the evaluation criterion calculation mechanism 3 are well-known, and a grouping function (attribute value conversion) of attribute groups that can be grouped (similar). It is also possible to use the function multiple times).

  In the above configuration, an information extraction process (partial relation extraction process) using an evaluation standard (partial relation strength) calculation mechanism will be described as a basic operation in the embodiment of the present invention.

  This is because when there are multiple fields in the learning data and it is required to order or evaluate the strength of the relationship between fields, the relationship between the fields in the learning data is noted and appears in the field group. Even if the relationship calculated based on all attribute values and their frequency distribution is low (the relationship is sparse), a “subset” of attribute values appearing on the field (or its frequency distribution, Based on a fixed threshold (pre-fixed, automatically calculated based on overall relationship strength, number of attribute types, etc., specified by the user, etc.) In the case of strongness, the relationship strength between “subsets” is adopted as the relationship between corresponding fields by using the evaluation score between fields or the presence or absence of the relationship between “subsets”.

  As a request cause when it is required to assign an order or evaluation score to the strength of the relationship between fields, for example, because there are many fields, it is difficult to consider the relationship between all fields in the process. Therefore, when it is necessary to extract related fields based on some criteria, there is a problem in considering the influence of other fields on a certain field equivalently, and ordering or scoring the influence of each field May be necessary.

  In other words, the partial relationship extraction processing performs a new relationship determination by recalculating the relationship determination processing based on the specified information criterion, focusing only on the above-mentioned “subset”, and the relationship between the entire field and the entire field When the field hidden in the field takes a specific value, the relationship with another field is extracted.

Between the fields from which the relationship is extracted, the relationship is avoided from disappearing due to the following two points.
1) Avoid the “summarization process” that is executed when there are a large number of low-frequency attribute groups.
2) Use an evaluation value based on the extracted partial relationship instead of the evaluation value of the whole relationship (assuming there is a relationship).

  However, the situation in which the process 1) is required is based on the premise that the number of attributes of the corresponding field is large, and in this state, it is inevitable that the model becomes complicated. Means for avoiding this will be described later.

  In this partial relationship extraction process, since there are a plurality of ways of taking “subsets”, there may be a plurality of “subsets” having an evaluation value that exceeds the evaluation value of the overall relationship. For this, for example, the following means are used to specify one.

1) Perform a full search on the designated search target and find the one with the maximum evaluation value.
2) A standard (absolute value, relative value to the overall relationship evaluation value, accuracy increase (estimated) value, etc.) is given in advance, and if a value exceeding this standard is found, it is adopted.
3) Use the first one found.

  As for which of the above methods to adopt, consider system characteristics (particularly time performance and usable virtual space size) and characteristics of data to be analyzed (number of fields, number of attributes, number of records). And decide.

  In the present specification, “record” means an element constituting data and indicates a certain state. For example, one horizontal line in FIG. 5 (information of one person to be surveyed) corresponds.

  There is also a method in which two or more of the above items are mounted, and a mechanism for instructing (inputting) a method that the user desires to use and a mechanism for selectively executing the instructed process are available.

  Furthermore, when two or more of the above items are implemented, it is not an instruction from the user (or if there is no instruction from the user), the time at which the relevant program (for this purpose) will operate. There is also a method in which a mechanism that determines a means to be used based on a system load (as a predicted value) and data characteristics is used, and a method determined by this mechanism is adopted.

The following conditions are also considered for the search order of “subset”.
1) The whole thing can be done comprehensively and in order.
2) When a “subset” having a large relational evaluation value is found, give priority to the direction of trying to find a larger evaluation value by giving a difference to this.
3) Look at the attribute value distribution of the field group and incorporate it into the subset in order from the smallest (the larger distribution is considered to have a large effect on the evaluation value of the overall relationship. Search).

Also in these respects, the method determining mechanism by the user instruction is effective.
Note that this function is meaningless when the previous search target is the entire area (all searches) (because it is executed in any case, only the execution timing is changed and the result is not affected).

  On the other hand, when the “subset” is adopted when a certain condition is satisfied, it functions effectively (and the inference performance is affected because it affects the structure of the output model). For this reason, a method of setting the search order of the “subset” when the previous search range is determined other than the full search is also conceivable.

  Here, a means for preventing the above-described model from becoming complicated will be described.

  First, as one means (first integration means), the constituent elements (field group attribute value combinations) of the extracted “subset” are classified into several set groups (for example, depending on their tendency), The attribute values are integrated (and used as new learning data if necessary).

  This means implements a reduction in the number of attributes without losing the relationship by integrating only the attributes where the similar relationship is established from the plurality of extracted relationships. It can be applied to the model generated by the extracted relationship.

  Prior to model generation, integration processing may be performed on learning data (known data), and model generation may be performed using the attribute integration processed data.

  Further, as another means (second integration means), the attribute values corresponding to the exclusive part of the acquired “subset” are collected into a constant value (and this is used as new learning data as necessary). This can be achieved. This also contributes to the construction of a simpler model by reducing the number of attributes.

  The difference between the other means and the above-described means is that they are collected together without confirming the similarity. Originally, the summarization process ignoring the similarity relationship corresponds to the process of extinguishing the relation held by the known data pointed out as a problem in the known function, but here, the attribute to be summarized is the above-described attribute. Therefore, there is no relationship that disappears due to the summary, and only the contribution due to the simplicity of the model due to the summary remains.

  As described above, the difference from the conventional method is the concept that the attribute value distribution of the field is not used as it is in determining whether there is a relationship, but the portion is used.

  Furthermore, in this embodiment, the processing range of the apparatus can be specified by an instruction from the outside. Therefore, in an environment where two or more relation strength calculation methods are implemented, each of a plurality of fields As a target, it is possible to specify (input) the relationship strength calculation method and the adoption conditions adopted for each field from the outside of the implementation module of this function (including user interaction) (corresponding to the first designation means) I have to.

  In addition, prior to the execution of the partial relationship extraction process described above, a list of attribute value groups that may belong to the “subset” can be specified (input) from outside the module that implements this function (including user interaction) ( It corresponds to the second designation means).

  In the present embodiment, model construction by a known calculation method has been ignored. That is, when viewed from the whole model, a small relationship is extracted to improve the inference performance of the corresponding part. In the previous example, when the model incorporating the present embodiment is used, the part where the inference performance is improved is limited to the part of a person from a country other than U.S.A. Of course, processing time is required.

  Therefore, there are the following cases where it is preferable to avoid the above-described processing in the present embodiment for the following reasons.

1) The existence of a relationship between fields that is clearly irrelevant (and already recognized by the user) with the purpose of model generation (the reasoning target in the reasoning process) has been clarified in advance (so that it can be placed in this part) If no extraction is required).
2) When high-performance inference performance is not expected for reasons such as shortening the processing time.
3) When the purpose of generating the model is to grasp the whole image and the influence on the minority part (local) is unnecessary.

  By making the above-described designation input possible, it is possible to designate a case where there is a reason for avoidance as described above and to perform efficient processing.

  As described above, in the present embodiment, the field groups existing in the learning data, the attribute groups held therein, the attribute distribution in addition to the above depending on the situation, the relationship strength between the fields, the specific attribute (group) in the field The function of presenting the relationship strength between the nodes and the function of designating from the outside regarding what part and under what conditions are provided.

  Even when this function is used, there are cases where it is desired to specify a summary range.

  In the previous example, the first integration unit is used in the partial relation extraction process described above to improve the prediction performance for a minority country, but the second integration unit should not be used in consideration of later statistical processing. If it belongs to other, this is the case, for example, when it is divided into five continents.

  In addition, when it is desired to convert known data to a condition that satisfies a certain condition (for example, the number of attributes is 5 or less), a plurality of summary functions including the present embodiment cannot be realized by a single unit, but only by combining a plurality of functions. If possible, a function for specifying the application order and application conditions (corresponding to the third specifying means) is desired.

  In the present embodiment, in an environment where two or more data summarization methods including the first integration unit or the second integration unit are implemented, the user uses (interactively) from outside the module for implementing this function. A summary method (and its application conditions, application order) can be selected. This presents a list of usable summary functions and provides a function for inputting the used functions together with their order and conditions.

(Embodiment 1)
In the following, as a first embodiment, reasoning of high-income earners using a personal information DB in the USA will be described.

  Let us consider an example of a good US (high-income customer) enclosure by a US bank. Suppose that he wanted to extract high-income earners from all customer groups in order to propose recommended products for the wealthy. In general, unlike address, age, etc., it is difficult to investigate the annual income of general customers, so it is considered to predict higher income people than existing information.

  Here, the personal information DB in U.S.A mentioned above is referred to as existing. This DB lists age, gender, height, weight, educational background, country of origin, family relations, and various other personal information along with annual income, and applies the customer information held by the bank to the annual income prediction model created using this By doing so, the annual income of the customer can be predicted. The DB may be created from a collection (such as loan customers) of which the annual income of the bank is known, or a publicly available survey result such as government statistics may be used.

<Preliminary explanation of relationships to be extracted>
When the relationship between the country of origin and the other requirements is confirmed here, as shown in FIG. 10, it is understood that there is a relationship with educational background, although it is intended for limited countries.

  For example, in some European and East Asian countries, the percentage of highly educated people is high, while in other European and South American countries, the percentage is low.

  In this way, some of the countries of origin have a certain impact on educational background, and educational background is a strong influence factor on annual income, so this relationship between the country of origin and educational background exists in the annual income estimation model. I think that is preferable.

<Reason for difficulty in extracting relationships with existing functions>
However, when calculated by normal processing, the strength of the relationship between birth and education is actually not so strong.

  First, because of the variety of country names, it is less likely that the influence of the field of origin on other elements will be adopted when modeling. As an existing technology for solving this problem, rounding processing from country to country group is known, but the judgment of rounding processing is left to the user, and the region name that is executed when simply considered The summarization processing by continents, rounding processing by continent names, etc. will crush the height information of education indicated by the country name in order to ignore the level of education (Asian and European) and convert it to a regional name.

  In addition, since the educational backgrounds of the overwhelming majority of USA citizens are seen, when referring to the whole, the influence of the educational background (deviation) of a small number of specific countries is small. It is low.

<Extraction method>
Here, it is assumed that the following is used as a function for measuring the relationship strength between fields A and B.

(Formula (1))
Relationship strength between AB: f1 (model evaluation with AB relationship, model evaluation without AB relationship)
Relationship strength threshold: D1 (total number of records, number of attributes: changes depending on the degree of penalty for model complexity, etc.)
Part affected between specific fields: f2 (total number of records, appearance frequency)
f1 = D1 + Σ_ (each fieldX) Σ_ (related measurement target fieldY) Σ_ (various combinations of XY attribute values) [f2]

  In the above equation, if f1 is greater than a certain threshold, it is recognized that there is a relationship between AB.

  Here, f1 is an evaluation formula that considers the relationship between fields held by the entire model, and is calculated as follows, for example.

(Use of a known model evaluation formula)
f1 = MDL (model with AB relationship)-MDL (model without AB relationship),
f1 = AIC (model with AB relationship)-AIC (model without AB relationship)

  Here, MDL (X) is an MDL evaluation value when the model is in the X state, and AIC (X) is an AIC evaluation value when the model is in the X state.

(Original setting)
f1 (total number of records, number of fields, number of attributes, appearance frequency of each attribute, ...)

  On the other hand, the partial relation strength is calculated by the following calculation.

(Formula (2))
Partial field impact calculation correction between specific fields: f2 '= w · f2 (subset record number, appearance frequency) + D2
Weight of each term: w (total number of records, number of subset records)
Adjustment value considering relation strength threshold D1: D2
f1 '= D2 + Σ_ (each fieldX) Σ_ (related measurement target fieldY) Σ_ (various combinations of XY attribute values) [f2']

  In the above, when f1 ′ is larger than a certain threshold, it is recognized that there is a relationship between AB.

  Note that D2 varies depending on 1) the importance of the strength of the partial relationship in the entire model, 2) when uniquely determined in advance, 3) the original model freedom, partial relationship partial freedom, partial relationship classification after In some cases, automatic calculation is performed based on the degree of freedom or the like (in the simplest case, an intermediate value is adopted).

  Instead of f2 in equation (1), f1 ′ is executed instead of relationship strength calculation f1 by using f2 ′ in equation (2).

  While the original f2 is calculated when the entire training data exists, the same evaluation formula is used and the calculation is performed assuming that only a “subset” of the training data exists. The relationship strength of the “subset” is calculated by applying a weight w based on the number of learning data and the number of “subset” records.

  In other words, if the evaluation is performed using the evaluation formula f1 given by the system, f1 ′ is calculated when f1 is small. Here, “w: weighting” and “D2: clogs” mean a correction function to prevent this local relationship from becoming too important compared to the whole, and thus the basic equation f1 is changed. Balance with the evaluation part used.

  It should be noted that f2 in equation (2) may be in accordance with the concept of the original relationship strength calculation with the following as an argument, and need not be the same function as f2 in equation (1). In the text, a simplified case is presented.

  Here, the following values can be taken.

Overall relationship strength: eg) Overall relationship: Country of origin field-Education field calculated for relationship strength.
Strength of relationship between partial value and overall value (group): Example) Relationship between partial value and overall value: (Japan + Korea)-f2 calculation with educational background Partial strength vs. partial value relationship strength (group): Example) Partial value vs. partial Relationship of values: (Japan + Korea)-(University graduate + high school graduate) f2 calculation Other (depending on the situation) Various: Example) If the purpose is determined as in the annual income model, etc., the purpose field (here annual income) Correlation degree and the like can be considered.

  Execute the above processing for each existing field (age, gender, educational background, country of origin, family relationship, etc.), and whether there is a relationship between a subset of field values and other field values (above the relationship strength threshold) (Or, the following judgment) is executed. The target field is determined separately (for example, all unconditionally are limited by user designation, and the field value distribution satisfies a certain condition).

  It should be noted that the extracted relationship takes into account a plurality of fields. In the above example, whether to extract the relationship “Japan + Korea”-“University graduate + High school graduate” will disappear by this relationship extraction “Japan”-“Family composition”, “Korea”-“Age” Etc. are judged in combination with evaluation of the relationship between other fields.

Below, according to the schematic flowchart shown in FIG. 2, the measurement of the strength of the relationship between the country of origin and field education is exemplified.
The operation of the embodiment of the present invention when the learning data shown above is given (step S1) is as follows.

  Here, in an environment where two or more relationship strength calculation methods are implemented, the relationship strength calculation method to be adopted for each field and the adoption conditions are designated from the outside for each of a plurality of fields. It is assumed that it is specified that the relationship determination focusing on this “subset” is limited to fields that are considered unrelated in normal calculation.

If it is a normal model generation function, the strength of the relationship between the fields is measured (step S2), and if a value equal to or greater than the threshold value is obtained (step S3), this relationship is present (step S4).
Here, we take the relationship between two fields, country of origin and educational background as an example.

  First, the relationship strength f1 of the whole (country of origin) versus the whole (educational background) is measured using f2. If it is equal to or greater than a certain threshold (simply 0 is adopted here), the corresponding relationship is set as an output target without continuing the following processing (after step S8).

Next, a known relationship strength evaluation is performed, and when the value is less than a certain value (step S3, n), it is registered that partial relationship extraction processing is performed for all relationships.
For this reason, if it is less than a certain value (here, 0), the following processing is continued.

  For example, the relationship between the country of origin from the field and the educational background of the field is considered to be weak, so the relationship strength is less than or equal to the threshold value in the above existing function, and therefore, the step of measuring the relationship strength of this “subset” is entered.

  Next, it is confirmed whether there is a useful “subset” having a certain relationship strength, in other words, whether there is a “subset” in which f2 ′ is larger than f2 (step S5, FIG. 3). . Here, FIG. 3 shows details of step S5 (inter-field attribute subset relation strength measurement processing) of FIG.

  In FIG. 3, initialization is performed (step S51), a candidate list constituting a subset is extracted (step S52), and a subset having the maximum evaluation is extracted (steps S53 to S56).

  Thereby, the evaluation value of the relationship is recalculated, and the relationship order between fields is determined. When the evaluation of the relationship after recalculation is equal to or greater than the threshold value or within a certain order, the relationship between fields is determined to be present (step S6, y) and recorded (step S7).

  Here, although the overall relationship between country of origin and educational background is low, there are deviations in the high percentage of highly-educated personnel in Taiwan, India, etc., and low educational background in Mexico, Guatemala, Portugal, etc. If you select the set = "Taiwan, Iran", "India, France", "Mexico, Guatemala, Portugal", ... ", it will be found that f2 <f2 ', so the relationship strength will be corrected A new relationship is established. The generated model will include the relationship between country of origin and educational background.

  In this embodiment, the constituent elements of the subset can be limited in advance.

  For example, because U.S.A has a very high frequency of appearance, it may be desirable to make it independent (even if the relationship is strengthened by including it in the subset). In this case, a list of attribute value groups that may belong to the subset is designated, and for example, the fact that U.S.A is excluded from the constituent elements of the subset is registered. Then, with reference to the registered exclusion target list, the relationship evaluation (f2 ′) calculation of the subset including U.S.A is omitted.

  Thus, consider the situation where the following “subset” has been extracted.

  "Taiwan, India, Iran, France, Mexico, Guatemala, Portugal, El Salvador, Japan, United Kingdom, Hungary"

  This is classified based on the similarity of the distribution state. It should be noted here that the embodiment of the present invention considers the relationship between a plurality of field groups, and evaluates items other than the country name and educational background that are focused on here. is there.

  Therefore, the relation extracted in the partial relation extraction process (partial relation strength calculation) is classified in such a way that the relation existing with other field groups does not disappear, and "" Taiwan, Iran "," India " , France ”,“ Mexico, Guatemala, Portugal, El Salvador ”,“ Japan, England, Hungary ”.

Based on the extracted relationship, the field relationship is recorded (step S8), and then, as shown in FIG. 12, it is classified, integrated, and renamed (step S9).
Here it is as follows.

Mexico, Guatemala, Portugal, El Salvador: relate1
Taiwan, Iran: relate2
India, France: relate3
Japan, UK, Hungary: relate5
Nicaragua, Colombia, Ecuador: relate6
other

  In many cases, the number of attribute values does not decrease even with the above classification. In this example, as shown in FIG. 12, the number of country attributes is only halved from 40 to 20.

  In FIG. 12, the color coding indicates the distribution of educational background, and the countries of origin are summarized in 21 countries. In the figure, relateX is a collection of multiple country names that have the same tendency to be related to other fields. It can be seen that countries with high educational backgrounds are grouped as “relate2” and “relate3”, and countries with low educational backgrounds are grouped as “relate1”. That is, relate2 and relate3 are similar at first glance, but are classified as a result of internal judgment (judgment based on relationships with other fields).

  Here, this processing is the result of the summary processing of the partial relationship extraction processing, but in fact, prior to the use of the processing or similarity calculation processing, the USA uses a “subset” to specify the target attribute value group. However, there are countries that have been excluded from consideration of the summary process even though they are similar. In addition, there is a case where the summarization process is denied because it is similar if only the educational background is seen, but the relationship by another summarization method in other field groups is larger than that.

  Therefore, if it is desired to further reduce the number of attributes, those that are excluded from the above integration process (however, those that should have existed independently due to their unique nature, The number of attributes can be reduced by collecting “others” except those designated by the user as “others” (step S11).

  Note that the execution load of the above processing is basically equivalent to acquiring the brute force correlation of the value of each attribute of each field, and in actual operation, it is better to make a certain prediction in advance and limit the search range desirable.

  This means includes the following means.

1) In addition to directly limiting the range of fields that are known to have no relationship, the user directly determines the range of use and determines the relationship strength by using existing relationships such as clustering and trees. Specify the method and adoption conditions and exclude it from the “subset” calculation.
2) If there is a record number above a certain level, even if it is selected as “subset” in the strength calculation of the sub-relationship, it is not a target of the similarity calculation of the sub-relationship. By specifying (inputting) a list of attribute value groups that may belong to, it is possible to save labor in similarity calculation.

3) In addition to the partial function calculation function for obtaining an exact solution, a relational determination function with a lower calculation load (for example, MBR impact calculation) is prepared, and the partial function is used only when a value within a certain range is taken with a function with a lower calculation load. Designates that the strength is calculated (recognizes that there is nothing else), and realizes labor saving of similarity calculation.
4) “Subset” extraction by partial relation strength calculation, similarity calculation of partial relation (summarization processing), and summarization processing by exclusive part of extracted “subset”, the number of attributes of all fields is 5 or less Repeat until the relation strength threshold D2 is changed within a certain range. If the purpose cannot be achieved even if D2 falls outside the allowable range, the number of attributes can be forced by specifying that the summarization by the frequency distribution for the initial data (top 4 attributes + other) is executed. Perform reduction.

<Influence on inference processing (expected impact)>
The relationship model extracted above has a different structure from the model based on the normal relationship strength, and the relationship exists locally (between specific attributes) between fields that are considered to be less related in normal relationship measurement. And this relationship can be displayed on the model (step S11).

When this model is used for inference processing, highly accurate inference using the extracted relationship is possible.
As explained earlier, when it comes to the relationship between origin and educational background, the relationship is actually not so strong.
This is due to variations in the educational background of the majority of Americans. In addition, according to the frequency distribution, in the case of USA and other cases, the composition is made up of a high-education USA citizen and a low-educated other person by averaging.

  In reality, however, there is a bias in education among people outside U.S.A, especially those that are very highly educated compared to the U.S.A average. When inferring items that have an educational background, existing methods cannot recognize this correctly.

  In this case, the influence cannot be correctly determined by simply recalculating the strength with the low strength element deleted. This is because the influence of the number of samples is distorted. This function extracts this relationship, and using it, for example, makes it possible to more accurately infer the annual income of a person with unknown educational background or the educational background (using the country of origin).

  The model created by the initial relationship strength calculation (Fig. 8) had a weak relationship between birth and annual income. On the other hand, in the model created according to the present embodiment, as shown in FIG. 13, the relationship of birth → education → annual income is expressed, which makes it possible to estimate the annual income of customers who are not from USA more accurately. As a result, it will be possible to more reliably propose high-income products for high-income customers.

That is, in FIG. 13, since a group with a high educational background is extracted as “relate2” and a group with a low educational background is extracted as “relate1,” the influence due to the difference in educational background remains on the model. In Figure 7, the country of origin contributed little to the annual income (the relationship was not linked). Further, FIG. 10 expresses a useless relationship. Here, due to the correct relationship with educational background, the effect on the annual income is great, which helps to improve the reasoning performance (numbers are not mentioned here).
(Example)

  Hereinafter, with respect to the embodiment, a more detailed understanding of the distribution of high-income earners in U.S.A will be described. This is to make a more accurate grasp of the high-income earner (determining whether or not the customer is a high-income earner) when “recommending” a high-priced financial product to the customer.

  Unless you are asked to manage your assets, it is difficult to understand the customer's asset status. However, since asset status tends to be related to educational background and occupation, asset status and the accumulation of other information that is relatively available (this is used as data) can be used to predict asset status and suit high-priced products. Think about what you'd recommend to your opponent.

  For the above purpose, an embodiment will be described in which the present embodiment is applied to an annual income prediction apparatus and the relationship between the asset state and other information is extracted.

<Acquired information>
Here, for the sake of simplification, consider the case where customer information is composed of race, origin, annual income, age, and number of years of education (hereinafter, the above five are called fields). In addition, although the annual income, age, etc. are numerical values, the annual income is assumed to be two types of “high, low”, and the age and the number of years of education are classified into five types.

  The source of this information is assumed to be something like a census. Then, using such information, consider the construction of a system that predicts annual income when the race, origin (nationality), age, and number of years of education are known. Using this result, the bank can increase the operating efficiency by explaining the product only to customers who are predicted to have “annual income = high” (customers who can buy high-value financial products).

<Known system outline>
Here, as a model showing the relationship between annual income and other than that, here is considered a Bayesian Model (which is determined probabilistically from the value of another field related (= limited) related to the value of one field).
For model construction, the following model is adopted.

1) Evaluation criteria: BDe remodeling 2) Addition of relationship: When there is a relationship between fields, when the evaluation of the whole model goes up, that relationship is adopted.
3) Model growth method: Growing into Greedy (This is to get a solution close to the optimal solution in a short time)
4) Stop growth when there is no longer a relationship that raises evaluation

  According to the evaluation criteria (BDe remodeling), when the value of one fieldA depends on the value of another fieldB and C, B and C are called the parents of A. Based on this parent-child relationship, The evaluation value of the entire model is determined by the calculation formula of 14 (BDe score calculation method error! Reference source is not found.). It is basically an analog of AIC. Compared with MDL, it can take a relatively fine relationship in its original state. This is because the penalty for model complexity is low.

  However, when this calculation formula is simply calculated, problems such as floating point number overflow occur. Therefore, this time, a value using the “BDe remodeling formula” shown in FIG. 15 is presented. Actually, in order to obtain a more correct value, the log addition part may be devised.

<Problem of known system>
Since model generation in existing systems summarizes trends in the data used, the characteristics of a small number of parts are often ignored.
The relationship extracted when the model is constructed by the calculation formula shown in the “BDe remodeling formula” shown in FIG. 15 is as follows.

Initial state: Each field (node) is completely independent
Add relationships to Greedy
Number Node
node [0] = race
node [1] = from
node [2] = annual income
node [3] = age-div5
node [4] = years of education-div5

      Calculate the evaluation value.

(All elements 47902)
[child, parent]: = [1, 0] (The first one found was node [1] as a child of node [0])
add scoreDiff_i = 4911.672547034621 (This increased the evaluation value by 4911)
[child, parent]: = [3, 2] (The next one found was node [3] as a child of node [2])
add scoreDiff_i = 4120.6223854812415 (Evaluation value increased by 4120)
[child, parent]: = [4, 2]
add scoreDiff_i = 4092.57636431673
[child, parent]: = [3, 4]
add scoreDiff_i = 1416.8170142548188
[child, parent]: = [2, 0]
add scoreDiff_i = 330.7434124116626
[child, parent]: = [4, 0]
add scoreDiff_i = 87.20574294828111

  Here, since the evaluation value exceeding the relation strength threshold D1 = 0 (and this time setting) is not found, the model generation is stopped.

  Score f1 = 4911 + 4120 + 4092 + 1416 + ... (score difference between all nodes completely independent state and final state)

  After all, the model is as follows.

node [0],,,,, (node [0] has no parent)
node [1] x,,,,,
node [2] x,,,,, (node [2] is parented by node [0])
node [3],, x,, x, (node [3] is the parent of node [2] and node [4])
node [4] x,, x,,,

  Thus, node [2]: annual income is related to node [0]: race, node [3]: age, node [4]: years of education.

  This relationship is statistically natural considering that the majority of the constituents are from the USA, but the tendency of a small number of foreigners is ignored (Figure 16). Then, node [2]: In the prediction of annual income, first, the node [0] race is judged (white is high and colored race is low). If the race is unknown, node [2] is used to estimate the race. 1] Although the country can be used, for example, Taiwanese people are colored, and colored people are treated as having low annual income (in fact, they are biased toward high income).

<Improvement by Embodiment of the Present Invention>
By the way, when looking at the relationship between the country and the annual income, there are actually countries where there are many high-income people (see “Relationship between the country and the annual income”). From this, it can be inferred that it is better to use the country of origin in the prediction of annual income.

  However, because only 5fields were used here, the person was able to grasp the relationship between the country of origin and the annual income, but it is difficult to judge this from a large amount of data. Consider extracting this minority characteristic. At this time, attention should be paid so as not to ignore the entire system focusing on a small number.

  Here, the evaluation and correction for the subset are considered. It is confirmed whether or not there is nothing really related to what is not related by a known method. Since it takes a long time to obtain an optimal solution by brute force, some priority is given to the search. It may wait for an instruction from the user or may be limited mechanically. When it is executed mechanically, priority is given to a highly dispersed relationship between the two fields (if the relationship between the country and race does not come out in a known manner, the dispersion is severe = only whites depending on the country, There are countries such as black and black only, so it will be the first to be tested), judging from the purpose (in this case, because the annual income forecasting machine is created, the relationship between annual income and certain attributes is given priority) .

  You can check only the priority (XX from the beginning) without considering the processing time, or you can get the specification of the required time and execute in order as much as possible within the time. May be.

  Here, the example shows the relationship between the country and annual income. Actually, the check is made using the above evaluation formula for multiple types such as country, annual income, and race.

  {Country: America, Canada, Brazil, ...} and nothing else related to race. That is, the evaluation value was small. Then, it is considered whether there is no relationship in the part set.

The subset considered here is {country: groupA, GroupB, GroupC}.
All countries belong to one of A, B, C, D (not considered). The maximum number of Gtoups is the number of elements (here, the number of countries), and the minimum is 2.

  First, decide whether to check all subsets, or prioritize some. There are the following means.

A1) User specified
A2) Dispersion-dependent machine processing: Average country is extracted in relation to annual income (naturally USA here), and it is classified into those that are within XX% and those that are not (up and down)
A3) Number-dependent machine processing: Focus on the effects of a small number (in this case Hungary, Honduras, ...) because a large number should affect known models
A4) Combination of the above

  There is no point in putting all values into the same set. As shown above, the BDe evaluation formula looks at the change in probability that the child value is determined when the parent value is determined to be a specific value (deviation from the average distribution). There is no point in being (100% is 100% no matter how it changes).

  So we will go through the evaluation of the subsets in turn. Here, A3 is adopted as the search order, and A2 is used to determine grouping.

1) [(minimum) Hungary (31.58% distribution) and similar (with a difference of XX%), otherwise]
Here, change XX and repeat. Here, since the average is 23.9, (31.58-23.9) / 5 * n: n = 1 to 5 is repeated 5 times. Remember the group that corresponds to XX with a high rating xx.

2) [Honduras (10% distribution) and similar (with YY% difference), otherwise]
Here, repeat YY. However, those similar to the above-mentioned another (here, only hangar) are not treated. Again, remember YY with the highest yy rating.

3) [Group that issued xx, group that issued yy (do not duplicate), otherwise]
[Group with xx, group with yy (not duplicated)]
For the first time, since the number of elements exceeds 2 (minimum required number), “other than that” can be selected as out of consideration.

  In this case, since the number of populations is reduced by “other than that”, Nij in the previous mathematical formula changes, and if it is incorporated into a known method as it is due to a change in the denominator, it becomes a problem. A correction coefficient (raising by getter) is required.

4) [Group with xx, group with yy, group with zz, etc.]

The calculation is performed in the same manner as known. The following is an example.
{Country: Above grouping}, [Race], [Annual income], etc. First, the evaluation is re-executed normally (in order of the above groups).

(All elements 47902)
[child, parent]: = [0, 1]
add scoreDiff_i = 2821.2557579191407
[child, parent]: = [1, 4]
add scoreDiff_i = 541.032507935146 / * Evaluation value of relationship between birth and years of education in a certain state * /
[child, parent]: = [4, 2]
add scoreDiff_i = 438.6759072416644
[child, parent]: = [2, 3]
add scoreDiff_i = 234.67135951875343
...
Score = 2881 + 541 + 438 + 234 +…

The model is
node [0] = race
node [1] = from
node [2] = annual income
node [3] = age-div5
node [4] = years of education-div5
node [0], x,,,,
node [1],,,, x, / * Country of origin is related to years of education * /
node [2],,, x,,
node [3],,,,,
node [4],, x,,,

  Next, a new relationship (country of origin and years of education: score 541) is extracted.

By the way, since this is an evaluation by looking at the part, in the situation of (1), it becomes a factor that destroys the whole relationship (only seeing too fine).
If you simply incorporate the relationship of score541 into the original model, you will see something too fine.

Adopting the new model will now ignore the relationship seen in the original model (this can also be seen from the fact that the relationship between annual income and the number of years of education falls from 4092 to 438. In the worst case, it disappears. == Below threshold value = In this case, it may be 0 or less).
Therefore, the use of the relationship is determined after correcting the new relationship of the new model.

  Although it may be adopted when the user finds a favorite relationship (a relationship expected from previous experience), in many cases, automatic determination is performed. In that case,

Correction formula
Score f1 '= D2 + w (upper score group)

  Use the one that has a large f1 ′, the one that is above a certain level, the one that was first discovered, the one that is above a certain level, and that meet the conditions specified in advance.

  Here, w and D2 are correction conditions that are determined according to the tendency of the data and the user's desire (how much detailed relationship to take). “w” indicates how important the relationship is in the scale of this subset, and indicates, for example, how much the relationship is compared to the maximum strength relationship. In the above example, the simplest decision method is

  w (Upper score group) = Ratio when compared with the maximum relationship * Size of the corresponding model = 541/2821 * 4911 * (47902/47902) = 0.19 * 4911 * 1 = 933

  This is the importance of the relationship that appears when a certain subset is adopted from the perspective of that subset.

For example, in the above example:
Years of education → Origin → Race --- (A)
In this relationship, origin score = 541.032507935146
W = 933

  , But if we calculate for another subset,

(All elements 28543: There are about 19000 records that are not calculated)
[child, parent]: = [2, 1]
add scoreDiff_i = 4115.2899534634635
[child, parent]: = [1, 3]
add scoreDiff_i = 4091.6699646188354
[child, parent]: = [4, 0]
add scoreDiff_i = 2135.7989910176548
[child, parent]: = [2, 3]
add scoreDiff_i = 1412.705668559298
[child, parent]: = [4, 3]
add scoreDiff_i = 883.041889777498
[child, parent]: = [0, 1]
add scoreDiff_i = 325.8174577299578
[child, parent]: = [0, 3]
add scoreDiff_i = 93.33244284622924
node [0] = race
node [1] = annual income
node [2] = age-div5
node [3] = years of education -div5
node [4] = from
node [0], x,, x,,
node [1],,, x,,
node [2], x,, x,,
node [3],,,,,
node [4] x,,, x,,
Race → Birth & Years of Education → Birth (In short, parents are two parents) --- (B)
Origin score = 2135 + 883 = 3018
W = 3018/4115 * 4911 * (28543/47902) = 0.733 * 4911 * 0.595 = 2141

If w is strong, it is the case of the latter (B) that makes w strong, and if it is included, it is determined that the relationship (B) should be adopted.
Apart from this, the problem of whether or not this should be put into the original relationship arises in the first place.

  In order to solve this problem, the concept D2 is introduced and used as a footing material. The complexity of the model obtained by D2 changes.

  Although it is possible to determine a fixed value and register it, create multiple models using multiple D2s (a small number that can be compared and evaluated by humans) and allow users to select or use as a predictor If this is the case, it is considered appropriate to perform cross-validation using learning data given in advance using each model and adopt the one with the best performance. Simply compare the strength of the relationship with the strength of the original model,

f1 '= D2 + W, f1> max (maximum relationship evaluation value of the original model)
D2 = k * max ((maximum relational evaluation value of the original model), k = 0, 0.1, 0.2,… 0.9

  A means such as the above can be considered.

  In this case, for simplification, it was the country that was the target of subset creation, and distribution was considered only for annual income, but in general, combinations of subsets or distribution-considered sets were considered. Another distribution is also considered.

(Embodiment 2)
Hereinafter, as Embodiment 2 of the present invention, a disease name integration process in medical office information will be described as a simple and easily recognizable medical management model presentation. FIG. 4 is a flowchart showing the main part of the second embodiment.

<Data overview>
Present a medical management model.
The final goal will be management analysis (predicting occupancy rate, profits, etc.).

  For the purpose of this analysis support, here is an example of a part of the process of creating an influence model for the number of days required for hospitalization (the number of days remaining after hospitalization given information) (== target is defined as the number of days remaining in hospital).

  In general, it is believed that the most influential factor on medical expenses and hospitalization days is the name of the disease, which is no mistake. However, there are a wide variety of disease names, and when analyzing patient trends in a single hospital, there is no way to obtain statistics for each disease name here based on the information accumulated throughout the year. Therefore, it is necessary to collect them into a certain level of classification. In this analysis, management problem = day relation = is considered, so classification by the same case is not essential.

  As this classification, the diagnosis department, international disease classification, etc. can be considered, but the diagnosis department does not necessarily summarize the name of the disease (complex disease, existence of pediatrics), and the classification is classified according to the problem site. Because it is difficult to measure the weight, it is close to the number of disease name classification in the small classification, and the statistically meaningful number cannot be collected in the single hospital aggregation, so an original way of summarizing is necessary.

  Moreover, since there are very many types of medicines and treatments, it is effective to reduce the common number for the purpose of assisting recognition by the user.

  Therefore, it is considered to make it easier to grasp the current state by compressing these various field value types, creating a simpler model, and presenting it to the user.

  The integration target is a disease name, a drug used (history), and a treatment (history), and is acquired from the learning data database (step S21).

  Here, the relationship strength calculation method and the employment conditions adopted for each field are specified, and the subset extraction target is limited to the disease name, the medicine used, and the treatment (step S22). In extracting relationships, all fields (field groups such as age, sex, hospital room location, etc.) are considered, but for example, age and sex are not subject to subset extraction.

  The extracted relationship will be, for example, [Pneumonia A, Drug B, Treatment C, Age], and [Pneumonia A, Drug B, Treatment C, Older] etc. will control only a specific part regarding age, It is not extracted here.

  Regarding the field group {disease name, drugs used, treatment}, the relationship between each attribute and the other field group is judged. Regarding the field group {age, sex, hospital room location}, each field and each other field group And the subset extraction is executed (steps S23 to S25).

  Adopting the similar situation of the number of days remaining on the hospitalization lapse date as the relation f1 in the partial relation strength calculation (partial relation extraction), partial relation extraction is performed, and this is attribute integration processing using a subset component (similarity calculation) Is executed (step S26).

  Using the similarity calculation, it is possible to extract that there are multiple disease names called “pneumonia”, but only some of them have similarities in the tendency of other fields (for example, injections used exclusively) The trend of hospitalization days varies depending on the period of use of internal medicines). Since this is summarized as related1, etc., it is renamed pneumonia A etc. and presented to the user.

  In addition, regarding the disease name group belonging to the above pneumonia A, it is possible to perform the similarity calculation inside the model creation process as in the case of the country name integration described above, but this is a general medical knowledge Can specify a list of attribute value groups, specify a group of disease names that will be summarized in pneumonia A, and determine whether or not the relationship extraction by the summary processing of the specified disease name group is valid on the corresponding data Confirm by using the method of strength calculation (partial relationship extraction) or subset component (similarity calculation) (when the specified disease name group is assumed as a subset in partial relationship strength calculation (partial relationship extraction)) Confirm that the relationship strength is stronger than before, or that a subset of all the subsets is created using a subset component (similarity calculation)), etc. The method is also conceivable.

After this, the same procedure may be repeated to create pneumonia B and pneumonia C, or it may be grouped as `` other pneumonia '' using a function that combines the attribute values corresponding to the exclusive part of the subset into a constant value. Also good.
By creating a model using the collected data, a simple model reflecting the knowledge can be obtained.

  As described above, according to the embodiment of the present invention, by extracting the relationship partially established between fields in the learning data and adopting it in the model, this was measured as an influence on the entire model. In some cases, a small relationship but a large relationship when used as a specific field relationship can be used for the inference processing, thereby improving the inference processing performance.

  Even if it is judged that the relationship between fields is sparse due to the relationship extracted by counting the existing mechanical and simple data count (it is judged that it is not necessary to adopt as a model component), There is a situation where the data relationship becomes strong when focusing on specific attribute values between fields.

  According to the embodiment of the present invention, it is possible to extract a relationship between specific attribute values in such a situation, and to build a model with the extracted relationship, thereby using the model. Can improve inference performance.

  In addition, the following is illustrated as what corresponds to the model here.

(Bayesian Network construction)
A network is constructed by combining related field groups, and inference processing is executed based on this network.

Therefore, inference performance depends on this network, and relationship extraction becomes important.
At this time, it is not desirable to define the relation of all fields as a brute force in terms of model or inference processing using the model (required memory amount and required time).

  In this generally known network construction method, a determination based on a certain amount of information criterion is performed, the strength of the relationship between the two fields is determined based on this determination, and only the strong one is valid.

  Currently, MDL, AIC, etc., which are generally used as information standards for BN construction tools, calculate using the entire distribution of attribute values appearing in the corresponding field. Not performed.

  The MBR determines the influence degree of the corresponding field (strength of the relationship with the target value) based on the attribute value distribution of the field group and the distribution of the target field value (MBR influence degree determination).

  In this determination process, if a part of the field value distribution is excluded from the calculation target or is processed into a specific value, a value different from the currently output impact level is calculated and different. An inference result is obtained.

  If the relationship between the target value calculation part of the original field and target is sufficiently strong, a more accurate result may be obtained.

  In addition, there are a large number of fields whose relations are unknown, and only a group of related fields is extracted from the fields, and a function expression using these as parameters is constructed (variable extraction in the model expression). Also in this parameter determination, it is possible to extract a relationship only for a specific attribute value group between fields.

In the above-described embodiment of the present invention, the steps shown in the respective flowcharts are stored as a structure analysis program in a computer-readable recording medium, whereby the structure analysis method can be executed by the computer. In the present invention, the computer-readable recording medium is a portable storage medium such as a CD-ROM, a flexible disk, a DVD disk, a magneto-optical disk, an IC card, a database holding a computer program, or other Computer and its database, and also a transmission medium on a line.
(Supplementary note 1) A partial relation extraction device between attributes in a model creation device that constitutes a predetermined model based on attribute values in a plurality of fields each of a plurality of data and the appearance frequency of the attribute values,
An overall relationship strength calculating means for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution;
Based on a subset of attribute values appearing in the field group for the predetermined attribute, partial relationship strength calculating means for calculating the strength of the partial relationship between the fields to which the attribute value belongs;
An inter-attribute partial relationship extraction device comprising: inter-field relationship setting means for determining a relationship between fields based on the strength of the partial relationship between fields calculated by the partial relationship strength calculating means.
(Additional remark 2) In the partial relationship extraction apparatus between the attributes of Claim 1,
The partial relation strength calculating means calculates the strength of the partial relation when the strength of the whole relation calculated by the whole relation strength calculating means is a predetermined threshold value or less. Extraction device.
(Supplementary Note 3) In the partial relation extraction device between attributes described in Supplementary Note 1 or Supplementary Note 2,
The partial relationship strength calculation means includes a subset acquisition means for acquiring a subset of the attribute values,
The partial relationship extracting unit is a partial relation extracting device between attributes, wherein the subset acquisition means performs a full search on a search target and acquires a maximum evaluation value.
(Supplementary Note 4) In the partial relationship extraction apparatus between attributes described in Supplementary Note 1 or Supplementary Note 2,
The partial relationship strength calculation means includes a subset acquisition means for acquiring a subset of the attribute values,
An apparatus for extracting a partial relationship between attributes, wherein the subset acquisition unit performs a full search on a designated search target and acquires a component having the maximum partial relationship strength.
(Supplementary Note 5) In the partial relation extracting device between attributes described in Supplementary Note 1 or Supplementary Note 2,
The partial relationship strength calculation means includes a subset acquisition means for acquiring a subset of the attribute values,
The apparatus for extracting a partial relationship between attributes, wherein the subset acquisition means acquires a partial relationship having a strength higher than a predetermined reference value.
(Supplementary note 6) In the partial relation extraction device between attributes described in any one of supplementary notes 1 to 5,
The inter-field relationship setting means, when the strength of the partial relation between fields calculated by the partial relation strength calculating means is greater than or equal to a predetermined threshold, the relation strength between the subsets as the relation between the fields A partial relation extracting device between attributes, characterized in that an evaluation point between fields is used, or whether or not there is a relation between subsets.
(Supplementary note 7) In the partial relation extraction device between attributes described in any one of supplementary notes 1 to 6,
The field group attribute value combinations that are constituent elements of the subset in which the strength of the partial relationship between fields calculated by the partial relationship strength calculating means is equal to or greater than a predetermined threshold are classified into a plurality of set groups, and each classification An apparatus for extracting a partial relationship between attributes, comprising first integration means for integrating attribute values in units.
(Supplementary note 8) In the partial relation extraction device between attributes described in any one of supplementary notes 1 to 7,
A second integration unit that integrates attribute values corresponding to an exclusive portion of the subset in which the strength of the partial relationship between fields calculated by the partial relationship strength calculation unit is equal to or greater than a predetermined threshold, to a constant value; A device for extracting a partial relationship between feature attributes.
(Supplementary note 9) In the partial relation extraction device between attributes described in any one of supplementary notes 1 to 8,
An apparatus for extracting a partial relationship between attributes, characterized by comprising a first designating means for designating a relation strength calculation method adopted for each field and a adoption condition for each of a plurality of fields.
(Supplementary note 10) In the partial relation extraction device between attributes described in any one of supplementary notes 1 to 9,
An apparatus for extracting a partial relationship between attributes, comprising: a second designating unit capable of designating a list of attribute value groups that may belong to the subset.
(Additional remark 11) In the partial relationship extraction apparatus between the attributes of Additional remark 7,
The apparatus for extracting a partial relationship between attributes, wherein the first integration unit includes a third specification unit that can specify an application condition or an application order in integration.
(Additional remark 12) Between attribute which makes a computer perform the partial relationship extraction method between the attributes in the model creation apparatus which comprises a predetermined model based on the attribute value in the some field which each of some data has, and the appearance frequency of this attribute value A partial relationship extraction program of
An overall relationship strength calculating step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution;
A partial relationship strength calculating step for calculating the strength of the partial relationship between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute;
An inter-attribute partial relationship extraction program comprising: an inter-field relationship setting step for determining a relationship between fields based on the strength of the partial relationship between fields calculated in the partial relationship strength calculating step.
(Additional remark 13) In the partial relationship extraction program between the attributes of Claim 12,
The partial relationship strength calculating step calculates the partial relationship strength when the strength of the overall relationship calculated by the overall relationship strength calculating step is equal to or less than a predetermined threshold. Extraction program.
(Supplementary Note 14) In the partial relationship extraction program between attributes described in Supplementary Note 12 or Supplementary Note 13,
The partial relationship strength calculation step includes a subset acquisition step of acquiring a subset of the attribute values,
The subset acquisition step is a program for extracting a partial relationship between attributes, characterized in that a search is performed for all search targets and the largest evaluation value is acquired.
(Supplementary Note 15) In the partial relationship extraction program between attributes described in Supplementary Note 12 or Supplementary Note 13,
The partial relationship strength calculation step includes a subset acquisition step of acquiring a subset of the attribute values,
The subset acquisition step is a program for extracting a partial relationship between attributes, characterized in that a full search is performed on a designated search target, and a component having the maximum partial relationship strength is acquired.
(Supplementary Note 16) In the partial relationship extraction program between attributes described in Supplementary Note 12 or Supplementary Note 13,
The partial relationship strength calculation step includes a subset acquisition step of acquiring a subset of the attribute values,
The partial set extraction step is a program for extracting a partial relationship between attributes, wherein a subset having a partial relationship strength exceeding a predetermined reference value is acquired.
(Supplementary Note 17) In the program for extracting a partial relationship between attributes described in any one of the supplementary notes 12 to 16,
The field relationship setting step includes a relationship strength between the subsets as a relationship between the fields when the strength of the partial relationship between fields calculated by the partial relationship strength calculation step is equal to or greater than a predetermined threshold. A program for extracting a partial relationship between attributes, characterized by defining an evaluation point between fields or determining whether or not there is a relationship between subsets.
(Supplementary note 18) In the partial relation extraction program between attributes described in any one of supplementary notes 12 to 17,
The field group attribute value combinations that are constituent elements of the subset in which the strength of the partial relationship between fields calculated by the partial relationship strength calculation step is equal to or greater than a predetermined threshold are classified into a plurality of set groups, and each classification A program for extracting a partial relationship between attributes, comprising a first integration step of integrating attribute values in units.
(Supplementary note 19) In the partial relationship extraction program between attributes described in any one of Supplementary notes 12 to 18,
A second integration step of integrating attribute values corresponding to exclusive portions of the subset in which the strength of the partial relationship between fields calculated in the partial relationship strength calculation step is equal to or greater than a predetermined threshold value to a constant value. A program for extracting partial relationships between feature attributes.
(Supplementary Note 20) A method for extracting a partial relationship between attributes in a model creation device that constitutes a predetermined model based on attribute values in a plurality of fields included in each of a plurality of data and the appearance frequency of the attribute values,
An overall relationship strength calculating step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution;
A partial relationship strength calculating step for calculating the strength of the partial relationship between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute;
A method of extracting a partial relationship between attributes, comprising: a step of setting a relationship between fields that determines a relationship between fields based on the strength of a partial relationship between fields calculated by the partial relationship strength calculation step.

It is a block diagram explaining the concept in embodiment of this invention. It is a flowchart which shows the operation | movement of Embodiment 1 of this invention. It is a flowchart which shows the operation | movement which performs a subset relationship evaluation (subset similarity calculation). It is a flowchart which shows the operation | movement of Embodiment 2 of this invention. It is a figure which shows the example of data of the personal information in U.S.A. It is a figure which shows the origin distribution in the personal information model in U.S.A. It is a figure which shows a U.S.A. personal information model. It is a figure which shows the origin country distribution in the personal information model in U.S.A. It is a figure which shows the country of origin and educational relations in the personal information model in U.S.A. It is a figure which shows the origin and educational background relationship in a U.S.A. personal information model. It is a figure which shows the country of origin and educational relationship in the U.S.A. personal information model obtained in this Embodiment. It is a figure which shows the origin country and educational relationship in the personal information model in U.S.A. in this Embodiment. It is a figure which shows the personal information model in U.S.A. in this Embodiment. It is a figure which shows the BDe score calculation method. It is a figure which shows a BDe remodeling image. It is a figure which shows the origin-annual income in a well-known system. It is a figure which shows the relationship between the country in this Embodiment, and annual income.

Explanation of symbols

  1 learning data input mechanism (DB), 2 model learning mechanism, 3 evaluation criteria calculation mechanism (overall relationship strength calculation means), 4 evaluation criteria calculation mechanism (partial relationship strength calculation means), 5 relationship strength similarity calculation mechanism ( (First and second means), 6 partial relation calculation object / condition designation mechanism (first, second, and third designation means), 7 inference mechanism.

Claims (8)

An apparatus for extracting a partial relationship between attributes in a model creation device that constitutes a predetermined model based on an attribute value in a plurality of fields included in each of a plurality of data and an appearance frequency of the attribute value,
An overall relationship strength calculating means for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a predetermined attribute and its frequency distribution;
A partial relation strength calculating means for calculating the strength of a partial relation between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute;
An inter-field relationship setting means for determining a relationship between the fields based on the strength of the partial relationship between fields calculated by the partial relationship strength calculating means ;
The partial relationship intensity calculating means, partial relation extraction device between attributes strength of the calculated overall relationship by the overall relationship strength calculation means that to calculate the strength of the parts involved in the case of less than a predetermined threshold value. In the partial relationship extraction device between attributes according to claim 1,
The inter-field relationship setting means, when the strength of the partial relation between fields calculated by the partial relation strength calculating means is equal to or greater than a predetermined threshold, the relation strength between the subsets as the relation between the fields. Is an evaluation point between fields, or determines whether or not there is a relationship between subsets. In the partial relationship extraction device between the attributes according to claim 1 or 2,
An apparatus for extracting a partial relationship between attributes, characterized by comprising designation means for designating a relationship strength calculation method and an employment condition adopted for each field for each of the plurality of fields. Extracting partial relations between attributes causing a computer to execute a method for extracting a partial relation between attributes in a model creation device that constitutes a predetermined model based on attribute values in a plurality of fields included in each of a plurality of data and the appearance frequency of the attribute values A program,
An overall relationship strength calculating step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution;
A partial relation strength calculating step for calculating the strength of a partial relation between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute;
A field relationship setting step for determining a relationship between the fields based on the strength of the partial relationship between fields calculated by the partial relationship strength calculating step ;
The partial relationship intensity calculating step, partial relationship extraction program between attributes strength of the calculated overall relationship by the overall relationship strength calculation step you calculate the strength of the parts involved in the case of less than a predetermined threshold value. In the partial relationship extraction program between the attributes according to claim 4 ,
In the inter-field relationship setting step, when the strength of the partial relationship between the fields calculated in the partial relationship strength calculating step is equal to or greater than a predetermined threshold, the relationship strength between the subsets is set as the relationship between the fields. Is an evaluation point between fields, or determines whether or not there is a relationship between subsets. In the partial relationship extraction program between the attributes according to claim 4 or 5,
A program for extracting a partial relationship between attributes, comprising a specifying step for specifying a relationship strength calculation method and a use condition adopted for each field for each of the plurality of fields. A method for extracting a partial relationship between attributes in a model creation device that constitutes a predetermined model based on an attribute value in a plurality of fields included in each of a plurality of data and an appearance frequency of the attribute value,
An overall relationship strength calculating step for calculating the strength of the overall relationship between fields based on all attribute values appearing in the field group for a given attribute and its frequency distribution;
A partial relation strength calculating step for calculating the strength of a partial relation between fields to which the attribute value belongs based on a subset of attribute values appearing in the field group for the predetermined attribute;
A field relationship setting step for determining a relationship between the fields based on the strength of the partial relationship between fields calculated by the partial relationship strength calculating step ;
The partial relationship intensity calculating step, portions relation extraction method between attributes strength of the calculated overall relationship by the overall relationship strength calculation step you calculate the strength of the parts involved in the case of less than a predetermined threshold value. In the partial relationship extraction method between the attributes according to claim 7,
A method for extracting a partial relationship between attributes, comprising: a designation step for designating a relationship strength calculation method and an employment condition adopted for each field for each of the plurality of fields.

Images