JP2016091343A - Information processing system, information processing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently execute an operation to an attribute in a data analysis using a statistical model.SOLUTION: A data analysis system 100 includes a statistic amount change calculation part 140 and an output part 180. The statistic amount change calculation part 140 calculates a change in a statistic amount related to a model generated by using a set of selection attributes obtained by a selection operation to each of a plurality of attributes from a statistic amount related to a model generated by using a set of present selection attributes among the plurality of attributes. The output part 180 outputs the calculated change in the static amount in association with each of the plurality of attributes.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing system, an information processing method, and a program, and more particularly, to an information processing system, an information processing apparatus, an information processing method, and a program that perform data analysis.

  In data analysis using a statistical model, it is necessary to select an attribute to be used for analysis from many attributes that represent data. Here, such attribute selection is referred to as attribute selection.

  FIG. 5 is a diagram illustrating an example of data used in data analysis. The data in FIG. 5 is an example of data related to rent for each region. In the example of FIG. 5, the data includes the region ID (Identifier), the number of people, the number of households, the average age, the male ratio, the marriage rate, the number of properties, the vacancy rate, the average annual income, and the average rent as attributes.

  When a regression analysis model is used as a statistical model for such data, for example, a linear function formula that predicts the average rent among the attributes is the other attributes (people, households, average age, male ratio, Marriage rate, number of properties, vacancy rate, average annual income) are expressed as variables.

  If the number of data (learning data) used to generate the statistical model is less than the number of attributes, the generated statistical model is not suitable for unknown data (test data). May happen. In order to avoid over-learning, various information criteria such as BIC (Bayesian Information Criterion) and algorithms for estimating an optimal solution of a set of attributes based on the information criterion have been proposed. . A statistical model without overlearning can be obtained by selecting some of the many attributes using such an algorithm.

  However, the attribute set obtained by the attribute selection by the algorithm as described above may not conform to the intention of the analyst. In other words, the algorithm selects attributes based on given learning data, but the analyst has general knowledge, domain knowledge, intuition, etc. as other knowledge. May be considered appropriate. For example, if an attribute is a characteristic related to a specific area and the analyst is familiar with the specific area, the analyst can obtain the algorithm based on the difficulty of acquiring the attribute value. Attribute may be deleted, or other attributes may be added.

  Generally, if the analyst's knowledge as described above can be expressed in advance for all attributes, they can be reflected in attribute selection by an algorithm. However, for example, if the analyst does not know how to express knowledge, the knowledge of the analyst cannot be reflected. In addition, when the number of attributes is large, it is not always possible to reflect knowledge about all attributes. Furthermore, even when new knowledge is born when the analyst sees the result of attribute selection by the algorithm, the knowledge of the analyst cannot be reflected in advance.

  Therefore, operations such as addition and deletion of attributes by an analyst are required for a set of attributes obtained by attribute selection by an algorithm.

  For example, Patent Literature 1 and Patent Literature 2 disclose techniques relating to attribute manipulation by an analyst in data analysis using a statistical model.

  The input variable selection support device described in Patent Literature 1 generates sensitivity for output variables of each input candidate variable, and selects and presents input variables based on the generated sensitivity. The data analysis method described in Patent Document 2 presents the relationship between the description attribute and the target attribute as the analysis result when the description attribute and the target attribute are selected in the regression analysis.

  As a related technique, Non-Patent Document 1 discloses a technique using a greedy method that is an approximation algorithm.

JP 2010-282547 A JP 2004-139199 A

Ji Liu et al. "Forward-Backward Greedy Algorithms for General Convex Smooth Functions over A Cardinality Constraint", [online], Journal of Machine Learning Research, [October 29, 2015 search], Internet <http://jmlr.org /proceedings/papers/v32/liub14.pdf>

  In the techniques of Patent Document 1 and Patent Document 2 described above, the sensitivity of the input variable to the output variable and the relationship between the explanation attribute and the target attribute are presented. However, in the techniques of Patent Literature 1 and Patent Literature 2, the analysis is performed when an operation such as addition or deletion of an input variable or explanation attribute is performed on the currently selected input variable or explanation attribute. It is not possible to know how the statistical model to be changed will change.

  An object of the present invention is to provide an information processing system, an information processing method, and a program capable of solving the above-described problems and efficiently executing operations on attributes in data analysis using a statistical model.

  The information processing system according to the present invention is obtained by a selection operation for each of the plurality of attributes from a predetermined statistic based on a predetermined model generated using a set of currently selected attributes among the plurality of attributes. A calculating means for calculating a change in the predetermined statistic relating to the predetermined model generated using a set of selection attributes; and the calculated predetermined statistic in association with each of the plurality of attributes. Output means for outputting a change in quantity.

  The information processing method of the present invention is obtained by a selection operation for each of the plurality of attributes from a predetermined statistic based on a predetermined model generated using a set of currently selected attributes among the plurality of attributes. A change in the predetermined statistic relating to the predetermined model generated using a set of selection attributes is calculated, and the calculated change in the predetermined statistic is associated with each of the plurality of attributes. Output.

  The program of the present invention is obtained by a selection operation for each of the plurality of attributes from a predetermined statistic associated with a predetermined model generated using a set of currently selected attributes among the plurality of attributes. A change in the predetermined statistic related to the predetermined model generated using a set of selected attributes, and the calculated change in the predetermined statistic in association with each of the plurality of attributes. Is output and the process is executed.

  An effect of the present invention is that operations on attributes can be efficiently executed in data analysis using a statistical model.

It is a block diagram which shows the characteristic structure of the 1st Embodiment of this invention. It is a figure which shows the structure of the data analysis system 100 in the 1st Embodiment of this invention. It is a block diagram which shows the structure of the data analysis system 100 implement | achieved by the computer in the 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the data analysis system 100 in the 1st Embodiment of this invention. It is a figure which shows the example of the data used in data analysis. It is a figure which shows the example of the calculation result of the value of an evaluation function in the 1st Embodiment of this invention. It is a figure which shows the example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the example of the transition of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the calculation result of the value of an evaluation function in the 1st Embodiment of this invention. It is a figure which shows the other example of the transition of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the calculation result of the value of an evaluation function in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the attribute display screen 200 in the 1st Embodiment of this invention. It is a figure which shows the other example of the transition of the attribute display screen 200 in the 1st Embodiment of this invention.

(First embodiment)
A first embodiment of the present invention will be described.

  First, the configuration of the first exemplary embodiment of the present invention will be described. FIG. 2 is a diagram showing a configuration of the data analysis system 100 in the first embodiment of the present invention. The data analysis system 100 is an embodiment of the information processing system of the present invention.

  Referring to FIG. 2, the data analysis system 100 includes a data storage unit 110, an initial attribute selection unit 120, a selection attribute storage unit 130, a statistic change calculation unit 140 (or calculation unit), a model generation unit 150, and an analysis unit 160. , An input unit 170 and an output unit 180.

  The data storage unit 110 stores data to be analyzed (learning data and test data). The data storage unit 110 stores, for example, data as illustrated in FIG. Note that the data may be expressed in CSV (Comma-Separated Values) format, for example. In this case, typically, an attribute name is given to the first line.

  The initial attribute selection unit 120 selects an attribute from among a plurality of attributes of the analysis target data using an algorithm. Here, in the algorithm, an attribute is selected based on a predetermined type of statistic representing “goodness” of a predetermined statistical model (hereinafter simply referred to as a model) generated using the attribute. As the type of the statistic, an information criterion relating to the model, likelihood, prediction error (hereinafter simply referred to as an error) by the model, and the like are used.

  As the model, for example, a regression model is used. In this case, the selected attribute is used as an explanatory variable. Note that the model may be another model such as an autoregressive model or a neural network model as long as it is a model that predicts the value of a certain attribute using a plurality of attribute values.

  The selection attribute storage unit 130 stores the currently selected attribute. Hereinafter, among the plurality of attributes of the data to be analyzed, a selected attribute is referred to as a selection attribute, and an unselected attribute (an attribute other than the selection attribute) is referred to as a non-selection attribute.

  The statistic change calculation unit 140 calculates, for each of a plurality of attributes of the analysis target data, a set of currently selected attributes of a model statistic generated using a set of selection attributes obtained by a selection operation on the attribute. To calculate the change from the model statistic generated. Here, the selection operation includes a change of the current selection attribute to the non-selection attribute (attribute deletion) and a change of the current non-selection attribute to the selection attribute (attribute addition).

  The magnitude of the change in statistic calculated for each attribute indicates the value of the selection operation for that attribute. In the first embodiment of the present invention, a function representing the magnitude of a change in statistics is referred to as an evaluation function for selection operation. The statistic change calculation unit 140 calculates a change in statistic using the evaluation function. As the evaluation function, for example, a statistic difference or a ratio is used.

  Note that the statistic type used by the initial attribute selection unit 120 for attribute selection by the algorithm and the statistic type used by the statistic change calculation unit 140 for calculating the evaluation function are the same or different. May be. Further, the statistic change calculation unit 140 may change the type of the statistic used for calculating the evaluation function in response to a request from an analyst or the like.

  The model generation unit 150 generates a model instructed by the statistic change calculation unit 140 using the learning data, and calculates the statistic of the generated model.

  The analysis unit 160 performs a predetermined analysis such as prediction for new analysis target data using a model.

  The input unit 170 receives an input of a selection operation from an analyst or the like.

  The output unit 180 outputs (displays) the value of the evaluation function calculated by the statistic change calculation unit 140 in association with each attribute.

  Each of the data analysis systems 100 may be a computer that includes a CPU (Central Processing Unit) and a storage medium that stores a program and that operates by control based on the program.

  FIG. 3 is a block diagram showing a configuration of a data analysis system 100 realized by a computer according to the first embodiment of the present invention.

  Referring to FIG. 3, a data analysis system 100 includes a CPU 101, a storage device (storage medium) 102 such as a hard disk and a memory, a communication device 103 that communicates with other devices, an input device 104 such as a mouse and a keyboard, An output device 105 such as a display is included.

  The CPU 101 executes a computer program for realizing the functions of the initial attribute selection unit 120, the statistic change calculation unit 140, the model generation unit 150, the analysis unit 160, the input unit 170, and the output unit 180. The storage device 102 stores data in the data storage unit 110 and the selection attribute storage unit 130. The input device 104 receives an input of a selection operation from an analyst or the like. The output device 105 outputs the value of the evaluation function to an analyst or the like. Instead of the input device 104, the communication device 103 may accept an input of a selection operation from another apparatus. Further, instead of the output device 105, the communication device 103 may output the value of the evaluation function to another apparatus.

  In addition, the data analysis system 100 may be configured such that each component illustrated in FIG. 2 is distributed in a plurality of physical devices connected by wire or wirelessly.

  Next, the operation of the first exemplary embodiment of the present invention will be described. FIG. 4 is a flowchart showing the operation of the data analysis system 100 according to the first embodiment of the present invention.

  Here, it is assumed that data shown in FIG. 5 is stored in the data storage unit 110 as data to be analyzed. Further, as a model, when a regression model such as Equation 1 is generated with the attribute “average rent” in FIG. 5 as an objective variable and an attribute selected from attributes other than the attributes “region ID” and “average rent” as explanatory variables, Assume.

  Here, y′i is the estimated value of the objective variable in the i th data, x ji is the value of the j th explanatory variable in the i th data, and Aj (j = 1,..., M) is the j th The coefficient, M, for the explanatory variable is the number of explanatory variables.

  First, the initial attribute selection unit 120 performs attribute selection by an algorithm from among a plurality of attributes included in the analysis target data stored in the data storage unit 110 (step S101).

  For example, the initial attribute selection unit 120 selects the attributes “population”, “male ratio”, and “vacancy ratio” by an algorithm based on the information amount criterion.

  The initial attribute selection unit 120 stores the selected attribute in the selection attribute storage unit 130 (step S102).

  The statistic change calculation unit 140 instructs the model generation unit 150 to generate a model using the current set of selected attributes and to calculate the statistic of the generated model (step S103).

  The model generation unit 150 generates a model using the current set of selection attributes based on the learning data stored in the data storage unit 110. And the model production | generation part 150 calculates the statistics of the produced | generated model using the data memorize | stored in the data storage part 110. FIG.

  For example, the model generation unit 150 generates a regression model using the current set of selected attributes {“population”, “male ratio”, “vacancy rate”}. Here, when using an error as a statistic, the model generation unit 150 generates, for example, the root mean square error (RMSE) of the regression model, which is expressed by Equation 2, and the error of the generated model. Calculate as

  Here, N is the number of data used for error calculation, and yi is the observed value of the objective variable in the i-th data. For example, learning data is used to calculate the error. In addition, test data or both learning data and test data may be used for calculating the error. In addition, the model generation unit 150 may calculate the error of the generated model by other methods, such as the maximum value of the absolute value of error (MAE: Max Absolute Error), without being limited to Equation 1.

  Next, the statistic change calculation unit 140 extracts one of the currently selected attributes (step S104).

  For example, the statistic change calculation unit 140 extracts the attribute “population” from the current selection attribute.

  The statistic change calculation unit 140 generates a model using the set of selection attributes obtained by deleting the extracted attributes from the current set of selection attributes, and the generated model Is instructed to calculate the statistic (step S105).

  For example, the model generation unit 150 generates a regression model by using a selection attribute set {“male ratio”, “vacancy rate”} obtained by deleting the attribute “population” from the current selection attribute set. The model generation unit 150 calculates an error of the generated regression model by using Formula 2.

  The statistic change calculation unit 140 calculates the value of the evaluation function related to the extracted attribute (step S106).

  Here, in the case of using a statistic (error) difference as the evaluation function, the statistic change calculation unit 140 subtracts the error calculated in step S103 from the error calculated in step S106, thereby obtaining the value of the evaluation function. Is calculated.

  The statistic change calculation unit 140 repeats the processing from step S104 for all the currently selected attributes (step S107).

  For example, the model generation unit 150 generates a regression model, calculates an error, and uses a set of selection attributes obtained by deleting each of “male ratio” and “vacancy rate” from the current selection attribute, and The value of the evaluation function is calculated.

  Next, the statistic change calculation unit 140 extracts one of the current non-selected attributes (step S108).

  For example, the statistic change calculation unit 140 extracts the attribute “number of households” from the current non-selected attribute.

  The statistic change calculation unit 140 generates a model using a set of selection attributes obtained by adding the extracted attribute to the current selection attribute to the model generation unit 150, and relates to the generated model. The calculation of statistics is instructed (step S109).

  For example, the model generation unit 150 uses a set of selection attributes {“population”, “male ratio”, “vacancy rate”, “number of households”} obtained by adding the attribute “number of households” to the current selection attribute, Generate a regression model. The model generation unit 150 calculates an error of the generated regression model by using Formula 2.

  The statistic change calculation unit 140 calculates the value of the evaluation function related to the extracted attribute (step S110).

  Here, the statistic change calculation unit 140 calculates the value of the evaluation function by subtracting the error calculated in step S103 from the error calculated in step S110.

  The statistic change calculation unit 140 repeats the processing from step S108 for all the current non-selected attributes (step S111).

  For example, the model generation unit 150 uses a set of selection attributes obtained by adding each of “average age”, “married ratio”, “number of properties”, and “average annual income” to the current selection attribute, and uses the regression model. Generation, error calculation, and evaluation function value calculation.

  FIG. 6 is a diagram illustrating an example of the calculation result of the value of the evaluation function in the first embodiment of the present invention. For example, as shown in FIG. 6, the value of the evaluation function is calculated for each of the currently selected attribute and the non-selected attribute.

  Next, the output unit 180 selects the attributes included in the analysis data from the attribute display screen 200 arranged in the order corresponding to the selected state (selected attribute or non-selected attribute) and the calculated evaluation function value. Is generated and output (displayed) to an analyst or the like (step S112).

  Here, each attribute is arranged on the attribute display screen 200 in the order according to the following first rule and second rule.

  The first rule is the order according to the selection state (selected attribute or non-selected attribute). That is, on the attribute display screen 200, the attributes are arranged in the order of selected attributes and non-selected attributes.

  The second rule is the order of the magnitudes of the evaluation function values. That is, on the attribute display screen 200, the selected attribute and the non-selected attribute are arranged in the order of descending order of the evaluation function values (descending order) or in ascending order (ascending order).

  Here, when a smaller model represents a better model, selection attributes are arranged in descending order of evaluation function values, and non-selection attributes are arranged in ascending order of evaluation function values. On the other hand, when a larger model represents a better model, selection attributes are arranged in ascending order of evaluation function values, and non-selection attributes are arranged in descending order of evaluation function values.

  As a result, the change in the statistic due to the selection operation from the statistic of the model generated using the current set of selected attributes at the boundary between the selected attribute and the non-selected attribute is minimized (indicating a peak). The selection attribute and the non-selection attribute are arranged. In this case, even if a selection operation is performed on an attribute close to the boundary between the selection attribute and the non-selection attribute, the goodness of the model generated using the set of selection attributes obtained by the selection operation is determined based on the current set of selection attributes. It does not drop much from the model generated using it. On the other hand, when a selection operation is performed on an attribute far from the boundary between the selected attribute and the non-selected attribute, the goodness of the model is greatly reduced.

  FIG. 7 is a diagram showing an example of the attribute display screen 200 in the first embodiment of the present invention.

  For example, the output unit 180 displays the attribute display screen 200 as shown in FIG. 7 based on the value of the evaluation function shown in FIG.

  In FIG. 7, rectangles indicating the attribute names of the respective attributes are arranged in order in the form of a grid, with the top of the attribute display screen 200 having the first priority and the left side having the second priority. The current selection attribute is distinguished by a bold line. And according to the above-mentioned 1st rule, an attribute is displayed in order of the present selection attribute and a non-selection attribute. Further, according to the second rule described above, the current selection attributes “population”, “male ratio”, and “vacancy rate” are arranged in descending order of evaluation function values (error difference). In addition, the non-selected attributes “number of households”, “average age”, “married rate”, “number of properties”, and “average annual income” are arranged in ascending order of evaluation function values (difference in error). .

  An analyst or the like refers to the attribute display screen 200 of FIG. At this time, the analyst or the like can grasp the degree of decrease in the goodness of the model when the selection operation is performed on each attribute according to the order in which the attributes are arranged. In other words, analysts and others can create a new set of selected attributes without significantly degrading the model from the model for the current selected attribute by deleting and adding attributes close to the boundary between the selected attribute and the non-selected attribute. Can be obtained.

  Next, the input unit 170 receives an input of a selection operation for the attribute on the attribute display screen 200 from an analyst or the like (step S113).

  Here, for example, when the selection attribute is clicked with the mouse, the input unit 170 determines that the selection attribute is changed to a non-selection attribute (attribute deletion). If the non-selected attribute is clicked with the mouse, it is determined that the non-selected attribute is changed to the selected attribute (attribute addition).

  In accordance with the selection operation, the input unit 170 updates the current selection attribute and stores it in the selection attribute storage unit 130 (step S114). Further, the output unit 180 reflects the result of the selection operation on the attribute display screen 200.

  Thereafter, by repeatedly executing the processing from step S103, the attributes on the attribute display screen 200 are rearranged according to the new selection state and the value of the evaluation function.

  8 and 10 are diagrams showing an example of transition of the attribute display screen 200 in the first exemplary embodiment of the present invention. 9 and 11 are diagrams showing another example of the calculation result of the evaluation function value in the first exemplary embodiment of the present invention.

  For example, when the selected attribute “population” is clicked (deleted) on the attribute display screen 200 in the state (a) in FIG. 8, the output unit 180 displays the attribute display screen 200 in the state (b) in FIG. 8. To do.

  The model generation unit 150 generates a regression model using a set of new selection attributes {“male ratio”, “vacancy rate”}, and calculates an error. The model generation unit 150 generates a regression model, calculates an error, and evaluates using a set of selection attributes obtained by deleting each of the new selection attributes “male ratio” and “vacancy rate”. Calculate the value of the function. Further, the model generation unit 150 adds the non-selection attributes “population”, “number of households”, “average age”, “marital ratio”, “number of properties”, and “average annual income” to the new selection attributes. A set of selection attributes obtained is used to generate a regression model, calculate an error, and calculate an evaluation function value. As shown in FIG. 9, the statistic change calculation unit 140 calculates the value of the evaluation function for each of the new selection attribute and the non-selection attribute. Then, the output unit 180 rearranges the attributes on the attribute display screen 200 based on the value of the evaluation function in FIG. 9 as in the state (c) in FIG.

  Similarly, for example, when the non-selected attribute “average age” is clicked (added) on the attribute display screen 200 in the state (a) in FIG. 10, the output unit 180 displays the attribute in the state (b) in FIG. 10. A display screen 200 is displayed.

  The model generation unit 150 generates a regression model using a new set of selection attributes {“population”, “male ratio”, “vacancy rate”, “average age”}, and calculates an error. In addition, the model generation unit 150 uses the set of selection attributes obtained by deleting each of the new selection attributes “population”, “male ratio”, “vacancy rate”, and “average age”, and Generation, error calculation, and evaluation function value calculation are performed. Further, the model generation unit 150 uses a set of selection attributes obtained by adding each of the non-selection attributes “number of households”, “married ratio”, “number of properties”, and “average annual income” to the new selection attributes. The regression model is generated, the error is calculated, and the evaluation function value is calculated. The statistic change calculating unit 140 calculates the value of the evaluation function for each of the new selection attribute and the non-selection attribute, as shown in FIG. Then, the output unit 180 rearranges the attributes on the attribute display screen 200 based on the value of the evaluation function in FIG. 11, as in the state (c) in FIG.

  Thus, the operation of the first exemplary embodiment of the present invention is completed.

  In the above description, the attributes are arranged in the order according to the value of the evaluation function on the attribute display screen 200. However, the present invention is not limited to this, and the attributes may be arranged so that the value of the evaluation function related to each attribute can be identified.

  12, 13, 14, 15, 16, and 17 are diagrams showing another example of the attribute display screen 200 in the first exemplary embodiment of the present invention.

  For example, in the attribute display screen 200, the output unit 180 may arrange the attributes at intervals according to the difference in the value of the evaluation function between the attributes as shown in FIG.

  Further, the output unit 180 may arrange each attribute at a position corresponding to the value of the evaluation function on the one-dimensional scale on the attribute display screen 200 as shown in FIG. In this case, the one-dimensional scale does not need to be linear, and may be, for example, a spiral.

  Further, the output unit 180 may arrange each attribute on the attribute display screen 200 such that the distance from the center on the two-dimensional plane corresponds to the absolute value of the evaluation function as shown in FIG. In this case, the output unit 180 may adjust the position of each attribute using a spring model.

  Moreover, the output part 180 may show the statistic used for calculation of the value of the said evaluation function with the value of the evaluation function of each attribute on the attribute display screen 200 as shown in FIG. Here, for example, “RMSE: 0.7-> 1.1 (+0.4)” is “0.7” for the model generated using the current set of selection attributes, and the set of selection attributes obtained by the selection operation is used. It indicates that the RMSE of the generated model is “1.1” and the difference between the RMSE is “+0.4”. In this case, as shown in FIG. 16, when the mouse 201 is overlapped with an attribute, the statistics related to the attribute may be displayed in the pop-up area 202. In addition, as shown in FIG. 17, when the mouse 201 is overlapped with an attribute, the statistics related to the attribute may be displayed in the common display area 203.

  Each attribute may be displayed in a color corresponding to the value of the evaluation function (hue, brightness, saturation, etc. depending on the value of the evaluation function) so that the value of the evaluation function related to the attribute can be identified. Good. Each attribute may be displayed in a shape, size, or the like according to the value of the evaluation function.

  Further, the output unit 180 may output a predicted value based on the model instead of indicating the value of the evaluation function related to each attribute or the statistic used to calculate the value of the evaluation function.

  FIG. 18 is a diagram showing still another example of the attribute display screen 200 in the first exemplary embodiment of the present invention.

  In the example of FIG. 18, a two-dimensional graph 204 is arranged on the attribute display screen 200, where the horizontal axis is the data identifier (“region ID”) and the vertical axis is the objective variable “average rent”. In the graph 204, the observed value of the objective variable “average rent” included in the data and the predicted value based on the regression model are shown. In this case, for example, when the mouse 201 is not overlapped with any attribute as in the state (a) of FIG. 18, the objective variable “average” by the regression model generated using the current set of selected attributes is used. The predicted value of “rent” is displayed. When the mouse 201 is overlapped with an attribute as in the state (b) of FIG. 18, the objective variable “average” by the regression model generated using the set of selection attributes obtained by the selection operation for the attribute is used. The predicted value of “rent” is displayed. Thereby, an analyst etc. can grasp visually change of a predicted value by a model accompanying selection operation to an attribute.

  In the above description, the attributes are arranged on the attribute display screen 200 in the order according to the first rule and the second rule. However, the output unit 180 is not limited to this, and instead of applying the first rule, the output unit 180 displays the selection attribute and the non-selection attribute, for example, the attribute display screen 200 like the left half and the right half of the attribute display screen 200. You may arrange | position in the upper different area | region.

  In the above description, attribute selection by an analyst or the like is performed after attribute selection by an algorithm. However, not limited to this, attribute selection by the algorithm after the attribute selection by the analyst, etc., attribute selection by the analyst etc. during attribute selection by the algorithm, etc. The attribute selection may be performed in another order.

  In the above description, as the selection operation, the current selection attribute is changed to a non-selection attribute (attribute deletion), and the current non-selection attribute is changed to a selection attribute (attribute addition). However, the present invention is not limited to this, and as a selection operation, a change (attribute replacement) of the current non-selection attribute to the selection attribute may be performed simultaneously with the change of the current selection attribute to the non-selection attribute.

  In this case, for example, the input unit 170 distinguishes attribute replacement from attribute deletion or attribute addition by switching modes (using toggle buttons or the like) or keyboard combinations (clicking the mouse while pressing a shift key or the like).

  When attribute replacement is designated, the input unit 170 changes the clicked selection attribute (replacement source attribute) to a non-selection attribute, and changes the clicked non-selection attribute (replacement destination attribute) to a selection attribute.

  When attribute replacement is designated, the arrangement of attributes on the attribute display screen 200 may be updated by the same processing as steps S103 to S112 described above each time a selected attribute or non-selected attribute is clicked. Further, instead of updating the arrangement of attributes every time the attribute is clicked, the arrangement of attributes on the attribute display screen 200 may be updated after the combination of the replacement source attribute and the replacement destination attribute is determined.

  FIG. 19 is a diagram showing another example of the transition of the attribute display screen 200 in the first exemplary embodiment of the present invention.

  For example, when the attribute replacement is designated on the attribute display screen 200 in the state (a) in FIG. 19 and the selection attribute “population” and the non-selection attribute “average age is” are clicked, the output unit 180 displays The attribute display screen 200 in state 19 (b) is displayed. The statistic change calculation unit 140 includes new selection attributes “male ratio”, “vacancy rate”, “average age”, and non-selection attributes “population”, “number of households”, “average age”, “married rate” ”,“ Number of properties ”, and“ average annual income ”, the value of the evaluation function is calculated. Then, the output unit 180 rearranges the attributes on the attribute display screen 200 based on the calculated value of the evaluation function, for example, as in the state (c) of FIG.

  Next, a characteristic configuration of the first exemplary embodiment of the present invention will be described. FIG. 1 is a block diagram showing a characteristic configuration of the first embodiment of the present invention.

  Referring to FIG. 1, the data analysis system 100 (information processing system) includes a statistic change calculation unit 140 (calculation unit) and an output unit 180. The statistic change calculation unit 140 calculates a set of selection attributes obtained by a selection operation for each of the plurality of attributes from a statistic related to a model generated using the current set of selection attributes of the plurality of attributes. The change of the statistic related to the model generated by the calculation is calculated. The output unit 180 outputs a change in the calculated statistic in association with each of the plurality of attributes.

  According to the first embodiment of the present invention, it is possible to efficiently perform operations on attributes in data analysis using a statistical model. The reason is that the statistic change calculation unit 140 calculates the change in the statistic of the model generated for the selection attribute after the selection operation from the statistic related to the model generated for the current selection attribute. This is because the output unit 180 outputs a change in statistics for each of the plurality of attributes. This provides the analyst etc. with information for supporting the attribute selection operation (changes in model statistics after the selection operation for each attribute). Can be operated easily.

(Second Embodiment)
Next, a second embodiment of the present invention will be described.

  In the first embodiment of the present invention, when an attribute selection operation is performed by an analyst or the like, the value of the evaluation function is recalculated for a new selection attribute set or a selection attribute set obtained by the selection operation. The attribute display screen 200 has been updated. However, in the recalculation of the evaluation function, it is necessary to repeatedly generate a model and calculate a statistic for the set of selected attributes. For this reason, depending on the number of attributes and the number of data, the calculation time becomes enormous, resulting in a burden on the analyst.

  In the second embodiment of the present invention, in the calculation of the statistic in the above-described steps S103, S105, and S109, the statistic change calculation unit 140 uses the set of the selected attribute set and the calculated statistic as a statistic. Cache in a quantity storage (not shown). Then, the statistic change calculation unit 140 uses the cache in recalculation of the evaluation function after the selection operation by the analyst or the like. That is, the statistic change calculation unit 140 searches the statistic cached in the statistic storage unit by using a set of selection attributes for which the statistic is to be calculated as a key in the recalculation of the evaluation function (steps S103 to S111). . When the statistic exists, the statistic change calculation unit 140 recalculates the value of the evaluation function using the statistic.

  According to the second embodiment of the present invention, it is possible to reduce the time for recalculating the evaluation function after the selection operation by an analyst or the like. The reason is that the statistic change calculation unit 140 caches a set of the attribute set used to generate the model and the calculated statistic, and recalculates the evaluation function after the selection operation by an analyst or the like. This is because the cached statistics are used. Thereby, the burden accompanying recalculation of the evaluation function after the selection operation by the analyst or the like is reduced.

(Third embodiment)
Next, a third embodiment of the present invention will be described.

  In the third embodiment of the present invention, the output unit 180 presents the recalculation process to the analyst as another method for reducing the burden associated with the recalculation of the analyst.

  When the recalculation of the evaluation function value is performed in steps S103 to S111 described above, the output unit 180 displays that fact for the attribute for which the calculation of the evaluation function value has been completed.

  For example, for the attribute for which the calculation of the value of the evaluation function has been completed, the output unit 180 displays the color of the rectangle indicating the attribute name and the thickness of the frame in a color different from the attribute that has not been completed and the thickness of the frame. . Further, the output unit 180 may add an icon such as a flag mark to the attribute for which the calculation of the value of the evaluation function has been completed.

  The output unit 180 completes the rearrangement (rearrangement) in accordance with the order according to the first rule and the second rule from the attribute for which the calculation of the value of the evaluation function is completed. The attribute that has not been performed may be rearranged at a position different from the completed attribute. In this case, the attribute rearrangement occurs every time the calculation of the value of the attribute evaluation function is completed. Further, the movement of the attribute position on the attribute display screen 200 may be displayed as an animation so that the change in the order of the attributes due to the rearrangement can be visually grasped.

  Further, the input unit 170 may accept a selection operation from the completed attribute even if there is an attribute for which the calculation of the value of the evaluation function has not been completed. When the attributes are arranged according to the first rule and the second rule described above, as described above, the attribute is close to the boundary between the selected attribute and the non-selected attribute (position where the change in statistics is minimized). On the other hand, a selection operation is often performed. Therefore, it is desirable that the statistic change calculation unit 140 predicts an attribute that minimizes the change in the statistic, and calculates the value of the evaluation function in order from the attribute.

  Here, depending on the statistic, it is known that the statistic can be calculated by an approximate calculation with a shorter calculation time. For example, Non-Patent Document 1 discloses that when the likelihood of a model is used as a statistic, the likelihood can be approximated using a vector gradient. The statistic change calculation unit 140 calculates the value of the evaluation function using such approximate calculation. Then, the statistic change calculation unit 140 estimates the attribute close to the boundary between the selected attribute and the non-selected attribute from the value of the evaluation function calculated by the approximate calculation, and calculates the above-described attribute from the attribute estimated to be close to the boundary. The evaluation function is calculated in steps S103 to S111.

  According to the third embodiment of the present invention, it is possible to reduce the burden on the analyst and the like due to recalculation of the evaluation function after the selection operation by the analyst and the like. The reason is that when the evaluation function value is being recalculated, the output unit 180 presents the attribute for which the calculation of the evaluation function value has been completed, or the attribute from which the calculation of the evaluation function value has been completed. This is for rearrangement.

  While the present invention has been described with reference to the embodiments, the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

100 data analysis system 101 CPU
102 storage device 103 communication device 104 input device 105 output device 110 data storage unit 120 initial attribute selection unit 130 selection attribute storage unit 140 statistic change calculation unit 150 model generation unit 160 analysis unit 170 input unit 180 output unit 200 attribute display screen 201 Mouse 202 Pop-up area 203 Display area 204 Graph

Claims (13)

Generated using a set of selection attributes obtained by a selection operation for each of the plurality of attributes from a predetermined statistic based on a predetermined model generated using a set of currently selected attributes among the plurality of attributes Calculating means for calculating a change in the predetermined statistic relating to the predetermined model to be performed;
Output means for outputting the calculated change in the predetermined statistic in association with each of the plurality of attributes;
Information processing system with The output means outputs, on the output screen, the change in the predetermined statistic using at least one of a position, a size, a color, and a shape according to the amount of the change;
The information processing system according to claim 1. The selection operation for each of the plurality of attributes includes at least one of a change of the current selection attribute to each non-selection attribute and a change of the current non-selection attribute to each selection attribute.
The information processing system according to claim 1 or 2. The selection operation for each of the plurality of attributes further includes changing the current non-selected attribute to each selected attribute when any of the current selected attributes is changed to a non-selected attribute, and the current Including at least one of changes to each non-selected attribute of the current selected attribute when any non-selected attribute is changed to a selected attribute;
The information processing system according to claim 3. The output means changes the predetermined statistic with respect to each of the currently selected attributes, and changes the predetermined statistic with respect to each of the current non-selected attributes in ascending or descending order of the amount of the change. Are arranged on the output screen in the opposite order of the current selection attribute of the amount of the change,
The information processing system according to claim 3 or 4. The output means is generated using a predicted value based on the predetermined model generated using the current set of selected attributes and a set of selected attributes obtained by a selection operation for each of the plurality of attributes. Outputting a predicted value according to the predetermined model;
The information processing system according to claim 1. Furthermore, an input unit that receives a selection operation for each of the plurality of attributes and updates the current set of selection attributes with a set of selection attributes obtained by the selection operation;
The information processing system according to claim 1. The current selection attribute is selected in advance by an algorithm based on the predetermined statistic.
The information processing system according to claim 1. The predetermined statistic is one of an information criterion value related to the predetermined model, a likelihood related to the predetermined model, and a prediction error due to the predetermined model.
The information processing system according to claim 1. Furthermore, the storage unit stores the predetermined statistic related to the predetermined model calculated by the calculation unit in association with a set of selection attributes used for the calculation of the statistic.
The calculation means uses the predetermined statistic corresponding to a set of selection attributes obtained by a selection operation for each of the plurality of attributes, or the set of current selection attributes, stored in the storage means. Calculating a change in the predetermined statistic,
The information processing system according to claim 1. The output means displays, on the output screen, whether or not the calculation of the predetermined statistic change by the calculation means has been completed for each of the plurality of attributes.
The information processing system according to claim 1. Generated using a set of selection attributes obtained by a selection operation for each of the plurality of attributes from a predetermined statistic based on a predetermined model generated using a set of currently selected attributes among the plurality of attributes Calculating a change in the predetermined statistic relating to the predetermined model to be performed;
Outputting the calculated change in the predetermined statistic in association with each of the plurality of attributes;
Information processing method. On the computer,
Generated using a set of selection attributes obtained by a selection operation for each of the plurality of attributes from a predetermined statistic based on a predetermined model generated using a set of currently selected attributes among the plurality of attributes Calculating a change in the predetermined statistic relating to the predetermined model to be performed;
Outputting the calculated change in the predetermined statistic in association with each of the plurality of attributes;
A program that executes processing.

Images