Behavioral analysis device, behavioral analysis program, and behavioral analysis method

Abstract

To make it possible to analyze an action while considering a time series of an action patten.SOLUTION: An action analysis device categorizes results obtained by combining events on the basis of history information containing actions composed of combinations of events performed in time series and results of the actions. The action analysis device creates a decision tree for categorizing results of actions by selecting, in the history information, one of events as a parent event and an event performed after the parent event in a first reference time frame as a selection event, and with the parent event as a start point and the selection event as a branch condition.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a behavior analysis device, a behavior analysis program, and a behavior analysis method. [Background technology]

[0002] Behavioral analysis is widely used to classify typical behavioral patterns based on the behavioral history of many users. A known type of behavioral analysis is decision tree analysis, which displays, in a flow chart, the types of behavior (explanatory variables) that affect a result (objective variable) based on a behavioral history that indicates the result (objective variable) produced by a combination of types of behavior (explanatory variables) (Non-Patent Document 1). By applying decision tree analysis, it becomes possible to analyze, for example, the relationship between a behavioral pattern performed while playing an electronic game and the result obtained by that behavioral pattern. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] https: / / surveroid.jp / mr-journal / terms / Atckv Summary of the Invention [Problem to be solved by the invention]

[0004] However, in conventional behavioral analysis that applies decision tree analysis, the time sequence of behaviors is not taken into account when constructing a decision tree from behavioral history. Therefore, even if a combination of multiple behaviors has a different order, it is analyzed as the same behavioral pattern, which reduces the interpretability of the behavioral analysis.

[0005] Furthermore, there is a problem in that the load of analytical processing becomes very large when a decision tree is constructed manually. [Means for solving the problem]

[0006] One aspect of the present invention is a behavioral analysis device that classifies the results obtained when events are combined based on historical information including behaviors consisting of a combination of events that occur in chronological order and the results of the behaviors, and is characterized by having a decision tree generation means that selects, as a selected event, one of the events in the historical information as a parent event, an event that occurred within a first reference time frame after the parent event, and generates a decision tree that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition.

[0007] Here, it is preferable that the decision tree generation means uses the events that occurred after the parent event in the history information within a second reference time frame different from the first reference time frame as events to be classified, and generates a decision tree that classifies the results of the behavior with the parent event as a starting point and the selected event as a branching condition.

[0008] Preferably, the second reference time frame is longer than the first reference time frame.

[0009] It is also preferable to perform natural language processing on the combination order of the branching events in the decision tree, thereby pruning the combination order of the branching events that are similar to each other from the decision tree.

[0010] It is also preferable that the decision tree generating means generates the decision tree using a Gini coefficient calculated from the probability of occurrence of the result with the parent event as a starting point and the selected event as a branching condition as a branching index.

[0011] It is also preferable that the decision tree generating means generates the decision tree using entropy calculated from the occurrence probability of the result with the parent event as a starting point and the selected event as a branching condition as a branching index.

[0012] Another aspect of the present invention is a behavioral analysis program that classifies the results obtained when events are combined based on historical information including behaviors consisting of a combination of events that occur in chronological order and the results of the behaviors, and is characterized by causing a computer to function as decision tree generation means that selects, as a selected event, one of the events in the historical information that is a parent event, an event that occurred within a first reference time frame after the parent event, and generates a decision tree that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition.

[0013] Another aspect of the present invention is a behavioral analysis method for classifying the results obtained when events are combined based on historical information including behaviors consisting of a combination of events that occur in chronological order and the results of the behaviors, characterized in that the computer is caused to process a decision tree generation step of selecting, as a selected event, one of the events in the historical information as a parent event, an event that occurred within a first reference time frame after the parent event, and generating a decision tree that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition. [Effects of the Invention]

[0014] According to the present invention, it is possible to provide a behavior analysis device, a behavior analysis program, and a behavior analysis method that enable behavior analysis that takes into account the time series of behavior patterns. Other objects of the embodiments of the present invention will become apparent by reading this specification as a whole. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a diagram illustrating a configuration of a behavior analysis apparatus according to an embodiment of the present invention. [Figure 2] FIG. 4 is a diagram showing an example of history information according to an embodiment of the present invention. [Figure 3] FIG. 1 is a diagram showing an example of a decision tree generated by behavioral analysis in an embodiment of the present invention. [Figure 4]FIG. 1 is a diagram showing an example of a decision tree generated by behavioral analysis in an embodiment of the present invention. [Figure 5] FIG. 10 is a diagram showing an example of a decision tree pruned by natural language processing in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] <Configuration of behavior analysis device> As shown in Fig. 1, behavior analysis apparatus 100 according to an embodiment of the present invention includes processing unit 10, storage unit 12, input unit 14, and output unit 16. In other words, behavior analysis apparatus 100 has the basic configuration of a computer and can be a general PC or the like. Furthermore, some of the functions of behavior analysis apparatus 100 may be realized by an external server connected via a communication means such as the Internet.

[0017] The processing unit 10 includes a means for performing arithmetic processing, such as a CPU. The processing unit 10 executes a behavior analysis program stored in the memory unit 12 to realize the behavior analysis processing of this embodiment. The memory unit 12 includes storage means, such as a semiconductor memory or a hard disk. The memory unit 12 is accessible and connected to the processing unit 10, and stores the behavior analysis program, data used in the behavior analysis processing, and the like. The input unit 14 includes a means for inputting information to the behavior analysis device 100. The input unit 14 includes, for example, a touch panel or a keyboard for receiving input from a user. The input unit 14 also includes a network interface or the like for receiving information from outside the behavior analysis device 100, and receives the behavior analysis program and data used in the behavior analysis processing. The output unit 16 includes a means for outputting information processed by the behavior analysis device 100. The output unit 16 includes, for example, a display. The output unit 16 also includes a network interface or the like for transmitting information outside the behavior analysis device 100, and transmits information to an external server or the like.

[0018] <Information used in behavioral analysis processing> The data input to the behavior analysis device 100 is history information that combines behavior patterns to be analyzed with the results (goal classes) obtained by the behavior patterns. When the analysis target is an electronic game, for example, as shown in Fig. 2, a history information database is used that contains history information that combines behavior patterns that combine events (play content) in the electronic game as branching conditions with the results of play (win or loss) obtained by the behavior patterns as goal classes. Each piece of history information stores actions performed in the electronic game as behavior patterns in chronological order.

[0019] <Behavioral analysis processing> Behavior analysis device 100 performs behavior analysis processing to classify results into objective classes for time-series history information such as that shown in FIG. 2. In this embodiment, behavior analysis is performed by automatically generating a decision tree in which events (behaviors) included in the history information are used as branching conditions. Branching criteria can be the same branching index as the criteria used when generating conventional decision trees. Branching criteria can be, for example, the Gini coefficient or entropy.

[0020] Specifically, a decision tree can be automatically generated by automatically calculating branching indices based on historical information and setting branching conditions based on the branching indices. That is, a decision tree is automatically generated by determining branching conditions until the impurity reaches a predetermined value so as to maximize information gain.

[0021] For example, when applying the Gini coefficient to generate a decision tree, the Gini coefficient Gini is calculated using formula (1). Here, c is the number of target classes (number of classes), and p(i|t) is the occurrence probability of target class i when the current branching condition (node) included in the history information is t. Here, when the number of event (action) samples is N and the number of data belonging to target class i is ni, p(i|t) = ni / N.

number

[0022] Using the Gini coefficient Gini, the information gain Δ can be calculated using equation (2). That is, the information gain Δ is defined as the difference between the Gini coefficients Gini of the parent node and the child node. Nj / N are the weights for the branching conditions after branching (the ratio of the amount of data before branching).

number

[0023] Behavior analysis apparatus 100 generates a decision tree that maximizes information gain Δ using each behavior included in the history information as a branching condition.

[0024] In this embodiment, a parent event (parent action) in the history information is set as the time point of interest, and actions that serve as candidate branching conditions for the decision tree are selected from events (actions) executed within a predetermined time frame (first reference time frame) from the parent event (parent action) based on the Gini coefficient and entropy. Actions included within the first reference time frame from the separation point of the parent node of the decision tree are extracted as candidates, and the action most suitable as a branching condition for the decision tree is selected from the candidate actions based on the Gini coefficient and entropy, to generate a decision tree with the parent event (parent action) as the starting point and the selected event (action) as the branching condition. After selecting one action as a branching condition in this way, a decision tree is automatically generated by sequentially repeating the process of searching for actions from the remaining actions based on the branching indicators.

[0025] Furthermore, in the process of selecting the most suitable action as a branching condition for the decision tree from the extracted events (actions), a branching index may be calculated taking into account events (actions) included in a second reference time frame that is different from the first reference time frame from the time of interest, and the selection may be made based on the branching index. The second reference time frame is preferably set to a time frame longer than the first reference time frame. For example, for each event (action) included in the first reference time frame from the time of interest, a Gini coefficient or entropy may be calculated taking into account events (actions) included in the second reference time frame from the time of interest, and a branching condition may be selected from the events (actions) included in the first reference time frame from the time of interest based on the Gini coefficient or entropy. This reduces the influence of slight differences in the order of events (actions) for each behavior pattern.

[0026] If the events (behaviors) included in the history information are actions that are executed in real time, the reference time frame is preferably a time frame of 1 minute, 5 minutes, 1 hour, etc. If the events (behaviors) included in the history information are actions that are executed step by step, the reference time frame may be a step frame consisting of 2 steps, 4 steps, etc.

[0027] Furthermore, in a decision tree, actions included in history information are branched, and it is preferable to generate branches for behavioral patterns that do not fit into any of the branches as "other." For example, branching conditions are selected such as a behavioral pattern in which behavior A was performed, a behavioral pattern in which behavior B was performed, and a behavioral pattern in which neither behavior A nor behavior B was performed. For example, if there are three or more branches for one branching condition, a search is made for an action that becomes a branching condition within a predetermined time frame from the branching condition one level above the remaining action, and the remaining action that does not fit into any of the branching conditions is branched as "other." In this case, the maximum number of branches (number of events) branching from one branching condition may be determined in advance. However, the number of branches (number of events) may be reduced depending on the number of events (actions) and the branching index.

[0028] FIG. 3 shows an example of a decision tree automatically generated for the history information in FIG. 2 when the first reference time frame is four steps. Action A in the first step 1 is set as the parent event, and branching conditions for the decision tree are selected based on the Gini coefficient and entropy from events (actions) included in steps 2 to 5 within four steps from the parent event. For example, action C is selected as the first branching condition in the decision tree from the actions included between steps 2 and 5, and other events (actions) are set as other branching conditions. In this case, when calculating the Gini coefficient and entropy, which serve as branching indicators, events (actions) included in a second reference time frame (e.g., six steps) different from the first reference time frame may be taken into consideration. Next, the newly selected action C is set as the new parent event, and branching conditions for the decision tree are selected based on the Gini coefficient and entropy from events (actions) within four steps from the parent event. In the behavior pattern of behavior history 1, the actions included in steps 3 to 6 within four steps from action C in step 2 are set as the branching conditions, and in the behavior pattern of behavior history 2, the actions included in steps 4 to 7 within four steps from action C in step 3 are set as the branching conditions. Then, based on the Gini coefficient and entropy, action G is selected as the branching condition for the next decision tree from the events (behaviors) set as the branching conditions, and other events (behaviors) are set as other branching conditions. By repeating this process, a decision tree is automatically generated for the behavior pattern containing the history information.

[0029] It should be noted that branches with a degree of achievement to the objective class lower than a reference value or branches with fewer applicable events (actions) than a reference value may be pruned from the decision tree.

[0030] As described above, according to the present embodiment, it is possible to provide a behavior analysis device, a behavior analysis program, and a behavior analysis method that enable behavior analysis that takes into account the time series of behavior patterns.

[0031] <Modification> In the behavior analysis process in the above embodiment, branches containing similar behavior patterns may be generated. For example, as shown in Fig. 4, a decision tree may be automatically generated that has a branch on the left end containing a behavior pattern of behavior C → behavior G and a branch on the right end containing a behavior pattern of behavior other than behavior C → behavior C → behavior G. Therefore, it is preferable to prune such similar branches.

[0032] Therefore, in this modification, natural language processing is applied to the behavioral patterns included in the decision tree to extract branches containing similar behavioral patterns, and the extracted branches are pruned. For example, Word2Vector or the like can be applied as natural language processing.

[0033] In the example of Figure 4, natural language processing is applied to the leftmost behavior pattern "Action A, Action C, Action G..." and the behavior pattern "Action A, Other than Action C, Action C, Action G..." to define expression vectors for these behavior patterns. The similarity between the defined expression vectors is calculated, and branches of behavior patterns whose similarity is equal to or greater than a predetermined threshold are extracted as branches with equivalent meanings. Then, as shown in Figure 5, a pruning process is applied to remove from the decision tree those branches with low goal achievement.

[0034] In this way, by performing natural language processing on the order of combinations of branching events (actions) in a decision tree, the order of combinations of branching events (actions) that are similar to each other can be pruned from the decision tree, thereby improving the interpretability of the decision tree. [Explanation of symbols]

[0035] 10 processing unit, 12 memory unit, 14 input unit, 16 output unit, 100 behavior analysis device.

Claims

1. A behavior analysis device that classifies results obtained when combining events based on history information including behaviors each consisting of a combination of events that occur in time series and the results of the behaviors, the device comprising: A behavior analysis device characterized by comprising a decision tree generation means for selecting, as a selected event, one of the events in the history information as a parent event, an event that occurred within a first reference time frame after the parent event, and generating a decision tree that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition.

2. The behavioral analysis device according to claim 1 , The behavior analysis device is characterized in that the decision tree generation means uses the events that occurred after the parent event in the history information within a second reference time frame different from the first reference time frame as events to be classified, and generates a decision tree that classifies the results of the behavior with the parent event as a starting point and the selected event as a branching condition.

3. The behavioral analysis device according to claim 2, The behavior analysis device is characterized in that the second reference time frame is longer than the first reference time frame.

4. The behavioral analysis device according to any one of claims 1 to 3, A behavior analysis device characterized by pruning mutually similar combination orders of branching events from the decision tree by natural language processing the combination order of branching events in the decision tree.

5. The behavioral analysis device according to any one of claims 1 to 4, The behavioral analysis device is characterized in that the decision tree generation means generates the decision tree using the Gini coefficient calculated from the occurrence probability of the result with the parent event as a starting point and the selected event as a branching condition as a branching index.

6. The behavioral analysis device according to any one of claims 1 to 4, The behavioral analysis device is characterized in that the decision tree generation means generates the decision tree using entropy calculated from the occurrence probability of the result with the parent event as the starting point and the selected event as the branching condition as a branching index.

7. A behavior analysis program that classifies results obtained when combining events based on history information including behaviors each consisting of a combination of events performed in chronological order and the results of the behavior, the program comprising: A behavioral analysis program that causes a computer to function as a decision tree generation means that selects, as a selected event, one of the events in the history information that occurs within a first reference time frame after the parent event as a parent event, and generates a decision tree that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition.

8. A behavior analysis method for classifying results obtained when combining events based on history information including behaviors each consisting of a combination of events performed in chronological order and the results of the behavior, the method comprising: A behavioral analysis method characterized by having a computer process a decision tree generation step in which one of the events in the history information is designated as a parent event, an event that occurred within a first reference time frame after the parent event is selected as a selected event, and a decision tree is generated that classifies the results of the behavior using the parent event as a starting point and the selected event as a branching condition.

Images