Detection device, detection method, and detection program

Abstract

A detection device according to the present disclosure is equipped with: a collection unit for collecting collection data pertaining to the number of event occurrences for each attribute which pertains to an event which occurred within a prescribed time period; a derivation unit which, by using the collection data, derives a first prediction accuracy which indicates the accuracy of the event occurrence likelihood upon considering the attributes, and a second prediction accuracy which indicates the accuracy of the event occurrence likelihood without considering the attributes; and a determination unit for determining whether a bias pertaining to the attributes is present in the collection data by using the degree of divergence between the first prediction accuracy and the second prediction accuracy.

Description

[Technical Field]

[0001] The disclosed technology relates to detection devices, detection methods, and detection programs. [Background technology]

[0002] For example, if there is a bias in the occurrence of an event in the training data used to generate a predictive model for predicting the occurrence of an event, using training data that takes this bias into account may improve the accuracy of the prediction by the predictive model. In other words, if there is a bias in the occurrence of an event, using uniform training data that does not take this bias into account may decrease the accuracy of the prediction by the predictive model.

[0003] Furthermore, for example, showing both the probability of an event occurring and the reason why that probability is high is due to that bias can improve the credibility (readability) of the probability of the event occurring.

[0004] Therefore, in recent years, when predicting the occurrence of an event, there are methods that use an index to determine the degree of bias in attributes such as the day of the week and date and time, and evaluate whether or not there is a bias in those attributes in the event in question. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Masayuki Morikawa, "Demand Fluctuations and Productivity in the Service Industry - Analysis Using Establishment Data," Research Institute of Economy, Trade and Industry, Discussion Paper 08-J-042. [Non-Patent Document 2] Yuki Kawano, Satoshi Yamada, Hironobu Abe, and Koichi Nakajima, "A Proposal for a Power Demand Forecasting Method for Consumers Using a Combination of Multiple Forecasting Methods," Transactions of the Information Processing Society of Japan, Consumer Devices & Systems, Vol. 3, No. 1, 53-63 (March 2013).

[0006] Non-patent document 1 discloses a method for deriving the degree of demand fluctuation by day of the week and by month using standard deviation, and for detecting the effects of these differences (whether or not there are differences in the events that occur).

[0007] Non-patent document 2 discloses a method for selecting data to be used for prediction based on the standard deviation. For example, if the standard deviation of the maximum temperature over the past three days is large, the latest data is used to predict the amount of electricity used without using historical data. If the standard deviation is low, the amount of electricity used is used to predict the amount of electricity used on similar days in historical data. [Overview of the project] [Problems that the invention aims to solve]

[0008] However, when detecting data bias using a method based on normalized standard deviation, there is a tendency to identify data with a small number of occurrences as biased. Conversely, when detecting data bias using a method based on standard deviation without normalization, there is a tendency to identify data with a large number of occurrences as biased.

[0009] In other words, when using standard deviation to detect data bias, it was not always possible to accurately detect data that was biased.

[0010] This disclosure is made in view of these circumstances and aims to propose a detection device, a detection method, and a detection program that can accurately detect biased data. [Means for solving the problem]

[0011] According to a first aspect of the present disclosure, there is provided a detection apparatus comprising: a collection unit that collects collected data relating to the number of occurrences of an event for each attribute related to an event that has occurred in a predetermined period; a derivation unit that uses the collected data to derive a first prediction accuracy indicating the accuracy of the occurrence probability of the event in consideration of the attribute, and a second prediction accuracy indicating the accuracy of the occurrence probability of the event without considering the attribute; and a determination unit that determines whether there is a bias related to the attribute in the collected data using a degree of deviation between the first prediction accuracy and the second prediction accuracy.

[0012] According to a second aspect of the present disclosure, there is provided a detection method in which a computer executes processing including: collecting collected data relating to the number of occurrences of an event for each attribute related to an event that has occurred in a predetermined period; using the collected data to derive a first prediction accuracy indicating the accuracy of the occurrence probability of the event in consideration of the attribute, and a second prediction accuracy indicating the accuracy of the occurrence probability of the event without considering the attribute; and determining whether there is a bias related to the attribute in the collected data using a degree of deviation between the first prediction accuracy and the second prediction accuracy.

[0013] According to a third aspect of the present disclosure, there is provided a detection program for causing a computer to function as the detection apparatus according to the first aspect.

Effect of the Invention

[0014] According to the technology of the present disclosure, data having a bias can be detected with high accuracy.

Brief Description of Drawings

[0015] [Figure 1] FIG. 1 is a block diagram showing a hardware configuration of the detection apparatus according to the present embodiment. [Figure 2] FIG. 2 is a block diagram showing an example of a functional configuration of the detection apparatus according to the present embodiment. [Figure 3] FIG. 3 is a data flow diagram showing an example of a data flow of detection processing according to the present embodiment. [Figure 4] FIG. 4 is a diagram showing an example of a distribution of occurrence presence / absence, which is provided for explanation of detection using occurrence presence / absence according to the present embodiment. [Figure 5] It is a diagram showing an example of the distribution of the number of occurrences provided for explaining detection using the number of occurrences according to the present embodiment. [Figure 6] It is a flowchart showing an example of detection processing according to the present embodiment. [Figure 7] It is a diagram showing an example of collected data provided for explaining the comparison of evaluation results obtained by each method according to the present comparative example. [Figure 8] It is a Venn diagram showing an example of inclusion relationships in items extracted by each method according to the present comparative example. MODE FOR CARRYING OUT THE INVENTION

[0016] Hereinafter, exemplary embodiments for carrying out the present disclosure will be described in detail with reference to the drawings.

[0017] First, the hardware configuration of the detection device 10 according to the present embodiment will be described with reference to FIG. 1. FIG. 1 is a block diagram showing the hardware configuration of the detection device 10 according to the present embodiment.

[0018] As shown in FIG. 1, the detection device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is communicably connected to each other via a bus 18. It should be noted that the configuration using the above-described CPU and memory is merely an example, and for example, the present invention may be implemented as a device dedicated to object detection equipped with a dedicated arithmetic circuit.

[0019] The CPU 11 is a central processing unit that executes various programs and controls each component. Specifically, the CPU 11 reads a program from the ROM 12 or storage 14 and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculations according to the program stored in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores a detection program for detecting data that shows bias for each attribute such as the day of the week from data collected about events that occurred over a predetermined period (hereinafter referred to as "collected data").

[0020] In this embodiment, the collected data is data collected on the presence or absence of events and the number of occurrences for each day of the week. However, it is not limited to this. The collected data may be collected by location and time of occurrence, or by daily, weekly, monthly, quarterly, seasonal, or yearly. Note that the day of the week is an example of an "attribute related to an event," and the location of occurrence is an example of a "condition." Furthermore, units related to time and period such as time of day, daily, weekly, monthly, quarterly, seasonal, and yearly are examples of "periodic units." In the following, we will describe collected data in which "day of the week" is applied as an attribute for detecting bias and "location of occurrence" is applied as a condition. Furthermore, the condition in this embodiment is a condition for sorting the collected data. For example, if the condition is "none," the presence or absence of bias in the entire acquired data is detected, and if the condition is "location of occurrence," the acquired data is sorted by location of occurrence to collect data, and the presence or absence of bias in the collected data is detected for each location of occurrence. Furthermore, the events in this embodiment are, for example, sales performance of products, traffic congestion occurrences, traffic accident occurrences, and reports to the police. For example, if the event is sales performance, the conditions would include the product, product category, and retail store. If the event is traffic congestion, traffic accidents, or police reports, the conditions could include the location, prefecture, and weather. For instance, detecting biases in each location would reveal which areas should have increased security on specific days of the week and at specific times. Similarly, detecting biases in each product would reveal which products should be promoted during specific seasons and which products should have increased stock on specific days of the week. Likewise, if it is found that accidents and reports are biased by weather, incorporating weather into security plans could help prevent accidents and incidents.

[0021] ROM12 stores various programs and data. RAM13 temporarily stores programs or data as a working area. Storage14 consists of a storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) and stores various programs, including the operating system, and various data.

[0022] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used for various types of input.

[0023] The display unit 16 is, for example, a liquid crystal display and displays various information. The display unit 16 may also function as an input unit 15 by employing a touch panel system.

[0024] The communication interface 17 is an interface for communicating with other devices such as display devices. For this communication, a wired communication standard such as Ethernet® or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi® may be used. The communication interface 17 acquires input data from an external device and transmits the detection results to the external device.

[0025] Next, the functional configuration of the detection device 10 will be described with reference to Figure 2. Figure 2 is a block diagram showing an example of the functional configuration of the detection device 10 according to this embodiment.

[0026] As shown in Figure 2, the detection device 10 has a functional configuration comprising an acquisition unit 21, a collection unit 22, an output unit 23, a determination unit 24, and a detection unit 25. The CPU 11 executes a detection program, causing the acquisition unit 21, collection unit 22, output unit 23, determination unit 24, and detection unit 25 to function.

[0027] As an example, as shown in Figure 3, the acquisition unit 21 acquires event data 31 related to events that occurred during a predetermined period. In this embodiment, the event data 31 will be described as data that includes the date, day of the week, and location where the event occurred.

[0028] The collection unit 22 uses the acquired event data 31 to collect data 32 for each day, day of the week, and location of occurrence. For example, the collection unit 22 uses the event data 31 to collect the data 32 shown in Figure 4 as an example. In this embodiment, the data 32 is described as being collected for the occurrence or non-occurrence of events over a four-year period. Furthermore, the breakdown of each day of the week is 209 days for Sunday, Monday, Tuesday, Wednesday, and Thursday, and 208 days for Friday and Saturday, for a total of 1461 days.

[0029] Here, Figure 4 shows the collected data 32, which is the result of collecting data on whether or not an event occurred on each day of the week and at each location. In Figure 4, location A shows that over the course of four years, the event occurred on Mondays, Tuesdays, and Fridays 0 days, on Wednesdays, Thursdays, and Sundays 1 day each, and on Saturdays 3 days. Location B shows that over the course of four years, the event occurred on Mondays 49 days, on Tuesdays 62 days, on Wednesdays 52 days, on Thursdays 51 days, on Fridays 67 days, on Saturdays 7 days, and on Sundays 1 day. Similarly, location C shows that over the course of four years, the event occurred on Mondays 0 days, on Tuesdays, Wednesdays, and Fridays 1 day each, on Thursdays 2 days, on Saturdays 7 days, and on Sundays 9 days.

[0030] The derivation unit 23 uses the collected data 32 for each location to derive a prediction accuracy 33 that takes the day of the week into consideration, and a prediction accuracy 34 that does not take the day of the week into consideration. Here, the prediction accuracy 33 that takes the day of the week into consideration is an example of the "first prediction accuracy," and the prediction accuracy 34 that does not take the day of the week into consideration is an example of the "second prediction accuracy."

[0031] Specifically, the derivation unit 23 uses the briar score (mean squared error) for each occurrence location to derive both a prediction accuracy 33 that takes the day of the week into account and a prediction accuracy 34 that does not take the day of the week into account. The briar score (mean squared error) is expressed by the following formula.

[0032]

number

[0033] Here, n is the number of days during a given period in which the event occurred, i is an identifier that identifies the day the event occurred, and p i y is the probability of an event occurring. i This value indicates whether or not an event has actually occurred, or the number of occurrences. Note that if the Briar score is calculated based only on the date the event occurred, y will always be used regardless of i. i = 1.

[0034] When the derivation unit 23 derives the prediction accuracy 33 that takes the day of the week into consideration, as shown in equation (1) above, it calculates the mean squared error for each day of the week and location where the event occurred, and divides the sum of the calculated mean squared errors by the number of days on which the event occurred.

[0035] For example, the derivation unit 23 calculates the probability of occurrence p for location A on Wednesday, Thursday, and Sunday. i Let each be 1 / 209, and the probability of occurrence on Saturday p i Set this to 3 / 208, and the value y indicates whether or not it occurs. i The mean squared error is calculated by setting it to 1. The derivation unit 23 divides the sum of the calculated mean squared errors by the number of days (6) on which the event occurred, and derives a prediction accuracy 33 of 0.98091 that takes into account the day of the week at location A.

[0036] Furthermore, the derivation unit 23 calculates the probability of occurrence p over the entire period. i Set this to 6 / 1461, and the value y indicates whether or not it occurs. i The mean squared error for each day in which the event occurred is calculated, with the mean squared error set to 1. The derivation unit 23 divides the sum of the mean squared errors for the six days by the number of days in which the event occurred (6 days) to derive a prediction accuracy 34 of 0.9918, which does not take into account the day of the week at location A.

[0037] It has been described above that the prediction accuracy (Brier score) according to the present embodiment is derived targeting only days on which an event occurs. However, the present invention is not limited thereto. The prediction accuracy (Brier score) may be derived targeting both days on which an event occurs and days on which no event occurs. In this case, n in the aforementioned formula (1) represents the number of days in the entire period instead of the number of days on which the event occurs, where y for a day on which the event occurs i =1, and y for a day on which no event occurs i =0. Further, according to the sparsity (degree of bias) in collected data 32, the prediction accuracy (Brier score) may be derived by multiplying the mean squared error for days on which the event occurs and the mean squared error for days on which no event occurs by respectively corresponding weight values. For example, when the weight value for days on which the event occurs is set to 1, the weight value for days on which no event occurs is set to 0, and these weight values are multiplied by the respectively corresponding mean squared errors, the result corresponds to the prediction accuracy 33 (Brier score) targeting only days on which the event occurs.

[0038] Further, it has been described above that the prediction accuracy (Brier score) according to the present embodiment is derived using occurrence / non-occurrence of an event. However, the present invention is not limited thereto. The prediction accuracy (Brier score) may be derived using the number of occurrences. Here, in the aforementioned formula (1), by applying the number of occurrences y i , the prediction accuracy (Brier score) for the number of occurrences is derived. It has been described in the above embodiment that the prediction accuracy is derived targeting only days on which an event occurs. However, the present invention is not limited thereto. The prediction accuracy is not limited to being derived in units of days, and may be derived in combination with other units, such as being derived by combining days and time slots.

[0039] As an example, as shown in Figure 5, the collection unit 22 uses the acquired event data 31 to collect the number of occurrences of events over a four-year period, broken down by day of the week and location of occurrence, as collected data 32. Figure 5 shows the collected data, which is the result of collecting the number of occurrences by day of the week and location of occurrence. Note that the collected data 32 shown in Figure 5 represents the number of occurrences in the first and second weeks of a given month, and it is explained that the number of occurrences in other weeks is 0.

[0040] Figure 5 shows that at location A, over a four-year period, there were days in the first week of a given month when there were 0 incidents on Monday, Tuesday, and Friday, 1 incident each on Wednesday, Thursday, and Sunday, and 2 incidents on Saturday. In the second week, there were 0 incidents from Monday through Friday and Sunday, and 1 incident on Saturday. At location B, over a four-year period, there were 2 incidents from Monday through Sunday in the first week of a given month, 1 incident each from Monday through Thursday and Sunday in the second week, and 0 incidents on Friday and Saturday.

[0041] The derivation unit 23 uses the above-described equation (1) to derive the prediction accuracy 33 considering the day of the week and the prediction accuracy 34 not considering the day of the week from the collected data 32 for each location. Here, y related to the above-described equation (1) i This value indicates the number of occurrences of the event that actually happened.

[0042] For example, the derivation unit 23 uses the location-specific data 32 shown in Figure 5 to determine the probability of occurrence p for Wednesday, Thursday, and Sunday of the first week. i Each of these is 1 / 209, and the probability of occurrence p related to the first Saturday is 1 / 209. i The probability of occurrence for the second Saturday of the 208th week is p i The value y represents the number of occurrences on Wednesday, Thursday, and Sunday of the first week, with the value set to 1 / 208. i 1 is the value representing the number of occurrences on the first Saturday of the week, and y is the value representing the number of occurrences on the first Saturday of the week. i 2. The value y represents the number of occurrences related to the second Saturday. iThe mean squared error is calculated by setting it to 1. The derivation unit 23 divides the sum of the calculated mean squared errors by the number of days (5) on which the event occurred, and derives a prediction accuracy 33 of 1.577048 that takes into account the day of the week at location A.

[0043] Furthermore, the derivation unit 23 calculates the probability of occurrence p over the entire period. i Set to 6 / 1461, y represents the number of occurrences on Wednesday, Thursday, and Sunday of the first week. i 1 is the value representing the number of occurrences on the first Saturday of the week, and y is the value representing the number of occurrences on the first Saturday of the week. i 2. The value y represents the number of occurrences related to the second Saturday. i The mean squared error for each day is calculated, with the mean squared error set to 1. The derivation unit 23 divides the sum of the calculated mean squared errors for the five days by the number of days (5) on which the event occurred, to derive a prediction accuracy 34 of 1.590161, which does not take into account the day of the week at location A.

[0044] As shown in Figure 3, the determination unit 24 uses the prediction accuracy 33 that takes the day of the week into account and the prediction accuracy 34 that does not take the day of the week into account, which are derived by the derivation unit 23, to determine the bias related to the day of the week. Specifically, the determination unit 24 uses the degree of deviation in prediction accuracy obtained by subtracting the prediction accuracy 34 that does not take the day of the week into account from the prediction accuracy 33 that takes the day of the week into account to determine the bias in the occurrence of events related to the day of the week.

[0045] For example, as shown in Figure 4, the determination unit 24 calculates a deviation of -0.0109 by subtracting the prediction accuracy 34 (0.9918) which does not consider the day of the week at location A from the prediction accuracy 33 (0.98091) which considers the day of the week at location A, and determines the bias in the occurrence of events related to the day of the week. The determination unit 24 determines that a bias related to the day of the week exists if the deviation is less than or equal to a predetermined threshold (for example, 0).

[0046] Here, the Briar score indicates that a smaller value indicates higher prediction accuracy. The degree of deviation is calculated by subtracting the prediction accuracy without considering the day of the week 34 from the prediction accuracy with the day of the week 33, thereby comparing the prediction accuracy with the day of the week 33 and the prediction accuracy without considering the day of the week 34. If the degree of deviation is 0 or less, it indicates that the prediction accuracy with the day of the week 33 is more accurate than the prediction accuracy without considering the day of the week 34, and that the collected data 32 has a bias related to the day of the week. In this embodiment, the degree of deviation is described as the difference between the prediction accuracy with the day of the week 33 and the prediction accuracy without considering the day of the week 34. However, it is not limited to this. The degree of deviation may also be calculated as the ratio of the prediction accuracy with the day of the week 33 and the prediction accuracy without considering the day of the week 34.

[0047] The detection unit 25 detects and outputs a detection result 35 according to the result determined by the determination unit 24. For example, if the detection unit 25 determines that there is a bias related to the day of the week, it outputs a detection result 35 indicating that there is a bias related to the day of the week in the acquired event data 31, and the degree of deviation. Also, if the detection unit 25 makes a determination using collected data 32 for each condition (e.g., location of occurrence, time of day, etc.), it outputs a detection result 35 for each condition indicating that there is a bias related to the day of the week, the degree of deviation, and the condition related to the collected data 32 (e.g., location of occurrence, etc.). For example, if there are multiple collected data 32 for each condition, the detection unit 25 detects the collected data in order of decreasing deviation and outputs each condition related to a predetermined number of collected data 32. The detection unit 25 may transmit the detection result 35 to an external device or display it on the display unit 16.

[0048] Next, the operation of the detection device 10 according to this embodiment will be described with reference to Figure 6. Figure 6 is a flowchart showing an example of the detection process according to this embodiment. The detection program shown in Figure 6 is executed when the CPU 11 reads the detection program from the ROM 12 or storage 14 and executes it. The detection program shown in Figure 6 is executed, for example, when event data 31 is input and an instruction to execute the detection process is input.

[0049] In step S101, the CPU 11 acquires event data 31.

[0050] In step S102, the CPU 11 uses the acquired event data 31 to collect collected data 32 by day, day of the week, and location of occurrence.

[0051] In step S103, the CPU 11 derives a prediction accuracy 33 (briar score) for each location, taking into account the day of the week.

[0052] In step S104, the CPU 11 uses the acquired event data 31 to collect data 32 for each day and location of occurrence.

[0053] In step S105, the CPU 11 derives a prediction accuracy 34 (briar score) for each location, without considering the day of the week.

[0054] In step S106, the CPU 11 calculates the degree of deviation in the collected data 32 by subtracting the prediction accuracy 34 (which does not consider the day of the week) from the prediction accuracy 33 (which considers the day of the week) for each location where the event occurred. Here, the calculated degree of deviation is associated with the location where the event occurred in the collected data 32 and stored in memory.

[0055] In step S107, the CPU 11 sorts the locations of the collected data 32 in descending order of the derived degree of deviation.

[0056] In step S108, the CPU 11 sequentially determines whether or not there is a bias related to the day of the week in the collected data 32. If there is a bias related to the day of the week in the collected data 32 (the degree of deviation is 0 or less) (step S108: YES), the CPU 11 proceeds to step S109. On the other hand, if there is no bias related to the day of the week in the collected data 32 (the degree of deviation is greater than 0) (step S108: NO), the CPU 11 terminates the detection process.

[0057] In step S109, the CPU 11 outputs a detection result 35 indicating that there is a bias in the collected data 32, the degree of deviation, and the location where the collected data 32 occurred.

[0058] In step S110, the CPU 11 determines whether it has output a predetermined number of locations for the collected data 32. If it has output a predetermined number of locations for the collected data 32 (step S110: YES), the CPU 11 terminates the detection process. On the other hand, if it has not output a predetermined number of locations for the collected data 32 (step S110: NO), the CPU 11 proceeds to step S108 to determine whether there is a bias related to the day of the week for the next collected data 32.

[0059] As described above, according to this embodiment, biased data sets can be detected with high accuracy.

[0060] In the above embodiment, the method of applying the day of the week as an attribute for collecting the collected data 32 was described. However, it is not limited to this. The attribute may be the date, time of day, location, weather, and a category related to the location, or it may be a combination of multiple items from the date, time of day, location, weather, and a category related to the location.

[0061] Furthermore, in the above embodiment, the method of applying the location of occurrence as a condition for detecting the collected data 32 was described. However, it is not limited to this. The conditions may be categories related to the date, time of day, day of the week, weather, and location of occurrence, or they may be a combination of multiple categories related to the date, time of day, day of the week, weather, and location of occurrence.

[0062] Furthermore, in the above embodiment, the detection result 35 was described in which the presence of a bias in the collected data 32, the degree of deviation, and the location of occurrence related to the collected data 32 were output. However, the embodiment is not limited to this. The collected data 32 may also be output as the detection result 35.

[0063] Furthermore, in the above embodiment, a method for evaluating prediction accuracy was described in which the mean squared error is applied. However, the method is not limited to this. As a method for evaluating prediction accuracy, the mean absolute error may be applied, the mean absolute error rate may be applied, the root squared error may be applied, or any method may be applied as long as the prediction accuracy can be evaluated.

[0064] (Comparative example) In the above embodiment, a method for evaluating the bias in occurrence by location and day of the week by calculating the degree of deviation using the briar score was described. Below, a comparative example of the method proposed in the above embodiment (hereinafter referred to as the "proposed method") and a method using standard deviation will be described. The method using standard deviation will be described using a normalized standard deviation (hereinafter referred to as the "normalized standard deviation") and a standard deviation that is not normalized (hereinafter referred to as the "unnormalized standard deviation"). Furthermore, in the method using normalized standard deviation, the total number of days on which the event occurred for each occurrence location (item) was normalized to the number of occurrence days based on 100 days to calculate the strength of the bias. Furthermore, in the method using normalized standard deviation, the total number of days on which the event occurred for each occurrence location (item) was normalized to the number of occurrence days based on 100 days to calculate the standard deviation.

[0065] Figure 7, shown as an example, is the collected data used to illustrate the comparison between the evaluation results for bias using the proposed method and the evaluation results for bias using the normalized standard deviation method. Figure 7 shows the collected data extracted from the total number of occurrences over four years at 400 locations, aggregated by day of the week. Specifically, it shows the data for four locations where no bias was observed by day of the week, and the data for four locations where bias was observed by day of the week.

[0066] Figure 7 shows the collected data, including the number of occurrences collected by location and day of the week, the total number of occurrences at each location, and whether or not there is a bias. It also shows the standard deviation and evaluation results for the method using normalized standard deviation, and the degree of deviation and evaluation results for the proposed method. The collected data in Figure 7 is the number of occurrences for each day of the week at eight locations, A through H. Locations A through D, where no bias was observed, are marked as "unbiased," while locations E through H, where a bias was observed, are marked as "biased." In this comparative example, the collected data shown in Figure 7 was evaluated for the presence or absence of bias using the method using normalized standard deviation and the proposed method. Here, in the method using normalized standard deviation, locations (items) with a standard deviation of 3.9 or higher, indicating the top 50 locations, were evaluated as "biased," and locations (items) with a standard deviation of less than 3.9 were evaluated as "unbiased." Furthermore, in the proposed method, locations (items) with a deviation of -0.02 or less, indicating that they are among the top 50 locations, were evaluated as "biased," while locations (items) with a deviation exceeding -0.02 were evaluated as "unbiased."

[0067] The evaluation results using the normalized standard deviation method showed "bias" for occurrence locations A and D through G, while "no bias" was observed for occurrence locations B through D and H. In other words, there was a discrepancy between the presence or absence of bias in the collected data and the evaluation results using the normalized standard deviation method.

[0068] On the other hand, the evaluation results using the proposed method showed "no bias" for locations A through D, and "bias present" for locations E through H. In other words, there was no difference between the presence or absence of bias in the collected data and the evaluation results using the proposed method.

[0069] Furthermore, as shown in Figure 8 as an example, using data collected over four years from 400 occurrence locations, we extracted locations (items) with strong bias using the proposed method, the method using normalized standard deviation, and the method using unnormalized standard deviation, and compared the extracted locations (items). In Figure 8, the strength of bias was calculated using each method, and the top 50 locations with the strongest bias were extracted from the 400 occurrence locations (items), and locations (items) with a total number of occurrences of 20 or more were extracted for each method.

[0070] As an example, as shown in Figure 8, the number of items that appeared only in the proposed method (hereinafter referred to as "number of items") was 2, and the average total number of days for the extracted items (hereinafter referred to as "average total number of days") was 36.00 days. The number of items that appeared only in the normalized standard deviation was 33, and the average total number of days was 17.57 days. The number of items that appeared only in the denormalized standard deviation was 15, and the average total number of days was 100.33 days. In addition, the number of items that overlapped between the proposed method and the method using normalized standard deviation was 13, and the average total number of days was 21.46 days. In addition, the number of items that overlapped between the proposed method and the method using denormalized standard deviation was 31, and the average total number of days was 118.03 days. In addition, the number of items that overlapped across the three methods (proposed method, method using normalized standard deviation, and method using denormalized standard deviation) was 4, and the average total number of days was 42.25 days. Furthermore, there were zero overlapping items between the method using normalized standard deviation and the method using unnormalized standard deviation.

[0071] As shown in Figure 8, by extracting the top 50 occurrences (items) using methods employing normalized standard deviation and methods employing unnormalized standard deviation, it is possible to include the top 50 occurrences (items) extracted by the proposed method. However, both methods employing normalized and unnormalized standard deviation are highly susceptible to the influence of the number of days and tend to extract items with little bias. Extracting items with little bias can lead to problems, for example, when performing machine learning using biased training data, where unbiased training data is selected, potentially causing the learning model to miss out on training data that contributes significantly to learning. Therefore, extracting items with little bias may increase the learning burden, as machine learning requires considering not only recall but also precision during training.

[0072] With the proposed method according to the above embodiment, the prediction accuracy considering the day of the week and the prediction accuracy not considering the day of the week are compared, and the degree of deviation of each prediction is evaluated, so that the degree of bias in the collected data is evaluated without being affected by the number of days.

[0073] In addition, the detection process that the CPU reads and executes in the above embodiment may be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits that are processors with circuit configurations specifically designed to execute specific processes, such as ASICs (Application Specific Integrated Circuits). Furthermore, the detection process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0074] Furthermore, although the above embodiments describe a configuration in which the detection processing program is pre-stored (installed) in ROM 12 or storage 14, the invention is not limited to this. The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), DVD-ROM (Digital Versatile Disk Read Only Memory), or USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that is downloaded from an external device via a network.

[0075] The following additional information is disclosed regarding the embodiments described above.

[0076] (Additional note 1) Memory and At least one processor connected to the memory, Includes, The aforementioned processor, For each attribute related to an event that occurred during a predetermined period, collect data regarding the number of occurrences of the event. Using the collected data, a first prediction accuracy indicating the accuracy of the probability of the event occurring while considering the attributes, and a second prediction accuracy indicating the accuracy of the probability of the event occurring while not considering the attributes are derived. The degree of deviation between the first prediction accuracy and the second prediction accuracy is used to determine whether or not there is a bias in the attribute in the collected data. A detection device configured in such a way.

[0077] (Additional note 2) For each attribute related to an event that occurred during a predetermined period, collect data regarding the number of occurrences of the event. Using the collected data, a first prediction accuracy indicating the accuracy of the probability of the event occurring while considering the attributes, and a second prediction accuracy indicating the accuracy of the probability of the event occurring while not considering the attributes are derived. The degree of deviation between the first prediction accuracy and the second prediction accuracy is used to determine whether or not there is a bias in the attribute in the collected data. A non-temporary storage medium that stores programs that can be executed by a computer. [Explanation of Symbols]

[0078] 10 Detection device 11 CPU 12 ROM 13 RAM 14 Storage 15 Input section 16 Display section 17 Communication Interface 18 bus 21 Acquisition Department 22 Collection Department 23 Derivation part 24 Judgment section 25 Output section 31 Event Data 32. Collected Data Prediction accuracy considering 33 days of the week Prediction accuracy without considering the 34 days of the week 35 Detection Results

Claims

1. A collection unit collects data on the number of occurrences of events for each attribute related to an event that occurred during a predetermined period, A derivation unit that uses the collected data to derive a first prediction accuracy that shows the accuracy of the probability of the event occurring while considering the attributes, and a second prediction accuracy that shows the accuracy of the probability of the event occurring while not considering the attributes. A determination unit that determines whether or not there is a bias in the attribute in the collected data using the degree of deviation between the first prediction accuracy and the second prediction accuracy, A detection device equipped with the following features.

2. The derivation unit derives the first prediction accuracy and the second prediction accuracy from the collected data, focusing on the periodic units in which the event occurred. The detection device according to claim 1.

3. The derivation unit derives the first prediction accuracy and the second prediction accuracy using the mean squared error. The detection device according to claim 1.

4. The derivation unit derives the first and second prediction accuracies by integrating the weight values ​​corresponding to the days on which the event occurred and the weight values ​​corresponding to the days on which the event did not occur, each corresponding to the sparsity rate of the attribute during the period, with the mean square error. The detection device according to claim 3.

5. The collected data further includes at least one of the conditions of the location where the event occurred and the time period in which the event occurred. The derivation unit derives the first prediction accuracy and the second prediction accuracy for each of the conditions. The detection device according to claim 1.

6. If it is determined that there is a bias in the attribute in the collected data, the system further includes a detection unit that detects a predetermined number of the conditions in descending order of the degree of deviation. The detection device according to claim 5.

7. For each attribute related to an event that occurred during a predetermined period, collect data regarding the number of occurrences of the event. Using the collected data, a first prediction accuracy indicating the accuracy of the probability of the event occurring while considering the attributes, and a second prediction accuracy indicating the accuracy of the probability of the event occurring while not considering the attributes are derived. The degree of deviation between the first prediction accuracy and the second prediction accuracy is used to determine whether or not there is a bias in the attribute in the collected data. A detection method for when a computer will perform a process.

8. A detection program for causing a computer to function as a detection device according to any one of claims 1 to 6.

Images