JP6315698B2 - Index value calculation device, index value calculation method, and program - Google Patents

Description

  The present invention relates to an index value calculation device, an index value calculation method, and a program.

  In recent years, accumulation of purchase logs represented by POS (Point Of Sales system) data has made it possible to observe not only sales of the entire store but also purchase logs for each purchaser. In the marketing science field, trend analysis and forecasts using purchase logs are conducted to develop effective sales and advertising strategies. In existing technology, the observation of sales volume transition by visualization and customer segmentation by clustering are performed, but attention is paid to simple purchase frequency, and what considers how each purchase was made Few.

  For example, many of the conventional visualization methods focus on only the transition of the number of logs (see Non-Patent Document 2) and the periodicity of logs (see Non-Patent Document 6). There is not much consideration of habits that do not have periodicity, such as the degree of fit to habits. Also, in clustering performed in the marketing field, the number of purchases is often used (see Non-Patent Document 7), and consideration of purchase attitudes such as habits is not realized.

G.E.Box, G.M.Jenkins, and G.C.Reinsel, Time series analysis: forecasting and control-3rd edition, Prentice Hall, 1994 R. Chang, M. Ghoniem, R. Kosara, W. Ribarsky, J. Yang, E. Suma, C. Ziemkiewicz, D. Kern, and A. Sudjianto, Wirevis: Visualization of categorical, time-varying data from financial transactions , In Visual Analytics Science and Technology, 2007, IEEE Symposium on VAST 2007, pages 155-162, IEEE, 2007 A. Hayashi, T. Matsubayashi, and H. Sawada, Regular behavior measure for location based services, In Proceedings of the 2014 ACM conference on Web science, pages 299-300, ACM, 2014 C.D.Manning, P.Raghavan, and H.Schuetze, Introduction to information retrieval, volume1, Cambridge university press Cambridge, 2008 A. Shashua and T. Hazan, Non-negative tensor factorization with applications to statistics and computer vision, In Proceedings of the 22nd international conference on Machine learning, pages 792-799, ACM, 2005 M.Weber, M.Alexa, and W.Mueller, Visualizing time-series on spirals, In IEEE Symposium on Information Visualization, pages 7-13, IEEE Computer Society, 2001 Matsubayashi, Yukishima, Hayashi, Sawada, Brand selection analysis in purchasing behavior using non-negative tensor factorization, Network-developed Intelligence Study Group, 2014, JWEIN 2014, 2014 Tanaka Yo, Introduction to Consumer Behavior (6) 5, Attitude and Persuasion 1, Management Shibayashi Vol. 43 No. 3, 2006

  According to Non-Patent Document 8, “attitude” is regarded as important as one of the basic central concepts in consumer behavior theory. Attitude is a concept necessary for predicting and explaining purchasing and usage phenomena, such as why the product or brand is purchased or how the brand is purchased. In the case of purchasing, it is considered that the attitude differs depending on the combination of a plurality of purchasing conditions such as buyer, product, location, and time. In order to promote active consumption behavior, especially the initial purchase of unpurchased products and subsequent continuous purchases, an optimal sales and advertising strategy is required for each purchase condition. In addition, it is important to know the changes in attitude before and after the measures. In addition to internal factors such as interests and habits, purchasing attitudes are thought to be formed by multiple external factors such as trends, requests from others, the impact of measures such as discounts, and media influences.

  However, in particular, external factors are difficult to observe from the purchase log alone, and it is difficult to distinguish external factors from internal factors. Conventionally, attitudes were measured by questionnaire evaluation, but it was difficult to obtain subjective evaluations for all purchases.

  The present invention has been made in view of the above points, and an object thereof is to make it possible to calculate an index value indicating a tendency of an action from history data regarding the action.

Therefore, in order to solve the above problem, the index value calculation device is recorded before the first history data and the first history data of the set of history data recorded for each predetermined action by the individual, In addition, the first action is based on the history data recorded in the cycle time to which the recording time of the first history data belongs, among a plurality of cycle times that visit each other at the same cycle and at a certain interval. Counts the number of times, or is recorded before the first history data and the first history data, and is recorded in a unit period to which the recording time of the first history data belongs among a plurality of unit periods. The second action count is counted based on the history data, or recorded on the individual based on the first history data and the history data recorded before the first history data. Most A totaling unit that counts the third number of actions in the entire period from the recording time of the history data to the recording time of the first history data, and based on the first number of actions, the first history data The first action probability in the cycle period to which the recording time belongs is calculated, or the second action probability in the unit period to which the recording time of the first history data belongs is calculated based on the second action count. Or calculating a third action probability in the entire period based on the third action number, the first action probability, the second action probability, or the third action probability; 1st calculation which calculates the product with the value based on 1 action frequency or the said 2nd action frequency or the said 3rd action frequency as an index value which shows the degree to which the said predetermined | prescribed action is applied to the habit and parts, the possess The totaling unit is based on history data recorded at the cycle time out of the first history data and history data recorded before the first history data for each of the plurality of cycle times. Then, the fourth action count in the period is counted, or, for each of the plurality of unit periods, the first history data and the history data recorded before the first history data Based on the history data recorded in the unit period, the fifth action count in the unit period is totaled, and the history data recorded before the first history data and the first history data Based on the above, the sixth action count in the entire period from the recording time of the first history data recorded for the individual to the recording time of the first history data is tabulated, Using the mutual information amount based on the action probability for the whole period based on the number of actions and the fourth action number for each period or the fifth action number for each unit period, the period time Or it has the 2nd calculation part which calculates the index value which shows the degree of influence with respect to the action of the unit period .

  It is possible to calculate an index value related to a behavior tendency from the history data.

It is a figure which shows the hardware structural example of the index value calculation apparatus in embodiment of this invention. It is a figure which shows the function structural example of the index value calculation apparatus in embodiment of this invention. It is a flowchart for demonstrating an example of the process sequence which an index value calculation apparatus performs. It is a figure which shows the structural example of a purchase log. It is a flowchart for demonstrating an example of the process sequence of the total process of a purchase log. It is a figure which shows an example of a purchase history. It is a flowchart for demonstrating an example of the process sequence of the calculation process of a custom degree. It is a figure which shows the example of calculation of a purchase probability and a custom degree. It is a diagram showing a calculation example of NMI _d. It is a figure which shows the example of calculation of an influence degree. It is a figure which shows an example of the relationship between the purchase day in six weeks, and the influence degree of the day of the last day. It is a figure which shows an example of the relationship between the purchase date in a year, and the monthly influence degree of the last day. It is a figure which shows the example of a change of the time series of the calculation result of a custom degree and an influence degree.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the present embodiment, attention is paid to purchasing habits and non-habitability as an example of an attitude that can be observed from the purchase log. Examples of highly addictive products include, for example, breakfast bread that is purchased every day. Examples of highly non-customary products include desserts that are purchased by external factors such as TV commercials and are not usually purchased.

  In the habit level calculation method related to location information described in Non-Patent Document 3, the habit level indicating the habit and the degree of application is defined and visualized in consideration of both periodic habits and aperiodic habits, and the effect It is shown. The habit is expressed by four kinds of multinomial distributions, which consider both or one of the day of the week and the time of day, or not, and is treated as a weighted linear sum, and the degree of habit is calculated. At this time, a unique assumption is made in the location information log that only stays at the same place at the same time, a frequency distribution is created such that the total probability for all places is 1, and the relative probability for other places The degree of habit is defined from the size of.

  However, regarding the purchase log, it is necessary to consider simultaneous purchase of a plurality of products. For example, [Case 1] When a buyer buys both bread and coffee for 80 days out of 100 days, if a relative distribution is created, the purchase probability of bread and coffee is 50% respectively. Become. On the other hand, when [Case 2] bread is bought 80 days out of 100 days and coffee is bought 20 days out of 100 days, the bread distribution probability is 80% and the coffee purchase probability is 20% in the relative distribution. In [Case 1] and [Case 2], both cases have a purchase probability of 80/100 days when viewed with bread alone, but [Case 2] is judged to have a higher bread purchasing habit. End up. In the case of location information, because it was not possible to stay at location A and location B at the same time, it was suitable to use the probability of being in location A or location B as a habit, but in the purchase log, Since bread and coffee can be purchased at the same time, it is necessary to define the habit for each product individually and expand the calculation to calculate the degree of fit to the habit without using relative values.

  As an example of an individual habit definition for each product, a Sarima model (Seasonal Auto-Regressive Integrated Moving Average Model) (see Non-Patent Document 1), which is a classic method for modeling time-series data that is not limited to purchase logs, has been proposed. ing. The stationarity extracted by taking the difference in logs assuming periodicity such as one week is used, but in the SARIMA model, it is appropriate to consider the duration of habit (= stationarity here) It is not done and tends to rely heavily on the latest logs. Also in the habit degree calculation method related to location information, since the habit is learned only once using only a log of a certain period, the duration and change of habits cannot be considered.

  In addition, the habit level calculation method for location information uses the number of logs at all locations unique to the buyer for each day of the week and time of day when calculating the strength of the influence of habits considering the day of the week as a weight. However, it is considered that it becomes a more effective index in the trend analysis by extending the log so as to take into account the log bias for each target (each product in the purchase log).

  Therefore, in this embodiment, habits are defined using individual purchase probabilities at the time of each purchase log used to calculate the degree of habits for each product and the number of purchase days, and habits that can also be applied to purchase logs. A degree calculation technique is disclosed. Considering how much the buyer has purchased each product in the past (including non-periodic habits) and how long the habit has lasted, the degree of fit to the habits (habitual degree R ) Is calculated.

In addition, the habits of each buyer for each product, and the day of the week or month or time, such as time-day impact indicating whether the periodicity is how much influence W _d and the moon degree of influence W _m is defined. As a result, it is possible to extract the tendency that the purchase of each product by each purchaser occurs particularly frequently under specific conditions such as weekly, every summer, and only recently. In addition, by applying the calculation result of habit level R to visualization technology, clustering technology, etc., it becomes possible to utilize the purchasing habits as an example of purchasing attitude in marketing such as the formulation of sales and advertising strategies. .

  FIG. 1 is a diagram illustrating a hardware configuration example of an index value calculation apparatus according to an embodiment of the present invention. The index value calculation device 10 in FIG. 1 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a CPU 104, an interface device 105, a display device 106, an input device 107, and the like that are mutually connected by a bus B. .

  A program that realizes processing in the index value calculation apparatus 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program need not be installed from the recording medium 101 and may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program and also stores necessary files and data.

  The memory device 103 reads the program from the auxiliary storage device 102 and stores it when there is an instruction to start the program. The CPU 104 realizes functions related to the index value calculation device 10 according to a program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network. The display device 106 displays a GUI (Graphical User Interface) or the like by a program. The input device 107 includes a keyboard and a mouse, and is used for inputting various operation instructions.

  FIG. 2 is a diagram illustrating a functional configuration example of the index value calculation apparatus according to the embodiment of the present invention. In FIG. 2, the index value calculation apparatus 10 includes an input reception unit 11, a log totaling unit 12, a habit level calculation unit 13, an influence level calculation unit 14, a calculation result output unit 15, a time series change visualization unit 16, and a clustering unit 17. Etc. Each of these units is realized by processing that one or more programs installed in the index value calculation apparatus 10 cause the CPU 104 to execute. The index value calculation device 10 also uses the purchase log storage unit 21. The purchase log storage unit 21 can be realized using a storage device that can be connected to the auxiliary storage device 102 or the index value calculation device 10 via a network.

  The purchase log storage unit 21 stores a purchase log group. The purchase log is an example of history data recorded every time a product is purchased by a purchaser. The purchase log may be input from a POS (Point Of Sales system), for example.

The input receiving unit 11 receives a selection of a purchase log as a calculation target of the habit level R and the influence level W from the user of the index value calculation device 10. One selected purchase log is hereinafter referred to as a “target log”. However, the latest purchase log may be automatically set as the target log. The log totaling unit 12 includes, for the purchase log group stored in the purchase log storage unit 21 and the purchase log group and the target log before the target log related to the buyer and the product related to the target log, the custom degree R and The parameters used for calculating the influence level W are aggregated. The habit level calculation unit 13 calculates the habit level R (full-time habit level R _t , day-of-week habit level R _d , monthly habit level R _m ), which will be described later, on the basis of the counting result by the log totaling unit 12. The custom degree R is an example of an index value indicating a degree of fit to a custom related to purchase of a specific product by a specific purchaser. The influence degree calculation unit 14 calculates an influence degree W (day of the week influence degree W _d , month influence degree W _m ) based on the aggregation result by the log aggregation unit 12. The influence degree W is an example of an index value indicating how much the periodicity such as the day of the week, the month, and the time affects the habit of the purchaser. The calculation result output unit 15 uses the time-series change visualization unit 16 or the clustering unit 17 to perform output processing on the habit level and the influence level. The time series change visualization unit 16 visualizes (for example, displays) the temporal tendency of the purchase log using the habit level R and the influence level W. The clustering unit 17 performs purchase log clustering using the custom degree R and the influence degree W.

  Hereinafter, a processing procedure executed by the index value calculation apparatus 10 will be described. FIG. 3 is a flowchart for explaining an example of a processing procedure executed by the index value calculation apparatus.

  In step S <b> 100, the input receiving unit 11 receives a selection of a target log from the purchase log stored in the purchase log storage unit 21 from the user. Note that the process of FIG. 3 may be executed each time a new purchase log is input (generated), for example. In this case, the new purchase log may be the target log.

  FIG. 4 is a diagram illustrating a configuration example of a purchase log. As shown in FIG. 4, one purchase log includes a purchaser ID, a product name, an item name, a date, a day of the week, and an elapsed month. The purchaser ID is identification information of the purchaser of the product. The product name is a specific name of the purchased product. In the present embodiment, each product is identified by the product name. The item name is a category to which the product belongs, such as “coffee”. The date is the date (date) when the product was purchased. The day of the week is the day of the week to which the date belongs. The elapsed month is the number of months that have passed since the month of the first purchase log by the purchaser including products other than the product. The elapsed month may be calculated based on the first purchase log in step S100, for example. In FIG. 4, the alphabets shown in parentheses for each item are identifiers that represent the value of each item in the following description.

  Subsequently, the log totaling unit 12 is a purchase log in which the purchaser ID and the product name match the purchaser ID and the product name of the target log among the purchase logs stored in the purchase log storage unit 21. Aggregation processing is executed on the purchase log (hereinafter referred to as “related log”) whose date is before the date of the target log, the target log, and the like (S200).

  Subsequently, the habit level calculator 13 calculates the habit level R based on the counting result (S300). Subsequently, the influence degree calculation unit 14 calculates the influence degree W based on the aggregation result (S400).

  Note that steps S100 to S400 may be executed with each related log as a target log. By doing so, it is possible to obtain a time series habit level and a time series influence level. Alternatively, steps S100 to S400 may be executed every time a purchase log is input. That is, the custom degree R and the influence degree W are calculated for each purchase log.

  Subsequently, the calculation result output unit 15 outputs the habit level and the influence level (S500). Step S500 may be executed after steps S100 to S400 are executed for a plurality of purchase logs. That is, step S500 may be executed asynchronously with respect to steps S100 to S400.

  Next, details of step S200 will be described. FIG. 5 is a flowchart for explaining an example of a processing procedure of a purchase log totaling process.

In step S201, the log totaling unit 12 totals various days of purchase for the target log and the related log based on the buyer u, the product g, the day of the week d, and the elapsed month m included in the target log. . The number of purchase days is the number of days on which purchases are made. The type of purchase _{dates, C u, g, C u} , g, di, C u, g, mj is there. C indicates the number of purchase days that meets the condition indicated by the subscript C. The subscript i of d indicates each day of the week. The subscript j of m indicates each elapsed month. Therefore, _{Cu, g} indicates the number of days of purchase of the product g by the purchaser u. _{Cu, g, and di} indicate the number of days of purchase of the product g by the buyer u for each day of the week. Further _{, Cu, g, mj} indicates the number of days of purchase for each elapsed month of the product g by the purchaser u.

Of _{Cu, g, and di} , the value corresponding to the day of the week d of the target log is denoted as _{Cu, g, d} . A value corresponding to the elapsed month m of the target log among _{Cu, g, and mj} is denoted as _{Cu, g, and m} .

Subsequently, the log totaling unit 12 totals various days (S202). The types of days include D _{u, t} , D _{u, di} , and D _{u, mj} . D is the number of days within the period (hereinafter referred to as “evaluation period”) from the date of purchase of the buyer u to the date of the target log, including products other than the product u of the target log. Indicates. D _{u, t} indicates the total number of days in the evaluation period. _{Du, di} indicates the number of days of each day of the week in the evaluation period. D _{u, mj} indicates the number of days in each elapsed month within the evaluation period.

Of _{Du, di} , the value corresponding to the day of the week d of the target log is denoted as _{Du, d} . _{Further, D u,} among _mj, a value corresponding to the elapsed month m of target _{log, D u,} denoted as _mj.

  A tabulation example is shown based on a specific example about various purchase days and various days. FIG. 6 is a diagram illustrating an example of a purchase history. FIG. 6 shows an example in which the date of the first purchase log of the purchaser u related to the target log is February 1 (month). The day marked with a black circle (●) is the day corresponding to the date of any related log, and the day marked with an asterisk (★) is the day corresponding to the date of the target log. is there. That is, the day with a black circle or a star is the date when the product g is purchased by the purchaser u. In the purchase history shown in FIG. 6, the total results of various C and various D are as follows.

D _{u, t} = 42: number of days in the period from February 1 to March 14 (evaluation period) C _{u, g} = 18: number of days with black circles or stars in the evaluation period (that is, the number of purchase days)
_{Du, d month} = 6: Number of days corresponding to Monday in the evaluation period _{Du, d Tuesday} = 6: Number of days corresponding to Tuesday in the evaluation period _{Du, d Water} = 6: Corresponds to Wednesday in the evaluation period Number of days _{Du, d Thursday} = 6: Number of days corresponding to Thursday within the evaluation period _{Du, d Friday} = 6: Number of days applicable to Friday within the evaluation period _{Du, d Saturday} = 6: Corresponding to Saturday within the evaluation period Number of days _{Du, d Day} = D _{u, d} = 6: Number of days corresponding to the day of the week of the star (Sunday) within the evaluation period C _{u, g, d month} = 1: Number of purchase days on Monday C _{u, g, d Tuesday} = 1: Tuesday purchase days C _{u, g, d water} = 1: Wednesday purchase days C _{u, g, d Thursday} = 5: Thursday purchase days C _{u, g, d Friday} = 1: Friday purchase days C _{u, g, d Sat} = 4: number of purchasing days on Saturday C _{u, g, d day} = C _{u, g, d} = 5: Number of days of purchase on Sunday (Sunday) with star _{symbol Du, m0} = 28: Number of days in month 0 (February) _{Du, m1} = _{Du, m} = 14: Month 1 Number of days in the month (month including the star (March)) C _{u, g, m0} = 15: Number of days purchased in the elapsed month (February) of the 0th month C _{u, g, m1} = C _{u, g, m} = 3: Number of purchase days in the elapsed month of the first month (the month including the asterisk (March)) Next, the details of step S300 in FIG. 3 will be described. FIG. 7 is a flowchart for explaining an example of the processing procedure of the custom degree calculation processing.

In step S <b> 301, the habit level calculation unit 13 determines the total purchase probability P _t according to the purchaser u, the product g, the day of the week d, and the elapsed month m included in the target log based on the aggregation result of FIG. 5. The day-of-week purchase probability P _d and the monthly purchase probability P _m are calculated based on the following equation (1).

なお、全体購買確率Ｐ_ｔ、曜日購買確率Ｐ_ｄ、月購買確率Ｐ_ｍの意味は、以下の通りである。

The meanings of the overall purchase probability P _t , the day of week purchase probability P _d , and the monthly purchase probability P _m are as follows.

P _t : Product purchase probability for the total number of days D _{u, t in the} evaluation period P _d : Product items on the day of the week for the same day D _{u, d} as the evaluation date (date of the target log) in the evaluation period Purchasing probability P _m : Purchasing probability of the product in the month with respect to the number of days D _{u, m in the} same month as the evaluation target day in the evaluation period. According to the notation in the present embodiment, the purchaser u and the product g The various purchase probabilities inherent to are P _{u, g, t} , P _{u, g, d} , P _{u, g, m} , but here they are described as above for convenience. The same applies to the custom degree R and the influence degree W.

Subsequently, the customs degree calculation unit 13 calculates the total-period customs degree R _t , day-of-week customs degree R _d , and monthly customs degree R _m based on various purchase probabilities P and purchase days C as follows. (S302).

R _t = P _t lnC _{u, g}
R _d = P _d lnC _{u, g, d}
R _m = P _m lnC _{u, g, m}
The reason for multiplying the purchase probability P by the logarithm of the number of days of purchase C is to make it possible to grasp a longer habit as a stronger habit, even with the same probability. Here, the logarithm is adopted in order not to increase the weight of the purchase days C linearly, but a method of multiplying the numerator by the purchase days C as it is (determining the product of the purchase probability P and the purchase days C) is also conceivable. . By multiplying the numerator by a value proportional to the number of purchase days C, it is considered that the duration of the habit can be taken into account.

  For example, when the purchase days C are not multiplied, [Case 1] product g is purchased for 1 day in 2 days, and [Case 2] product g is purchased for 100 days in 200 days. Both purchase probabilities Pt are 50%, but when considering the degree of whether or not it is a habit (habitability level R), it is reasonable to obtain a higher value for [case2]. Conceivable. According to the present embodiment, the custom degree of [Case 1] is 0, the custom degree of [Case 2] is 2.30, and [Case 2] has a higher value. In the present embodiment, since the custom degree R is calculated for each purchase log when purchase is made, the number of purchase days does not become zero.

  FIG. 8 is a diagram illustrating a calculation example of the purchase probability and the custom degree. FIG. 8 shows a calculation example of the purchase probability and the custom degree based on the purchase history shown in FIG.

Next, details of step S400 of FIG. 3 will be described. The purchase habits of purchasers may be biased with respect to specific days of the week or months, such as “I often buy every Saturday” and “I often buy in summer”. In the present embodiment, the intensity of this bias is quantified, and the day-of-week influence degree W _d and the month influence degree W _m for the day of the week d or the elapsed month m regarding the purchase of the product g by the purchaser u are calculated.

  For calculating the influence degree W, a normalized mutual information amount NMI (Normalized Mutual Information) (see Non-Patent Document 4) is used. NMI is an index used in the evaluation of the clustering method, and indicates whether the correct answer label Y can be properly classified by the clustering result X. 0 ≦ NMI ≦ 1, and becomes 1 when both (X and Y) completely match.

ｘ_ｉは、ｉ番目のクラスタに分類されることを、ｙ_ｊはｊ番目の正解ラベルが付されていることを示し、Ｐは、各事象の生起確率を示す。

x _i indicates that it is classified into the i-th cluster, y _j indicates that the j-th correct label is attached, and P indicates the occurrence probability of each event.

In the present embodiment, in order to investigate the degree of bias regarding the day of the week and the elapsed month for the purchase days related to the purchase of the product g by the purchaser u, X is used to calculate the day of the week influence degree W _d , _In the calculation of _m , the NMI is calculated by considering the classification based on the elapsed month, and Y as the presence / absence of purchase. NMI related to the classification based on the day of the week is denoted as NMI _d, and NMI related to the classification based on the elapsed month is denoted as NMI _m . Expressions (3) and (4) are transformed into the following expressions (5) and (6), respectively.

曜日影響度Ｗ_ｄの算出では、｜ｘ_ｉ｜＝Ｄ_ｕ，ｄｉ（ｄ_ｉは月曜日から日曜日の各曜日）、｜ｙ_１｜＝Ｃ_ｕ，ｇ、｜ｙ_２｜＝Ｄ_ｕ，ｔ−Ｃ_ｕ，ｇ、｜ｘ_ｉ∩ｙ_１｜＝Ｃ_{ｕ，ｇ，ｄｉ}、｜ｘ_ｉ∩ｙ_２｜＝｜ｘ_ｉ｜−Ｃ_{ｕ，ｇ，ｄｉ}として、式（２）、（５）、（６）に適用される。

In calculating the day-of-week influence W _d , | x _i | = D _{u, di} (d _i is each day of the week from Monday to Sunday), | y ₁ | = C _{u, g} , | y ₂ | = D _{u, t} − C _{u, g} , | x _i ∩y ₁ | = C _{u, g, di} , | x _i ∩y ₂ | = | x _i | −C _{u, g, di} , formulas (2), (5), Applies to (6).

FIG. 9 is a diagram illustrating a calculation example of NMI _d . 9 shows calculation examples of NMI _d calculated based on the purchase history shown in FIG. 6 is shown. As the values of NMI _d and NMI _m are larger, the classification based on the day of the week or the elapsed month appropriately separates purchase / non-purchase, that is, the purchase is frequently performed on a specific day of the week or elapsed month.

  A simple NMI value is affected by the data period, for example, the shorter the data period, the larger the value. Therefore, the influence calculation unit 14 calculates the NMI value as

を用いて正規化した相対値である、

Is a relative value normalized using

を以下のように求める。

Is obtained as follows.

図９には、ＮＭＩ_ｄを正規化した相対値の算出例が右側の矩形内に示されている。図９に例示されているように、

9 shows, example of calculating normalized relative values are shown on the right side of the rectangle NMI _d. As illustrated in FIG.

は、対象ログに含まれている曜日ｄや経過月ｍの全ての日のみに購買が行われたと仮定したＮＭＩである。

Is an NMI that is assumed to have been purchased only on all days of the day d and the elapsed month m included in the target log.

の算出では、対象ログの曜日がｄのとき、｜ｘ_ｉ｜＝Ｄ_ｕ，ｄｉ、｜ｙ_１｜＝Ｄ_ｕ，ｄ、｜ｙ_２｜＝Ｄ_ｕ，ｔ−Ｄ_ｕ，ｄ、ｄ_ｉ＝ｄならば、｜ｘ_ｉ∩ｙ_１｜＝Ｄ_ｕ，ｄｉ、｜ｘ_ｉ∩ｙ_２｜＝０、ｄ_ｉ≠ｄならば、｜ｘ_ｉ∩ｙ_１｜＝０、｜ｘ_ｉ∩ｙ_２｜＝Ｄ_ｕ，ｄｉとして、数式（２）、（５）、（６）に適用される。

In the calculation, when the day of the week of the target log is d, | x _i | = D _{u, di} , | y ₁ | = D _{u, d} , | y ₂ | = D _{u, t} −D _{u, d} , d _i If | d, then | x _i ∩y ₁ | = D _{u, di} , | x _i ∩y ₂ | = 0, and if d _i ≠ d, | x _i ∩y ₁ | = 0, | x _i ∩y ₂ | = D _{u, di} , and is applied to the equations (2), (5), and (6).

  In the present embodiment, the impact calculation unit 14

に、ステップＳ３００と同様に購買日数の対数を乗じて、習慣の期間を考慮して影響度Ｗを算出する。

Similarly to step S300, the influence degree W is calculated by multiplying the logarithm of the number of purchase days and taking into account the habitual period.

図１０は、影響度の算出例を示す図である。図１０には、図６に示される購買履歴に基づいて算出される影響度の算出例が示されている。

FIG. 10 is a diagram illustrating an example of calculating the influence degree. FIG. 10 shows an example of calculating the degree of influence calculated based on the purchase history shown in FIG.

As described above, the day-of-week influence degree W _d indicates whether or not the purchase (purchase related to the target log) is likely to occur on each day of the week. Here, in order to pay attention to the bias of purchase / non-purchase on each day of the week, it is possible to extract not only a weekly habit such as only every Monday but also habits such as frequent purchases on Mondays and Tuesdays.

FIG. 11 is a diagram showing an example of the relationship between the purchase date and the day of the week influence on the last day in 6 weeks. In the period of 6 weeks, (1) 3 times on every day, (2) 6 times only on Monday, (3) 6 times on Monday and Tuesday, (4) The value of NMI _{d and} the value of the day of week influence W _d in the purchase log on the last day are shown.

In (1), the purchase probability is 50% for all days and the presence / absence of purchase is ambiguous, so the day-of-week influence _Wd is very low. On the other hand, in the cases (2), (3), and (4) where there is little bias in the purchase probability on each day of the week, the value of the day-of-week influence degree W _d increases. Even compared to the example of FIG. 10, (2), (3), (4), the can be seen that large values of day influence _{W d.}

Further, the monthly influence degree W _m indicates whether there is a purchase presence / absence bias for each elapsed month. FIG. 12 is a diagram illustrating an example of the relationship between the purchase date and the month influence level on the last day in one year. FIG. 12 shows (1) half of the month (from the 14th to the 16th), (2) only in December, (3) only from October to December, and (4) every day in the period of one year. When purchased, the value of NMI _{m and} the value of monthly influence W _m at the time of the purchase log on December 31 are shown.

In (1), since the purchase probability is about 50% for all elapsed months, the monthly influence degree W _m is very low. On the other hand, in the cases (2), (3), and (4) where there is little bias in the purchase probability within each month, the value of the monthly influence degree W _m increases. As can be seen in (2) and (3), the impact of the month on purchases that increase only in a specific month such as every summer due to seasonality, or when there are many purchases only in the specific month such as the latest due to changes in habits. the value of the degrees _{W d} becomes higher. This is because in the month when there was almost no purchase, the purchase presence / absence is strongly biased toward “no”, and the value of NMI increases.

Note that the day influence W _d and moon influence W _m, the value when there is a purchase in daily in periods becomes particularly large, correct labels purchase chromatic showing a purchase whether there is a purchase daily This is because the value of NMI becomes very large. In fact, since very few products that continue to be purchasing every day in a long period of time, it is being extracted as the value is large commodity degree of influence W, in the case of a day of the week effect of W _d is, was bought only on certain days of the week items In the case of the monthly influence degree W _m , it is confirmed that the product is often purchased only in a specific month.

  In addition, when steps S200 to S400 are executed not only for the target log but also for each related log, a time series habit level R and a time series influence degree W are obtained. In this case, the calculation result output unit 15 may use the time series change visualization unit 16 to output information indicating changes in the habit level R and the influence level W in time series as illustrated in FIG. .

FIG. 13 is a diagram illustrating a time-series change example of the calculation result of the habit level and the influence level. FIG. 13 is based on the purchase history shown in FIG. That is, according to FIG. 6, the number of purchase logs (that is, C _{u, g} ) is 18. Therefore, there are 18 plots for each custom degree R and each influence degree.

The higher the value of the day-of-week customs degree R _d , the easier the purchase of the product g by the buyer u occurs on the day of the week d. Looking at the day habits of R _d of FIG. 13, Thursday, Saturday and for Sunday it can be seen that the day habits of R _d is increasing each time the week has elapsed.

Day influence W _d shown in FIG. _13, a value of 0 seen in calculation example of the moon influence W _m, the day influence W _d day initial purchase for each day is performed, moon influence W _m Then, since it is the month when the first purchase was made, the influence W cannot be calculated. With respect to other purchase logs, it is possible to observe how the influence degree W changes according to the number of days elapsed. For example, when the day-of-week influence degree W _d indicating whether or not the habit is particularly likely to occur on a specific day of the week is seen, it can be seen that the day tends to increase every day. In particular, in March 4 Thursday, which is surrounded by a circle, from the fact that all of the purchase on Thursday of up to it have been made, the value of the day of the week effect of W _d is larger.

  In the present embodiment, the same purchaser purchases the same product once a day. Therefore, regarding the same product by the same buyer, the number of purchase logs matches the number of purchase days. If the same buyer purchases the same product a plurality of times, the plurality of purchase logs may be combined into a single purchase log and processed.

As described above, according to the present embodiment, it is possible to calculate the habit level R and the influence level W in consideration of the habit of purchasing. That is, it is possible to calculate an index value indicating a purchase attitude tendency from the purchase log. As a result, in understanding the purchase attitude, for example, the following effects (1) to (4) can be expected.
(1) Improving sales and advertising strategies by improving visualization analysis accuracy (2) Considering attitudes in segmentation analysis (3) Considering attitude changes in measuring sales strategy effects (4) Real-time considering the latest attitudes Recommended products under high addictive conditions Introduce new products that are similar to products that are often purchased on a daily basis, and under low addictive conditions, products that are not normally purchased or generally non-addictive It can be used to introduce products that are easy to purchase at high prices.

  In the above description, marketing using a purchase log has been described as an example. However, the present embodiment can be applied to various time series logs (history data) that are not limited to purchase logs. For example, if you observe the trend of error occurrence date using the system error log, depending on the type of system and the type of error, whether the trouble is likely to occur during the hot summer season or dry winter season, for example, It is expected that the tendency of errors, such as whether a trouble is likely to occur on a specific day of the week when the number of accesses suddenly changes, becomes possible.

  In addition, when applied to an access log to a Web page, for example, it is possible to observe a tendency that a Web page on which materials related to lectures by university teachers are easily accessed at a specific time such as a specific day of the week or before an exam. It is possible to obtain knowledge that it is necessary to avoid these periods when modifying Web pages.

In addition, when applied to the observation of time-series changes in the amount of smoking and the amount of alcohol consumed, if the habit level R exceeds a certain value, a warning that there is a possibility of harming the health is given, or the influence level W _{d of} the day of the week. If the value exceeds a certain value, it may be possible to present advice such as “Don't drink too much just because it is Friday” or “It is better to have a rest day for two days a week”.

  Furthermore, observing the day when a person who develops an infectious disease appears, there is a possibility that a correlation between a weekly holiday and an onset date will appear, and a tendency such as seasonality will appear in the onset time. For such a log, it can be considered that when the custom degree R exceeds a certain value, it can be found that the user is in a pandemic (always prevalent) state.

  Moreover, in this Embodiment, although the example which calculates the custom degree R and the influence degree W according to a day of the week or every elapsed month was shown, the unit of time which calculates the custom degree R and the influence degree W depends on an observation object. May be appropriately changed. For example, the habit level R and the influence level W such as time, time zone, and minute unit may be calculated.

As described above, according to the present embodiment, it is possible to calculate the habit degree R, the day-of-week influence degree W _d, and the month influence degree W _m for various time series logs (history data) including a purchase log. Therefore, visualization and clustering analysis considering log habits can be performed. In the purchase log, it becomes possible to consider the purchase habit as an example of the purchase attitude in the time series trend analysis / clustering analysis for the development of the sales / advertisement strategy. In addition, it is expected to be used for measuring the effectiveness of sales and advertising strategies that take into account changes in habituality, and for real-time product recommendations that take into account the latest attitudes.

  In the present embodiment, the purchaser is an example of an individual. The purchase of the product g (specific product) by the purchaser u (specific purchaser) is an example of predetermined behavior. The purchase log is an example of history data. The target log is an example of first history data. The date, day of the week, and elapsed month of the purchase log are examples of recording times. A purchase log related to the same purchaser is an example of a set of history data. The log totaling unit 12 is an example of a totaling unit. The custom degree calculation unit 13 is an example of a first calculation unit. Each day of the week is an example of a plurality of cycle times that visit each other with a certain interval from each other at the same cycle. That is, the moon, fire, water, tree, gold, soil, and day visit each other at a regular interval with a certain gap from each other. In addition, time etc. may be an example of a period time. The elapsed month is an example of a unit period. An elapsed week or elapsed time may be an example of a unit period.

_{Cu, g, and d} are examples of the first action count. _{Cu, g, m} is an example of the second number of actions. P _d is an example of a first action probability. P _m is an example of a second action probability. The habit degree R is an example of an index value indicating the degree of fit of a predetermined action to a habit.

_{Cu, g, and di} are examples of the fourth action count. _{Cu, g, mj} is an example of the fifth action count. D _{u, t} is an example of the entire period. _{Cu, g} is an example of the third and sixth actions. The influence degree calculation unit 14 is an example of a second calculation unit. The term on the left side of log ₂ in Equation (3) and Equation (5) is an example of the third action probability and the action probability for the entire period in claim 2. The term on the right side of log ₂ in Expression (3) and Expression (5) is an example of the action probability in each cycle period or the action probability in each unit period. The influence degree W is an example of an index value indicating the degree of influence on behavior.

  As mentioned above, although the Example of this invention was explained in full detail, this invention is not limited to such specific embodiment, In the range of the summary of this invention described in the claim, various deformation | transformation・ Change is possible.

DESCRIPTION OF SYMBOLS 10 Index value calculation apparatus 11 Input reception part 12 Log totaling part 13 Habit level calculation part 14 Influence degree calculation part 15 Calculation result output part 16 Time series change visualization part 17 Clustering part 21 Purchase log storage part 100 Drive apparatus 101 Recording medium 102 Auxiliary Storage device 103 Memory device 104 CPU
105 interface device 106 display device 107 input device B bus

Claims (8)

The first history data of a set of history data recorded for each predetermined action by an individual and the history data recorded before the first history data, and having a certain interval between each other in the same cycle. The first action count is totaled based on the history data recorded in the cycle time to which the recording time of the first history data belongs among the plurality of cycle times visited, or the first history data, The second action count is totaled based on history data recorded before the first history data and recorded in a unit period to which the recording time of the first history data belongs among a plurality of unit periods. Or based on the first history data and the history data recorded before the first history data, the first history data from the recording time of the first history data recorded for the individual A counting unit that counts the third action number in the whole period until the recording time,
Based on the first number of actions, a first action probability in the cycle period to which the recording time of the first history data belongs is calculated, or based on the second action number, the first history calculating a second behavior probability in the unit period in which the recording time belongs data, or on the basis of the third action times, the calculated third behavior probability Lifetime, the first behavior probability or The product of the second action probability or the third action probability and the first action number or the second action number or the value based on the third action number is the predetermined action by the individual. A first calculation unit that calculates an index value indicating the degree of fit to the habit;
I have a,
The totaling unit is based on history data recorded at the cycle time out of the first history data and history data recorded before the first history data for each of the plurality of cycle times. Then, the fourth action count in the period is counted, or, for each of the plurality of unit periods, the first history data and the history data recorded before the first history data Based on the history data recorded in the unit period, the fifth action count in the unit period is totaled, and the history data recorded before the first history data and the first history data Based on the above, the sixth action count in the entire period from the recording time of the first history data recorded for the individual to the recording time of the first history data is tabulated,
Using the mutual information amount based on the action probability for the whole period based on the sixth action count and the fourth action count for each period or the fifth action count for each unit period An index value calculation apparatus comprising: a second calculation unit that calculates an index value indicating a degree of influence on the behavior during the period or the unit period . The first history data of a set of history data recorded for each predetermined action by an individual and the history data recorded before the first history data, and having a certain interval between each other in the same cycle. The first action count is totaled based on the history data recorded in the cycle time to which the recording time of the first history data belongs among the plurality of cycle times visited, or the first history data, The second action count is totaled based on history data recorded before the first history data and recorded in a unit period to which the recording time of the first history data belongs among a plurality of unit periods. Or based on the first history data and the history data recorded before the first history data, the first history data from the recording time of the first history data recorded for the individual A counting unit that counts the third action number in the whole period until the recording time,
Based on the first number of actions, a first action probability in the cycle period to which the recording time of the first history data belongs is calculated, or based on the second action number, the first history calculating a second behavior probability in the unit period in which the recording time belongs data, or on the basis of the third action times, the calculated third behavior probability Lifetime, the first behavior probability or The product of the second action probability or the third action probability and the logarithm of the first action number or the second action number or the third action number is the product of the predetermined action by the individual. A first calculation unit that calculates the index value indicating the degree of fit to the habit;
An index value calculation device characterized by comprising: The first history data and the first history data in the set of history data recorded for each predetermined action by an individual for each of a plurality of cycle periods that visit each other with a certain interval from each other in the same cycle. Based on the history data recorded in the cycle period, among the history data recorded before the history data, the fourth action count in the cycle period is totaled, or for each of the plurality of unit periods, Based on the history data recorded in the unit period, of the first history data and the history data recorded before the first history data, the fifth action count in the unit period is aggregated. And from the recording time of the first history data recorded for the individual based on the first history data and the history data recorded before the first history data. A counting unit for counting the sixth action count in the whole period until the recording time of the first history data,
Using the mutual information amount based on the action probability for the whole period based on the sixth action count and the fourth action count for each period or the fifth action count for each unit period a second calculation unit for leaving calculate the index value indicating the degree of influence on the behavior of the periodic timing or the unit period,
An index value calculation device characterized by comprising: A first purchase log in a set of purchase logs recorded every time a product is purchased by a purchaser, and is recorded before the first purchase log, and the recording time of the first purchase log belongs. The first purchase count is totaled based on the purchase log recorded on the day of the week, or recorded before the first purchase log and the first purchase log, and the recording time of the first purchase log Based on the purchase log recorded in the month to which the data belongs, or based on the first purchase log and the purchase log recorded before the first purchase log, A totaling unit that counts the third number of purchases in the entire period from the recording time of the first purchase log by the purchaser to the recording time of the first purchase log;
Based on the first number of purchases, a first purchase probability for a day of the week to which the recording time of the first purchase log belongs is calculated, or based on the second number of purchases, the first purchase log The second purchase probability in the month to which the recording time belongs is calculated, or the third purchase probability in the whole period is calculated based on the third purchase count, and the first purchase probability or the second purchase probability is calculated. The product of the purchase probability or the third purchase probability and the value based on the first purchase count, the second purchase count, or the third purchase count, to the purchase habit of the product by the purchaser A first calculator that calculates an index value indicating the degree of fit of
I have a,
For each day of the week, the counting unit determines whether the day of the week is based on the purchase log recorded on that day of the first purchase log and the purchase log recorded before the first purchase log. The fourth purchase count is counted, or, for each of a plurality of months, the purchase recorded in the month among the first purchase log and the purchase log recorded before the first purchase log. Based on the log, the fifth purchase count in the month is tabulated, and the purchaser is recorded based on the first purchase log and the purchase log recorded before the first purchase log. Totaling the sixth number of purchases over the entire period from the recorded time of the first purchase log to the time of recording the first purchase log,
Using the mutual information amount based on the purchase probability for the whole period based on the sixth purchase count and the fourth purchase count for each day of the week, or the fifth purchase count for each month, An index value calculation apparatus comprising: a second calculation unit that calculates an index value indicating a degree of influence on the purchase of a day of the week or the month . A first purchase log in a set of purchase logs recorded every time a product is purchased by a purchaser, and is recorded before the first purchase log, and the recording time of the first purchase log belongs. The first purchase count is totaled based on the purchase log recorded on the day of the week, or recorded before the first purchase log and the first purchase log, and the recording time of the first purchase log Based on the purchase log recorded in the month to which the data belongs, or based on the first purchase log and the purchase log recorded before the first purchase log, A totaling unit that counts the third number of purchases in the entire period from the recording time of the first purchase log by the purchaser to the recording time of the first purchase log;
Based on the first number of purchases, a first purchase probability for a day of the week to which the recording time of the first purchase log belongs is calculated, or based on the second number of purchases, the first purchase log The second purchase probability in the month to which the recording time belongs is calculated, or the third purchase probability in the whole period is calculated based on the third purchase count, and the first purchase probability or the second purchase probability is calculated. The product of the purchase probability or the third purchase probability and the logarithm of the first purchase count or the second purchase count or the third purchase count is added to the purchase habit of the product by the purchaser. A first calculator that calculates an index value indicating the degree of fit;
An index value calculation device characterized by comprising: Computer
The first history data of a set of history data recorded for each predetermined action by an individual and the history data recorded before the first history data, and having a certain interval between each other in the same cycle. The first action count is totaled based on the history data recorded in the cycle time to which the recording time of the first history data belongs among the plurality of cycle times visited, or the first history data, The second action count is totaled based on history data recorded before the first history data and recorded in a unit period to which the recording time of the first history data belongs among a plurality of unit periods. Or based on the first history data and the history data recorded before the first history data, the first history data from the recording time of the first history data recorded for the individual And a summary procedure to aggregate the third action number of times in the entire period of up to record time,
Based on the first number of actions, a first action probability in the cycle period to which the recording time of the first history data belongs is calculated, or based on the second action number, the first history calculating a second behavior probability in the unit period in which the recording time belongs data, or on the basis of the third action times, the calculated third behavior probability Lifetime, the first behavior probability or The product of the second action probability or the third action probability and the first action number or the second action number or the value based on the third action number is the predetermined action by the individual. A first calculation procedure for calculating as an index value indicating the degree of fit to the habit,
The execution,
The counting procedure is based on history data recorded at the cycle time out of the first history data and history data recorded before the first history data for each of the plurality of cycle times. Then, the fourth action count in the period is counted, or, for each of the plurality of unit periods, the first history data and the history data recorded before the first history data Based on the history data recorded in the unit period, the fifth action count in the unit period is totaled, and the history data recorded before the first history data and the first history data Based on the above, the sixth action count in the entire period from the recording time of the first history data recorded for the individual to the recording time of the first history data is tabulated,
The computer is
Using the mutual information amount based on the action probability for the whole period based on the sixth action count and the fourth action count for each period or the fifth action count for each unit period An index value calculation method comprising: executing a second calculation procedure for calculating an index value indicating a degree of influence on the behavior during the period or the unit period . Computer
The first history data of a set of history data recorded for each predetermined action by an individual and the history data recorded before the first history data, and having a certain interval between each other in the same cycle. The first action count is totaled based on the history data recorded in the cycle time to which the recording time of the first history data belongs among the plurality of cycle times visited, or the first history data, The second action count is totaled based on history data recorded before the first history data and recorded in a unit period to which the recording time of the first history data belongs among a plurality of unit periods. Or based on the first history data and the history data recorded before the first history data, the first history data from the recording time of the first history data recorded for the individual And a summary procedure to aggregate the third action number of times in the entire period of up to record time,
Based on the first number of actions, a first action probability in the cycle period to which the recording time of the first history data belongs is calculated, or based on the second action number, the first history calculating a second behavior probability in the unit period in which the recording time belongs data, or on the basis of the third action times, the calculated third behavior probability Lifetime, the first behavior probability or The product of the second action probability or the third action probability and the logarithm of the first action number or the second action number or the third action number is the product of the predetermined action by the individual. A first calculation procedure for calculating as an index value indicating the degree of fit to a habit;
The index value calculation method characterized by performing.   The program for functioning a computer as each part as described in any one of Claims 1 thru | or 5.

Images