Computer program, user behavior sequence prediction system based on payment data, user behavior sequence prediction method based on payment data, and learning model construction method

Abstract

This invention provides a computer program capable of predicting or generating transportation-related payment information that can be used to predict customer behavior, a user behavior sequence prediction system based on payment data, a user behavior sequence prediction method based on payment data, and a method for constructing a learning model. [Solution] The computer program acquires time-series data of the customer's transportation payment information up to a first point in time. When this time-series data of transportation payment information up to the first point in time is input, the acquired time-series data is input into a first learning model that predicts the transportation payment information at a second point in time after the first point in time. The computer then performs a process to predict the customer's transportation payment information at the second point in time.

Description

Technical Field

[0006] ,

[0001] The present invention relates to a computer program, a user behavior sequence prediction system based on settlement data, a user behavior sequence prediction method based on settlement data, and a learning model construction method.

Background Art

[0002] In recent years, the importance of one-to-one marketing, which targets many people with the same approach, has been increasing instead of mass marketing. One-to-one marketing reads the needs of each customer and proposes products and services that match the customer's tastes. In order to effectively conduct one-to-one marketing, it is necessary to collect information such as customer information, past purchase history, and behavior history to predict customer behavior.

[0003] Patent Document 1 discloses an apparatus that acquires customer information regarding a customer from a system of a service used by the customer, processes the acquired customer information into feature quantity information, and predicts customer behavior based on the processed feature quantity information.

[0004] On the other hand, in the field of POI (Point Of Interest) prediction, the customer's destination is predicted from the history information of SNS (for example, a check-in app) to recommend the customer's POI. There is also a technique for analyzing the customer's movement sequence using the GPS information of the smartphone carried by the customer.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] The customer information disclosed in Patent Document 1 is limited to information such as gender, age, residential area, occupation, and membership type. Furthermore, since POI prediction assumes that the customer themselves operates the check-in app, it does not necessarily cover all places the customer has visited, and the historical information is limited. In addition, GPS information is not linked to location information, resulting in poor interpretability of the customer's movement and making it difficult to predict customer behavior.

[0007] This invention has been made in view of the above circumstances, and aims to provide a computer program capable of predicting or generating transportation-related payment information that can be used to predict customer behavior, a user behavior sequence prediction system based on payment data, a user behavior sequence prediction method based on payment data, and a learning model construction method. [Means for solving the problem]

[0008] The present invention includes several means for solving the above problem, but to give one example, a computer program acquires time-series data of a customer's transportation payment information up to a first point in time, and when the time-series data of transportation payment information up to the first point in time is input, the computer is made to input the acquired time-series data into a first learning model that predicts transportation payment information at a second point in time after the first point in time, thereby causing the computer to perform a process to predict the customer's transportation payment information at the second point in time. [Effects of the Invention]

[0009] According to the present invention, it is possible to predict or generate transportation-related payment information. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of the configuration of the user behavior sequence prediction system based on payment data in this embodiment. [Figure 2] This figure shows an example of passenger boarding and alighting information. [Figure 3] This figure shows an example of predicting boarding and alighting information using a learning model. [Figure 4] It is a diagram showing an example of boarding and alighting information at each time point. [Figure 5] It is a diagram showing an example of the configuration of a learning model. [Figure 6] It is a diagram showing an example of the prediction result of boarding and alighting information. [Figure 7] It is a diagram showing an example of congestion prediction at a station. [Figure 8] It is a diagram showing an example of congestion prediction between stations. [Figure 9] It is a diagram showing an example of settlement information. [Figure 10] It is a diagram showing an example of the prediction of settlement information by a learning model. [Figure 11] It is a diagram showing an example of settlement information at each time point. [Figure 12] It is a diagram showing an example of the prediction result of settlement information. [Figure 13] It is a diagram showing an example of settlement prediction. [Figure 14] It is a diagram showing an example of the generation of boarding and alighting information by a learning model. [Figure 15] It is a diagram showing the first example of attribute information. [Figure 16] It is a diagram showing the second example of attribute information. [Figure 17] It is a diagram showing an example of environment information. [Figure 18] It is a diagram showing an example of customer behavior information and demand prediction based on the boarding and alighting information generated by a learning model. [Figure 19] It is a diagram showing an example of the comparison between the generation result of boarding and alighting information and the actual performance of boarding and alighting information. [Figure 20] It is a diagram showing an example of the generation of settlement information by a learning model. [Figure 21] It is a diagram showing an example of the procedure for predicting traffic settlement information by an information processing device. [Figure 22] It is a diagram showing an example of the procedure for generating traffic settlement information by an information processing device.

Modes for Carrying Out the Invention

[0011] Embodiments of the present invention will be described below with reference to the drawings. Figure 1 is a diagram showing an example of the configuration of a user behavior sequence prediction system based on payment data according to this embodiment. The user behavior sequence prediction system based on payment data includes an information processing device 50. The information processing device 50, as a user behavior sequence prediction system based on payment data, can be connected via a communication network 1 to a display device 10, a transportation-related payment information DB (database) 20, an attribute information DB 30, an environmental information DB 40, and a large-scale language model 100.

[0012] The Transportation Payment Information DB20 records transportation payment information associated with each customer. This information includes boarding and alighting information, as well as payment information. Boarding and alighting information is read by the automatic ticket gate when a customer passes through a station using their IC card ticket and recorded in the Transportation Payment Information DB20. IC cards have functions such as prepaid tickets, commuter passes, and electronic money. Payment information is read by the POS (Point of Sale) register or other device in the store or vending machine when a customer purchases goods (such as merchandise) using their IC card ticket and recorded in the Transportation Payment Information DB20. Further details on transportation payment information will be described later.

[0013] The attribute information DB30 records attribute information associated with each customer. Attribute information may be obtained from the customer, for example, when they register their IC card for transit. Details of the attribute information will be described later.

[0014] The environmental information DB40 contains environmental information. Details of the environmental information will be described later.

[0015] The large-scale language model 100 is a machine learning model that utilizes natural language processing techniques to learn language patterns from existing text data and can generate or understand text such as sentences and conversations. The large-scale language model 100 divides text data into small chunks (or tokens) and performs an embedding process on each divided chunk to generate a vector representation (embedding) of each chunk. The embedding process includes numerical representations of the chunk type and numerical representations of the chunk's positional relationship. As a result, each chunk is represented as a point in a high-dimensional space, where chunks with similar meanings are placed close to each other, and chunks with different meanings are placed far apart. The large-scale language model 100 can numerically capture the semantic relationships and similarities between chunks and understand the semantics, i.e., meaning, of text data. Furthermore, the large-scale language model 100 is equipped with an attention mechanism, and by appropriately adjusting the magnitude of the linear combination coefficient when linearly combining the numerical representations of each chunk calculated by the embedding process, it can obtain a numerical representation that expresses the complex structure of a sentence.

[0016] The large-scale language model 100 is composed of deep neural networks and can use models such as GPT-4, GPT-3.5, BERT, LaMDA, PaLM, LLaMA, NVIDIA, or new models that will be developed and put into use in the future.

[0017] The information processing device 50 includes a control unit 51 that controls the entire device, a communication unit 52, a memory 53, a display unit 54, an operation unit 55, and a storage unit 56.

[0018] The control unit 51 may be configured by incorporating a required number of CPUs (Central Processing Units), MPUs (Micro-Processing Units), GPUs (Graphics Processing Units), etc. Alternatively, the control unit 51 may be configured by combining DSPs (Digital Signal Processors), FPGAs (Field-Programmable Gate Arrays), etc.

[0019] The communication unit 52 is equipped with a communication module and has the function of communicating with the display device 10, the transportation payment information DB (database) 20, the attribute information DB 30, the environment information DB 40, and the large-scale language model 100 via the communication network 1. The communication unit 52 can acquire (receive) transportation payment information from the transportation payment information DB 20, acquire attribute information from the attribute information DB 30, and acquire environment information from the environment information DB 40.

[0020] The display unit 54 can be made of a liquid crystal panel or an organic EL (Electro-Luminescence) display, etc. An external display device may be provided instead of the display unit 54.

[0021] The operation unit 55 consists of a touch panel or the like, and allows for the operation of icons displayed on the display unit 54, cursor movement and manipulation, and input of text, etc. The operation unit 55 may also be a mouse or keyboard.

[0022] The memory unit 56 can be made up of semiconductor memory or a hard disk, and stores a computer program 57 (program product), a learning model 58 as a first learning model, a learning model 59 as a second learning model, and other necessary information.

[0023] The computer program 57 can be read by a recording medium (e.g., an optically readable disc storage medium such as a CD-ROM) M using a recording medium reading unit (not shown) and stored in a storage unit 56. The computer program 57 may also be read by a recording medium such as a storage device (semiconductor memory such as an SSD (Solid State Drive)) connected by a standard for connecting to a computer (e.g., USB (Universal Serial Bus) or other standards) and stored in a storage unit 56. Alternatively, the computer program 57 may be downloaded from an external device via a communication unit 52 and stored in a storage unit 56. Details of the learning models 58 and 59 will be described later.

[0024] The memory 53 can be composed of semiconductor memory such as SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), or flash memory. The computer program 57 is loaded into the memory 53, and the control unit 51 can execute the computer program 57. The control unit 51 can execute the processing defined in the computer program 57. In other words, the processing performed by the control unit 51 is also the processing performed by the computer program 57.

[0025] Next, I will explain the forecast for transportation-related payment information.

[0026] Figure 2 shows an example of boarding and alighting information. As shown in Figure 2, boarding and alighting information includes information such as entry or exit, travel time, day of the week, and station. Entry indicates that the customer entered the station through the station's automatic ticket gate. Exit indicates that the customer left the station through the station's automatic ticket gate. Entry means boarding, and exit means alighting; the distinction between entry and exit is also called the distinction between boarding and alighting. The distinction between entry and exit can be identified, for example, by two labels.

[0027] The travel time can be the time when the customer passes through the automatic ticket gate at the station, and can be expressed as, for example, 0:00, 1:00, 2:00, ..., 23:00 for a 24-hour day. The travel time can be identified, for example, by 24 labels.

[0028] The day of the week can be the day the customer passes through the station's automatic ticket gate, and can be represented by Sunday, Monday, Tuesday, ..., Saturday. The day of the week can be identified, for example, by seven labels.

[0029] A station can be the name of the station when a customer passes through the automatic ticket gate. A station can be a station where a customer's IC card ticket is usable. A station can be identified by a number of labels corresponding to the number of stations.

[0030] Figure 3 shows an example of boarding and alighting information prediction by the learning model 58. The control unit 51 acquires time-series data of boarding and alighting information up to the first point in time for the customer, and inputs the acquired time-series data into the learning model 58. In the example in Figure 3, boarding and alighting information for time points Z1, Z2, ..., Z(n-1) has been input into the learning model 58. Time point Z(n-1) corresponds to the first point in time (for example, the current time).

[0031] The learning model 58 is designed to predict boarding and alighting information for a second time point, which is after the first time point, given time-series data of boarding and alighting information up to the first time point as input. Therefore, as shown in Figure 3, the learning model 58 outputs boarding and alighting information for time point Z(n) as predicted information. Time point Z(n) corresponds to the second time point (for example, in the future).

[0032] The control unit 51 acquires time-series data of the customer's boarding and alighting information (transportation payment information) up to the first point in time, and training data including boarding and alighting information at a second point in time after the first point in time. Based on the acquired training data, it can build a learning model that predicts the boarding and alighting information at the second point in time when time-series data of boarding and alighting information up to the first point in time is input.

[0033] Figure 4 shows an example of boarding and alighting information at each point in time. Figure 4 shows a specific example of boarding and alighting information for points in time Z1, Z2, ..., Z(n-1) as illustrated in Figure 3. As shown in Figure 4, A1 represents whether it is an entry or exit at point Z1, T1 represents the travel time, W1 represents the day of the week, and L1 represents the station. The entry or exit A1 can be represented by two labels, for example, A1(1): entry, A1(2): exit. The travel time T1 can be represented by 24 labels, for example, T1(1): 0:00, T1(2): 1:00, T1(3): 2:00, ..., T1(24): 23:00. The day of the week W1 can be represented by seven labels, for example, W1(1): Sunday, W1(2): Monday, ..., W1(7): Saturday. Station L1 can be represented by a number of labels corresponding to the number of stations, for example, L1(1): Osaka Station, L1(2): Shin-Osaka Station, L1(3): Higashi-Yodogawa Station, L1(4): Suita Station, and so on.

[0034] As mentioned above, time-series data of boarding and alighting information can also be considered as word-series data. In other words, the learning model 58 can predict future travel and payment patterns of individual customers when given time-series data of their past travel and payment patterns.

[0035] Figure 5 shows an example of the configuration of the learning model 58. The learning model 58 can use a Transformer model and includes an encoder 581 and a decoder 582. The encoder 581 extracts features from the encoder input (input sequence) and outputs the extracted features to the decoder 582. The decoder 582 predicts the decoder output (output sequence) from the decoder input (input sequence) based on the features extracted by the encoder 581. The decoder 582 can output the probability (accuracy) of the predicted boarding and alighting information.

[0036] In the example shown in Figure 5, 99 time series (Z1, Z2, ..., Z99) are used as input, and boarding / alighting information for time Z100 is predicted. Specifically, time series data of boarding / alighting information for time Z1, Z2, ..., Z97 are input to encoder 581, and boarding / alighting information for time Z97, Z98, and Z99 is input to decoder 582. Note that the number of time points is not limited to 100.

[0037] The learning model 58 can be trained using a labeled dataset. For the training dataset, one should prepare the following: time-series data of boarding and alighting information at each point in time during the feature extraction period as input to the encoder 581; time-series data of boarding and alighting information at each point in time during the prediction period as input to the decoder 582; and a set of measured boarding and alighting values ​​during the prediction period as output data from the decoder 582.

[0038] The control unit 51 can train the learning model 58 using the input data from the encoder 581, the input data from the decoder 582, and the output data from the decoder 582, so as to minimize the error between the predicted value and the actual value of the learning model 58 during the prediction period.

[0039] Figure 6 shows an example of the prediction results for boarding and alighting information. The control unit 51 can predict boarding and alighting information at a second time point for each customer. As shown in Figure 6, for customer ID=3001, the boarding and alighting information at time point Z1 is "Station S11_17:00 entry_Friday", ..., and the boarding and alighting information at time point Z(n-1) is "Station S13_9:00 exit_Monday". The prediction for boarding and alighting information at time point Z(n) is "Station S15_10:00 entry_Monday".

[0040] For customer ID=3002, the boarding / alighting information at time Z1 is "Station S01_18:00 Exit_Tuesday", ..., and the boarding / alighting information at time Z(n-1) is "Station S05_16:00 Enter_Thursday". The predicted boarding / alighting information at time Z(n) is "Station S07_17:00 Exit_Thursday".

[0041] For customer ID=3003, the boarding / alighting information at time Z1 is "Station S10_10am entry_Sunday", ..., and the boarding / alighting information at time Z(n-1) is "Station S20_12pm entry_Sunday". The predicted boarding / alighting information at time Z(n) is "Station S20_4pm exit_Sunday".

[0042] For customer ID=3004, the boarding / alighting information at time Z1 is "Station S40_14:00 Exit_Wednesday", ..., and the boarding / alighting information at time Z(n-1) is "Station S35_13:00 Exit_Friday". The predicted boarding / alighting information at time Z(n) is "Station S30_15:00 Enter_Friday". The same applies to other customers.

[0043] As described above, the control unit 51 acquires time-series data of the customer's boarding and alighting information (transportation payment information) up to the first point in time. When time-series data of boarding and alighting information up to the first point in time is input, the control unit 51 inputs the acquired time-series data into a learning model 58 that predicts boarding and alighting information at a second point in time after the first point in time, thereby predicting the customer's boarding and alighting information at the second point in time. The control unit 51 can output the predicted boarding and alighting information to the display device 10. The display device 10 can display the predicted boarding and alighting information.

[0044] The above configuration makes it possible to predict transportation-related payment information (e.g., boarding and alighting information) that can be used to predict customer behavior.

[0045] Next, we will discuss congestion prediction, which is related to predicting customer behavior.

[0046] Figure 7 shows an example of station congestion prediction. Figure 7 also shows an example of a railway route map. In the figure, rectangles and circles represent stations. The control unit 51 classifies the predicted boarding and alighting information for each customer by time. For each classified time, the number of people entering and exiting at each station is aggregated to calculate the number of people boarding and alighting at each station at each time. A threshold Th1 for high congestion and a threshold Th2 for medium congestion are predetermined for each station, and depending on whether the calculated number of people boarding and alighting exceeds threshold Th1 or threshold Th2, the station congestion prediction can be illustrated on the route map.

[0047] Note that Figure 7 shows station congestion information (prediction) at a given time, but the congestion information (congestion prediction for each station) will change as time progresses.

[0048] As described above, the control unit 51 collects station boarding and alighting information for multiple customers at a second point in time, as predicted by the learning model 58, and can predict time-specific congestion information for each station based on the collected station boarding and alighting information. This makes it possible to predict customer behavior from the perspective of station congestion.

[0049] The control unit 51 can output guidance information to help avoid congestion based on predicted congestion information for each time of day. This guidance information can be output to, for example, a video display device (digital signage) installed at a train station or other location.

[0050] For example, the information provided could include maps showing crowded shops within the station, suggesting ways to pass the time by shopping or eating, thereby reducing congestion.

[0051] Figure 8 shows an example of congestion prediction between stations. The control unit 51 classifies the predicted boarding and alighting information for each customer by time. For each classified time, it calculates the number of boarding and alighting passengers at each station by aggregating the number of passengers entering and exiting at each station. A threshold Th1 for high congestion and a threshold Th2 for medium congestion are predetermined for each station, and the congestion level is calculated for each station depending on whether the calculated number of boarding and alighting passengers exceeds threshold Th1 or threshold Th2. By connecting stations for which the same congestion level has been calculated, the control unit 51 can illustrate the congestion prediction between stations (lines) on a route map.

[0052] Note that Figure 8 shows congestion information (prediction) between stations at a certain time, but the congestion information (congestion prediction between stations) will change as time progresses.

[0053] As described above, the control unit 51 collects station boarding and alighting information for multiple customers at a second point in time, as predicted by the learning model 58, and can predict time-specific congestion information between stations based on the collected station boarding and alighting information. This makes it possible to predict customer behavior from the perspective of line congestion.

[0054] The control unit 51 can output guidance information to help avoid congestion based on predicted congestion information for each time of day. This guidance information can be output to, for example, a video display device (digital signage) installed at a train station or other location.

[0055] The information provided includes, for example, transfer instructions to other lines on crowded routes, suggestions for boarding trains during off-peak hours, and offers for discounts on special trains. This helps alleviate the mental stress caused by congestion for customers and contributes to increased customer satisfaction.

[0056] Figure 9 shows an example of payment information. As shown in Figure 9, payment information includes information such as the date and time of payment, the amount of payment, and the store where the payment was made. The date and time of payment can be expressed as year, month, day, and time. The time can be expressed as, for example, 0:00, 1:00, 2:00, ..., 23:00 for a 24-hour day. The date and time of payment can be identified by, for example, 24 labels in addition to the year, month, and day labels. Note that the labels for months and days within a year can be identified by 365 labels.

[0057] The payment amount is the purchase price of the goods (items), and can be categorized as follows: 500 yen or less, 1000 yen or less, 2000 yen or less, 5000 yen or less, 10,000 yen or less, 10,000 yen or more, etc. The payment amount can be identified, for example, by six labels.

[0058] The payment store can be the name of the store where the customer purchased the goods. The payment store can also be a store where the customer's IC card is accepted. The payment store can be identified by a number of labels corresponding to the number of stores.

[0059] Figure 10 shows an example of payment information prediction by the learning model 58. The control unit 51 acquires time-series data of the customer's payment information up to the first time point and inputs the acquired time-series data into the learning model 58. In the example in Figure 10, payment information for time points Z1, Z2, ..., Z(n-1) is input into the learning model 58. Time point Z(n-1) corresponds to the first time point (for example, the current time).

[0060] The learning model 58 is designed to predict settlement information for a second time point, which is later than the first time point, given time-series data of settlement information up to the first time point as input. Therefore, as shown in Figure 10, the learning model 58 outputs settlement information for time point Z(n) as predicted information. Time point Z(n) corresponds to the second time point (for example, in the future).

[0061] The control unit 51 acquires time-series data of the customer's payment information (transportation-related payment information) up to the first point in time, and training data including payment information from a second point in time after the first point in time. Based on the acquired training data, it can build a learning model that predicts the payment information at the second point in time when time-series data of payment information up to the first point in time is input.

[0062] Figure 11 shows an example of payment information at each point in time. Figure 11 shows a specific example of payment information at points in time Z1, Z2, ..., Z(n-1) as illustrated in Figure 10. As shown in Figure 11, the payment store at point in time Z1 is represented by P1, the payment date and time is represented by D1, and the payment amount is represented by M1. The payment store P1 can be represented by a number of labels corresponding to the number of stores, for example, P1(1): Osaka store, P1(2): Kyoto store, P1(3): Kobe store, ... The time within the payment date and time D1 can be represented by 24 labels, for example, D1(1): 0:00, D1(2): 1:00, D1(3): 2:00, ..., D1(24): 23:00. The payment amount M1 can be represented by six labels, for example, M1(1): 500 yen or less, M1(2): 1000 yen or less, M1(3): 2000 yen or less, M1(4): 5000 yen or less, M1(5): 10,000 yen or less, and M1(5): 10,000 yen or more.

[0063] Figure 12 shows an example of the predicted payment information. The control unit 51 can predict the payment information at a second point in time for each customer. As shown in Figure 12, for customer ID=4001, the payment information at time Z1 is "April 10th 12:00_500 yen_Store ST01", ..., and the payment information at time Z(n-1) is "April 15th 16:00_1000 yen_Store ST04". The predicted payment information at time Z(n) is "April 16th 9:00_500 yen_Store ST03".

[0064] For customer ID=4002, the payment information at time Z1 is "May 20th 9am_2000 yen_Store ST02", ..., and the payment information at time Z(n-1) is "May 20th 7pm_1000 yen_Store ST05". The predicted payment information at time Z(n) is "May 21st 9am_1000 yen_Store ST03".

[0065] For customer ID=4003, the payment information at time Z1 is "November 8th 13:00_1000 yen_Store ST05", ..., and the payment information at time Z(n-1) is "November 20th 12:00_1000 yen_Store ST09". The predicted payment information at time Z(n) is "November 25th 15:00_1000 yen_Store ST06". The same applies to other customers.

[0066] As described above, the control unit 51 acquires time-series data of the customer's payment information (transportation-related payment information) up to the first point in time. When time-series data of payment information up to the first point in time is input, the control unit 51 inputs the acquired time-series data into a learning model 58 that predicts the payment information at a second point in time, which is after the first point in time, and can predict the customer's payment information at the second point in time. The control unit 51 can output the predicted payment information to the display device 10. The display device 10 can display the predicted payment information.

[0067] The above configuration makes it possible to predict transportation-related payment information (e.g., payment information) that can be used to predict customer behavior.

[0068] Next, we will discuss payment predictions related to predicting customer behavior.

[0069] Figure 13 shows an example of a payment forecast. In Figure 13, the vertical axis represents the payment amount, and the horizontal axis represents the store where the payment was made. The control unit 51 collects predicted payment information for each customer and classifies the collected payment information by store. This makes it possible to calculate the payment amount for multiple customers for each store. The payment amount in Figure 13 corresponds to the total payment amount for customers who purchased goods at the same store.

[0070] Figure 13 allows us to predict the total payment amount (the sum of the payment amounts for each individual customer who made a purchase at that store) for each store. Note that while Figure 13 shows the predicted payment amount at a specific time, the predicted payment amount will change as time progresses.

[0071] This allows us to understand which stores customers are likely to purchase products from, and to predict customer behavior from the perspective of product purchases.

[0072] Next, we will explain how to generate transportation-related payment information.

[0073] Figure 14 shows an example of the generation of boarding and alighting information by the learning model 59. The control unit 51 acquires time-series data of boarding and alighting information, attribute information, and environmental information up to the first point in time for the customer, and inputs the acquired time-series data into the learning model 59. In the example in Figure 14, boarding and alighting information, attribute information, and environmental information for time points Z1, Z2, ..., Z(n-1) are input into the learning model 59. Time point Z(n-1) corresponds to the first point in time (for example, the current time).

[0074] The control unit 51 acquires attribute information and environmental information at time Z(n), and inputs the acquired attribute information and environmental information into the learning model 59.

[0075] The learning model 59 is constructed to generate boarding and alighting information for a second time point, which is later than the first time point, when it receives time-series data of boarding and alighting information, attribute information, and environmental information up to the first time point, as well as attribute information and environmental information for time point Z(n). Therefore, as shown in Figure 14, the learning model 59 generates and outputs boarding and alighting information for time point Z(n). Time point Z(n) corresponds to the second time point (for example, in the future).

[0076] The control unit 51 acquires training data including time-series data of the customer's boarding and alighting information (transportation payment information), attribute information, and environmental information up to the first point in time, as well as boarding and alighting information for a second point in time after the first point in time. Based on the acquired training data, it can construct a learning model 59 that generates boarding and alighting information for a second point in time when time-series data of the customer's boarding and alighting information up to the first point in time, attribute information, and environmental information are input.

[0077] More specifically, the control unit 51 acquires time-series data of the customer's boarding and alighting information (transportation payment information), attribute information, and environmental information up to the first point in time, as well as training data including attribute information, environmental information, and boarding and alighting information at a second point in time after the first point in time. Based on the acquired training data, the control unit 51 can construct a learning model 59 that generates boarding and alighting information at a second point in time when time-series data of the customer's boarding and alighting information up to the first point in time, attribute information, and environmental information, and attribute information and environmental information at a second point in time are input.

[0078] The attribute information and environmental information at time Z(n) specify the conditions for generating boarding and alighting information at time Z(n) in the simulation. In other words, by virtually changing the attribute information and environmental information at time Z(n), it is possible to check how the boarding and alighting information at time Z(n) changes. The attribute information and environmental information at time Z(n) are set as appropriate by the person operating the information processing device 50, and the control unit 51 can acquire the attribute information and environmental information at time Z(n) by inputting the set attribute information and environmental information via the operation unit 55.

[0079] Figure 15 shows the first example of attribute information. As shown in Figure 15, attribute information is information associated with each customer and includes, for example, information such as the customer's gender, age, usage area, whether or not they have a commuter pass, preferences, and residential area. The usage area represents the customer's living area and range of activity, and in the Kansai region, it includes the central Osaka area, the Kyoto / Shiga area, the Kobe / Hanshin area, the Nara / Wakayama area, and the Hokusetu / Kitakawachi area. Note that attribute information is not limited to the example in Figure 15.

[0080] Attribute information may include at least one of the following: customer gender, age, preferences, whether or not they have a commuter pass, and the area where they use transportation-related payment information.

[0081] Figure 16 shows a second example of attribute information. As shown in Figure 16, attribute information is information associated with each customer and may be an embedded vector. The control unit 51 can input character information (string) to the large-scale language model 100, obtain the embedded vector output by the large-scale language model 100, and use the obtained embedded vector as attribute information. The character information representing the attributes of customer ID=01001 can be set to STR001, and the corresponding embedded vector to V0001. The character information STR01001 is a qualitative expression, for example, "Male in his 30s. Resides in Osaka Prefecture. Uses Umeda Station frequently on weekdays for regular commutes."

[0082] In the case of attribute information as shown in Figure 15, the attribute information becomes a quantitative parameter (for example, age 40), and it is not possible to provide any parameters other than this parameter when training the learning model 59. On the other hand, in the case of embedding vectors as shown in Figure 16, the attribute information is represented qualitatively, so the learning model 59 can learn the attribute information qualitatively, and the need for retraining can be reduced even if the attribute information changes.

[0083] Attribute information may include character data representing customer attributes, which is input into a large-scale language model, and the embedding vectors output by the large-scale language model.

[0084] Figure 17 shows an example of environmental information. Environmental information includes, for example, the month, weather, weekday / holiday status, events, service information, and fare information. The month refers to each month from January to December. Weather includes, for example, sunny, cloudy, rainy, snowy, etc. Events include, for example, event details and event dates and times. Service information includes, for example, delay times and cancellation information for each route. Fare information includes, for example, information on fare increases or decreases for each route. Note that environmental information is not limited to the example in Figure 17.

[0085] Environmental information, like attribute information, may also be in the form of embedded vectors. In this case, the control unit 51 inputs character information (string) to the large-scale language model 100, obtains the embedded vector output by the large-scale language model 100, and uses the obtained embedded vector as environmental information.

[0086] Environmental information may include at least one of the following: month, weather, temperature, whether it is a weekday or public holiday, events, service information, and fare information.

[0087] Environmental information may include embedding vectors output by a large-scale language model, which is created by inputting character data representing the environmental information into the large-scale language model.

[0088] As described above, the control unit 51 can acquire time-series data of the customer's boarding and alighting information (transportation payment information), attribute information, and environmental information up to the first point in time, input the acquired time-series data into the learning model 59, and generate the customer's boarding and alighting information (transportation payment information) at the second point in time.

[0089] More specifically, the control unit 51 acquires time-series data of the customer's boarding and alighting information (transportation payment information), attribute information, and environmental information up to the first point in time, acquires the customer's attribute information and environmental information at the second point in time, inputs the acquired time-series data and the attribute information and environmental information at the second point in time into the learning model 59, and can generate the customer's boarding and alighting information (transportation payment information) at the second point in time.

[0090] By using attribute and environmental information at a second point in time as simulation conditions, it becomes possible to understand how boarding and alighting information at that second point in time changes when the attribute and environmental information at that second point in time is virtually altered and input. This allows for the simulation of customer behavior under the virtual attribute and environmental information, and the results can be used for macro-level demand forecasting and understanding customer behavior.

[0091] Figure 18 shows an example of customer behavior information and demand forecasting based on boarding and alighting information generated by the learning model 59. As shown in Figure 18, for convenience, we consider the following as hypothetical attribute and environmental information to be input: (1) changes in the area of ​​use, (2) changes in the presence or absence of a commuter pass, (3) changes in the weather, (4) changes in the operation information, and (5) changes in the fare information. Note that the hypothetical attribute and environmental information to be input may be at least two or more combinations of changes in the area of ​​use, changes in the presence or absence of a commuter pass, changes in the weather, changes in the operation information, and changes in the fare information.

[0092] If the service area is changed, a change in the boarding and alighting stations is observed in the generated boarding and alighting information. In this case, the behavioral information of each individual customer (customer behavioral information) can be interpreted, for example, as a change in the section of travel. Furthermore, demand forecasts based on changes in the boarding and alighting information of multiple customers can be interpreted, for example, as a change in the boarding and alighting distribution. The boarding and alighting distribution is represented by the number of passengers boarding and alighting at each station.

[0093] If the presence or absence of a commuter pass is changed, and changes in boarding and alighting times and days of the week are observed in the generated boarding and alighting information, then customer behavior information can be interpreted, for example, as a change in the frequency of use. Furthermore, demand forecasting can be interpreted, for example, as a change in revenue (sales) for each route.

[0094] If the weather changes, and a change in boarding and alighting times is observed in the generated boarding and alighting data, then customer behavior information can be interpreted, for example, as a change in the number of boarding and alighting trips. Similarly, demand forecasting can be interpreted, for example, as a change in the number of passengers boarding and alighting.

[0095] If operational information is changed, the generated boarding and alighting information may show changes in boarding and alighting stations and boarding and alighting times. In this case, customer behavior information can be interpreted, for example, as a change in the time of day of travel. Furthermore, demand forecasting can be interpreted, for example, as a change in customer guidance. Customer guidance refers to, for example, the information provided to customers so that they can use public transportation smoothly.

[0096] If fare information is changed, and a change in boarding and alighting times is observed in the generated boarding and alighting data, then customer behavior information can be interpreted, for example, as a change in boarding frequency. Furthermore, demand forecasting can be interpreted, for example, as a change in revenue (sales) for each route.

[0097] As described above, the control unit 51 can generate customer behavior information based on the generated boarding and alighting information (transportation-related payment information). This makes it possible to use this information for macro-level demand forecasting and understanding customer behavior.

[0098] Next, we will explain the comparison between the generated boarding and alighting information and the actual boarding and alighting information.

[0099] Figure 19 shows an example of comparing the generated boarding / alighting information with the actual boarding / alighting information. The generated boarding / alighting information is, for example, the boarding / alighting information at the second time point generated by the learning model 59 by inputting attribute information and environmental information at the second time point into the learning model 59. The actual boarding / alighting information is the actual boarding / alighting information at the second time point. As shown in Figure 19, let's assume that the boarding / alighting information (actual) was station S. In this case, if we let a be the probability that the boarding / alighting information generated by the learning model 59 is station S, then the error information between the generated boarding / alighting information and the actual boarding / alighting information can be expressed as a degree of mismatch, and the degree of mismatch can be calculated as a probability of (1-a).

[0100] In other words, the control unit 51 can calculate the degree of discrepancy between the generated boarding and alighting information and the actual boarding and alighting information at the second point in time.

[0101] By comparing the generated boarding and alighting data with the actual boarding and alighting data, it can be used to (1) measure the effects of policies due to changes in customer behavior, and (2) visualize macro-level anomaly scores for railway lines.

[0102] Measuring the effectiveness of measures based on changes in customer behavior can be done, for example, as follows: When events or campaigns not considered in the learning model 59 are implemented, it is expected that there will be discrepancies between the boarding and alighting information generated by the learning model 59 and the actual boarding and alighting information for customers who participated in these events. From this error information, it is possible to understand how effective the event or campaign was. For example, suppose an event was held at station S. By simulating the movement of event participants, the closer the probability a of visiting station S is to 0, the more it means that customers who do not normally use station S have been attracted, indicating a high probability that they visited station S specifically for the event, and thus contributing to the revitalization of the area.

[0103] Visualizing macro-level anomaly scores for a railway line is as follows: When something unusual occurs on a railway line, it is expected that the boarding and alighting information generated by the learning model 59 will differ from the actual boarding and alighting information for affected customers, resulting in an error. By visualizing this error information across the entire railway line, it is possible to detect any anomalies occurring somewhere within the company at an early stage.

[0104] Figure 20 shows an example of payment information generation by the learning model 59. The control unit 51 acquires time-series data of payment information, attribute information, and environmental information up to the first point in time for the customer, and inputs the acquired time-series data into the learning model 59. In the example in Figure 20, payment information, attribute information, and environmental information for time points Z1, Z2, ..., Z(n-1) are input into the learning model 59. Time point Z(n-1) corresponds to the first point in time (for example, the current time).

[0105] The control unit 51 acquires attribute information and environmental information at time Z(n), and inputs the acquired attribute information and environmental information into the learning model 59.

[0106] The learning model 59 is constructed to generate settlement information for a second time point, which is later than the first time point, when it receives time-series data of settlement information, attribute information, and environmental information up to the first time point, as well as attribute information and environmental information for time point Z(n). Therefore, as shown in Figure 20, the learning model 59 generates and outputs settlement information for time point Z(n). Time point Z(n) corresponds to the second time point (for example, in the future).

[0107] The control unit 51 acquires training data that includes time-series data of the customer's payment information (transportation-related payment information), attribute information, and environmental information up to the first point in time, as well as payment information for a second point in time after the first point in time. Based on the acquired training data, it can construct a learning model 59 that generates payment information for the second point in time when time-series data of the payment information up to the first point in time, attribute information, and environmental information are input.

[0108] More specifically, the control unit 51 acquires time-series data of the customer's payment information (transportation-related payment information), attribute information, and environmental information up to the first point in time, as well as training data including attribute information, environmental information, and boarding / alighting information at a second point in time after the first point in time. Based on the acquired training data, the control unit 51 can construct a learning model 59 that generates payment information at the second point in time when time-series data of the payment information, attribute information, and environmental information up to the first point in time, as well as attribute information and environmental information at the second point in time, are input.

[0109] The attribute information and environmental information at time Z(n) specify the conditions for generating settlement information at time Z(n) in the simulation. In other words, by virtually changing the attribute information and environmental information at time Z(n), it is possible to check how the settlement information at time Z(n) changes. The attribute information and environmental information at time Z(n) are set as appropriate by the person operating the information processing device 50, and the control unit 51 can acquire the attribute information and environmental information at time Z(n) by inputting the set attribute information and environmental information via the operation unit 55.

[0110] As described above, the control unit 51 can acquire time-series data of the customer's payment information (transportation-related payment information), attribute information, and environmental information up to the first point in time, input the acquired time-series data into the learning model 59, and generate the customer's payment information (transportation-related payment information) at the second point in time.

[0111] More specifically, the control unit 51 acquires time-series data of the customer's payment information (transportation-related payment information), attribute information, and environmental information up to the first point in time, acquires the customer's attribute information and environmental information at the second point in time, inputs the acquired time-series data and the attribute information and environmental information at the second point in time into the learning model 59, and can generate the customer's payment information (transportation-related payment information) at the second point in time.

[0112] Figure 21 shows an example of the procedure for predicting transportation payment information by the information processing device 50. The control unit 51 acquires time-series data of the customer's transportation payment information up to the first point in time (S11), and inputs the acquired time-series data into the learning model 58 (S12).

[0113] The control unit 51 predicts the customer's transportation payment information at a second point in time (S13) and determines whether or not time-series data exists (S14). If time-series data exists (YES in S14), the control unit 51 continues processing from step S11 onwards; if time-series data does not exist (NO in S14), it terminates the process.

[0114] Figure 22 shows an example of the procedure for generating transportation payment information by the information processing device 50. The control unit 51 acquires time-series data of transportation payment information, attribute information, and environmental information up to the first point in time for the customer (S21). The control unit 51 acquires attribute information and environmental information for the customer at the second point in time (S22), and inputs the acquired time-series data, attribute information, and environmental information into the learning model 59 (S23).

[0115] The control unit 51 generates the transportation payment information for the second time point (S24) and determines whether or not to change the simulation conditions (S25). Whether or not to change the simulation conditions can be determined, for example, based on the operation of the person in charge of operating the information processing device 50. If the simulation conditions are to be changed (YES in S25), the control unit 51 continues the processing from step S22 onward; if the simulation conditions are not to be changed (NO in S25), the processing ends.

[0116] (Note 1) The computer program obtains time-series data of the customer's transportation payment information up to a first point in time, and when the time-series data of the transportation payment information up to the first point in time is input, the computer is instructed to input the obtained time-series data into a first learning model that predicts the transportation payment information at a second point in time after the first point in time, thereby predicting the customer's transportation payment information at the second point in time.

[0117] (Note 2) In Note 1, the computer program causes the computer to perform the following processing: the transportation payment information includes station boarding and alighting information, the first learning model predicts the boarding and alighting information of multiple customers at the station at the second point in time, and the computer generates time-specific congestion information for each station based on the collected station boarding and alighting information.

[0118] (Note 3) In Note 1 or Note 2, the computer program causes the computer to perform the following processing: the transportation payment information includes station boarding and alighting information, the first learning model predicts multiple customers' station boarding and alighting information at the second time point, and the computer generates time-specific congestion information between stations based on the collected station boarding and alighting information.

[0119] (Note 4) The computer program instructs the computer to perform a process that outputs guidance information for avoiding congestion based on the time-specific congestion information generated in Note 2 or Note 3.

[0120] (Note 5) The computer program, in any one of Notes 1 to 4, further acquires time-series data of attribute information and environmental information of the customer up to the first point in time, and when time-series data of transportation payment information, attribute information and environmental information up to the first point in time are input, it inputs the acquired time-series data into a second learning model that generates transportation payment information for the second point in time, thereby causing the computer to execute a process to generate transportation payment information for the customer at the second point in time.

[0121] (Note 6) The computer program, as described in Note 5, obtains the attribute information and environmental information of the customer at the second point in time, inputs the time-series data up to the first point in time, and the attribute information and environmental information at the second point in time into the second learning model, and causes the computer to execute the process of generating the customer's transportation payment information at the second point in time.

[0122] (Note 7) The computer program causes the computer to perform the following processing as described in Note 5 or Note 6: inputting character information including at least one of the attribute information and environmental information into a large-scale language model, obtaining the embedding vector output by the large-scale language model, and inputting the obtained embedding vector into the second learning model.

[0123] (Note 8) The computer program causes the computer to perform a process in any one of Notes 5 to 7 to generate customer behavior information based on the generated transportation payment information.

[0124] (Note 9) In any one of Notes 5 to 8, the computer program causes the computer to perform a process in which the transportation payment information includes station boarding and alighting information, and calculates the degree of discrepancy between the generated boarding and alighting information and the actual boarding and alighting information at the second point in time.

[0125] (Note 10) The computer program, in any one of Notes 5 to 8, states that the aforementioned transportation payment information includes station boarding and alighting information.

[0126] (Note 11) In any one of Notes 2, 3, 9, and 10, the computer program includes the station, the time of boarding and alighting at the station, and whether it is boarding or alighting.

[0127] (Note 12) In any one of Notes 2, 3, 9, 10, and 11, the computer program includes the day of the week in which the boarding and alighting information takes place.

[0128] (Note 13) In the computer program, in Note 1 and any one of Notes 5 to 8, the aforementioned transportation payment information includes payment information for the purchase of goods.

[0129] (Note 14) In Note 13, the computer program states that the payment information includes the date and time of payment, the store where the payment was made, and the amount of payment.

[0130] (Note 15) The computer program, in any one of Notes 5 to 10, includes at least one of the following attribute information: the customer's last name, age, preferences, whether or not they have a commuter pass, and the area where the transportation payment information is used.

[0131] (Note 16) The computer program, in any one of Notes 5 to 10, includes at least one of the following in its environmental information: month, weather, temperature, whether it is a weekday or a public holiday, events, service information, and fare information.

[0132] (Note 17) The user behavior sequence prediction system based on payment data includes a control unit, which acquires time-series data of transportation payment information up to a first point in time for the customer, and when time-series data of transportation payment information up to the first point in time is input, the acquired time-series data is input into a learning model that predicts transportation payment information at a second point in time after the first point in time, thereby predicting the transportation payment information of the customer at the second point in time.

[0133] (Note 18) The method for predicting user behavior sequences based on payment data involves obtaining time-series data of transportation payment information up to a first point in time for a customer, and then inputting the obtained time-series data into a learning model that predicts transportation payment information at a second point in time after the first point in time, thereby predicting the transportation payment information of the customer at the second point in time.

[0134] (Note 19) The learning model construction method involves acquiring time-series data of transportation payment information up to the first point in time for the customer, and training data including transportation payment information at a second point in time after the first point in time. Based on the acquired training data, the learning model is constructed to predict the transportation payment information at a second point in time after the first point in time when time-series data of transportation payment information up to the first point in time is input.

[0135] (Note 20) The learning model construction method involves acquiring time-series data of the customer's transportation payment information, attribute information, and environmental information up to the first point in time, as well as training data including transportation payment information at a second point in time after the first point in time. Based on the acquired training data, the learning model is constructed such that when time-series data of the transportation payment information, attribute information, and environmental information up to the first point in time is input, it generates transportation payment information at a second point in time after the first point in time.

[0136] The matters described in each embodiment can be combined with each other. Furthermore, the independent and dependent claims described in the claims can be combined with each other in any combination, regardless of the form of reference. In addition, the claims use a form in which claims referencing two or more other claims (multi-claim form), but are not limited to this. A form in which multi-claims referencing at least one multi-claim (multi-multi-claim) may also be used. [Explanation of symbols]

[0137] 1. Communication Network 10 Display device 20 Transportation-related payment information database 30 Attribute information DB 40 Environmental information DB 50 Information Processing Devices 51 Control Unit 52 Communications Department 53 memory 54 Display section 55 Operation section 56 Memory section 57 Computer Programs 58, 59 Learning Models 100 Large-Scale Language Models

Claims

1. We obtain time-series data of the customer's transportation-related payment information up to the first point in time. When time-series data of transportation payment information up to the first point in time is input, the acquired time-series data is input into a first learning model that predicts transportation payment information at a second point in time after the first point in time, thereby predicting the transportation payment information of the customer at the second point in time. A computer program that instructs a computer to perform a process.

2. We acquire time-series data of the customer's transportation payment information, attribute information, and environmental information up to the first point in time. The attribute information and environmental information of the aforementioned customer at a second point in time, which is after the first point in time, When time-series data of transportation payment information, attribute information, and environmental information up to the first point in time, and attribute information and environmental information from a second point in time after the first point in time are input, the acquired time-series data of transportation payment information, attribute information, and environmental information up to the first point in time, and attribute information and environmental information from the second point in time are input into a second learning model that generates transportation payment information for the second point in time, thereby generating the customer's transportation payment information for the second point in time. A computer program that instructs a computer to perform a process.

3. Based on the generated transportation payment information, the customer's behavior information is generated. A computer program according to claim 2 that causes a computer to perform a process.

4. Equipped with a control unit, The control unit, We obtain time-series data of the customer's transportation-related payment information up to the first point in time. When time-series data of transportation payment information up to the first point in time is input, the acquired time-series data is input into a learning model that predicts transportation payment information at a second point in time after the first point in time, thereby predicting the customer's transportation payment information at the second point in time. A user behavior prediction system based on payment data.

5. We obtain time-series data of the customer's transportation-related payment information up to the first point in time. When time-series data of transportation payment information up to the first point in time is input, the acquired time-series data is input into a learning model that predicts transportation payment information at a second point in time after the first point in time, thereby predicting the customer's transportation payment information at the second point in time. A method for predicting user behavior sequences based on payment data.

6. The system acquires time-series data of the customer's transportation payment information up to a first point in time, and training data including transportation payment information from a second point in time after the first point in time. Based on the acquired training data, a learning model is constructed to predict transportation payment information at a second time point, after the first time point, when time-series data of transportation payment information up to the first time point is input. Methods for building learning models.

7. Time-series data of the customer's transportation payment information, attribute information, and environmental information up to the first point in time, attribute information and environmental information of the customer at a second point in time after the first point in time, and training data including the transportation payment information at the second point in time are acquired. Based on the acquired training data, a learning model is constructed to generate transportation payment information for the second time point when time-series data of transportation payment information, attribute information, and environmental information up to the first time point, as well as attribute information and environmental information for the second time point after the first time point, are input. Methods for building learning models.

Images