Item analysis device and item analysis method

Abstract

This item analysis device comprises: an input unit that receives an action sequence composed of records including at least three elements, namely, a user attribute, a time, and an action label representing an action content; an action model unit that receives at least one action sequence as an input, converts said at least one action sequence into a sequence of vectors of an embedding representation in consideration of a time order by using a prescribed function, and outputs a score for each action label in consideration of a pattern of the sequence of the vectors; and a training unit that uses the action sequence converted into the vectors as training data to estimate a parameter of the action model unit by self-supervised training using the difference between each action label score and correct answer data as an error function, and outputs an estimated model parameter.

Description

Item analysis device and item analysis method

[0001] The disclosed technology relates to an item analysis device and an item analysis method.

[0002] Companies are conducting marketing activities that utilize customer data in order to improve the value of customer experience and increase revenue.

[0003] Increasing the average customer spend is one of the key points for maximizing a company's profits. One widely used method for achieving this is cross-selling, which provides additional value to customers who have already purchased the company's products. However, analyzing the products to be cross-sold is difficult, and currently this is done based on simple purchasing history and the marketer's experience and intuition.

[0004] Association rule analysis has been known as a method for discovering combinations of products that are likely to be purchased. Association rule analysis is a method for analyzing patterns of products purchased together using indices based on customer purchase histories. Indices include, for example, the rate at which a set of products is purchased simultaneously (support), the probability that another product will be purchased after the purchase of one product (confidence), and the strength of the correlation between purchased products (lift value).

[0005] As a recommendation technology using association analysis, for example, Patent Document 1 discloses a technology relating to a recommendation rule generation device that extracts useful association rules from a large number of association rules and uses the extracted rules as recommendation rules.

[0006] Patent No. 6018541

[0007] However, association rule analysis has two problems. The first problem is that it cannot take into account the order of time series relationships. Product combinations are evaluated using fragmented data without considering time information. The second problem is that product combinations are characterized using indices such as support, confidence, and lift value, making it difficult to determine similarities and comparisons between products.

[0008] The disclosed technology has been made in consideration of the above points, and aims to provide an item analysis device and an item analysis method that enable analysis of trends in item combinations taking time information into account.

[0009] The disclosed item analysis device includes an input unit that accepts behavioral sequences composed of records including at least three items: a user attribute, a time, and a behavioral label that represents the activity content; a behavioral model unit that accepts one or more behavioral sequences as input, converts the input into a vector sequence of embedded representations that takes time order into account using a predetermined function, and outputs a score for each behavioral label by taking into account the pattern of the vector sequence; and a learning unit that uses the behavioral sequences converted into vectors as training data to estimate parameters of the behavioral model unit by self-supervised learning using the difference between each behavioral label score and ground truth data as an error function, and outputs the estimated model parameters.

[0010] The item analysis method disclosed herein is performed by a computer, which performs the following processes: accepts an action sequence consisting of records including at least three items: a user attribute, a time, and an action label representing the action content; accepts one or more action sequences as input; converts the action sequences into a series of vectors of embedded representations that take time order into account using a predetermined function; outputs a score for each action label taking into account the pattern of the series of vectors; uses the action sequences converted into vectors as training data; estimates parameters used in the process of calculating the score for each action label by self-supervised learning using the difference between each action label score and ground truth data as an error function; and outputs the estimated model parameters.

[0011] The disclosed technology makes it possible to analyze trends in item combinations by taking time information into consideration.

[0012] FIG. 1 is a diagram showing an overall outline of the processing of the item analysis device of this embodiment. FIG. 2 is a block diagram showing the hardware configuration of the item analysis device. FIG. 3 is a block diagram showing the configuration of the item analysis device of this embodiment. FIG. 4 shows an example of accepted behavioral sequences and behavioral sequence data processing. FIG. 5 is a diagram showing the learning flow. FIG. 6 is an image in which the behavior of model parameters is projected into two dimensions. FIG. 7 is an example of an output item combination. FIG. 8 is a flowchart showing the flow of item analysis processing by the item analysis device. FIG. 9 is an example of analysis using behavioral sequence data of customer groups from different data sets. FIG. 10 is an example of an output item combination obtained by combining multiple data sets.

[0013] An example of an embodiment of the disclosed technology will be described below with reference to the drawings. Note that the same or equivalent components and parts in each drawing are given the same reference numerals. Also, the dimensional proportions in the drawings are exaggerated for the convenience of explanation and may differ from the actual proportions.

[0014] First, to address the above-mentioned problem, the technology disclosed herein derives semantic vectors of products from product purchase histories through self-supervised learning that takes chronological order into consideration, and lists combinations of products that tend to be similar to each other by calculating similarity using these semantic vectors. Note that in this embodiment, a case will be described where a product is used as an example of an item.

[0015] This allows for the consideration of temporal information in the process of acquiring product semantic vectors through self-supervised learning. The norm of a product semantic vector tends to increase as the purchase frequency increases, and has properties similar to the support level in association rule analysis. The dot product and Euclidean distance between product semantic vectors represent similar relationships regarding product purchases, and have properties similar to reliability.

[0016] The technology of this embodiment makes it possible to find combinations of products that sell well, taking into account not only simple frequency but also chronological order. In addition, suggestions for proposing substitute products based on product similarity can be obtained. Furthermore, by using different data sets, this technology can acquire different semantic vectors for each data set. In other words, suggestions can be obtained, such as differences in combinations of products that sell well depending on customer groups (good customers, new customers, etc.) and times (seasons, time periods, etc.). Obtaining these suggestions is expected to be useful in designing more effective marketing measures.

[0017] FIG. 1 is a diagram showing an overall outline of the processing of the item analysis device of this embodiment. In (1), a behavioral sequence of time-series behavioral data for each customer is input. The behavioral sequence is processed and input into a model. In (2), the learning process involves learning the association between behaviors using self-supervised learning. In (3), the calculation process involves calculating the similarity of purchasing trends between products. In the calculation, the product purchase history is used as a product vector. In (4), the output is a combination of products with similar purchasing trends.

[0018] The configuration of this embodiment will be described below.

[0019] FIG. 2 is a block diagram showing the hardware configuration of the item analysis device 100.

[0020] 2 , the item analysis device 100 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0021] The CPU 11 is a central processing unit that executes various programs and controls each part. That is, the CPU 11 reads the programs from the ROM 12 or the storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs stored in the ROM 12 or the storage 14. In this embodiment, the programs are stored in the ROM 12 or the storage 14.

[0022] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured by a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores various programs including an operating system and various data.

[0023] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0024] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may be a touch panel type and function as the input unit 15.

[0025] The communication interface 17 is an interface for communicating with other devices such as terminals, etc. For this communication, for example, a wired communication standard such as Ethernet (registered trademark) or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi (registered trademark) is used.

[0026] Next, a description will be given of each functional configuration of the item analysis device 100. Fig. 3 is a block diagram showing the configuration of the item analysis device of this embodiment. Each functional configuration is realized by the CPU 11 reading a program stored in the ROM 12 or storage 14, expanding it in the RAM 13, and executing it.

[0027] As shown in Fig. 3, the item analysis device 100 includes a behavioral process analysis unit 102 and a vector distance calculation unit 104. The behavioral process analysis unit 102 corresponds to the learning phase, and the vector distance calculation unit 104 corresponds to the calculation phase, and these may be configured as separate devices. The behavioral process analysis unit 102 includes an input unit 110, a behavioral model unit 112, and a learning unit 114. The vector distance calculation unit 104 includes a vector similarity calculation unit 120, a counting unit 122, and a display unit 124.

[0028] The input unit 110 receives a behavioral sequence. The behavioral sequence is composed of a record including at least three items: a user attribute, a time, and a behavior label representing the behavior content. Furthermore, the input unit 110 processes the data by grouping the behaviors of an arbitrary period in the behavioral sequence into one token or a set of tokens.

[0029] Figure 4 shows an example of accepted behavioral sequences and data processing of the behavioral sequences. As shown in (1), data on the time series of behavioral sequences for each customer is accepted. As shown in (2), the behavioral content is summarized into tokens. The behavioral content can be in any format, such as a token or a set of tokens.

[0030] The behavioral model unit 112 receives one or more behavioral sequences as input and converts them into a sequence of vectors of embedded representation that takes into account the time order using a predetermined function. The behavioral model unit 112 outputs a score for each behavioral label, taking into account the pattern of the sequence of vectors. The predetermined function may be any function that can be used for embedded representation, such as an Embedding function.

[0031] The learning unit 114 uses the vector-converted behavioral sequences as training data to estimate the parameters of the behavioral model unit 112 through self-supervised learning using the difference between each behavioral label score and the correct data as an error function, and outputs the estimated model parameters.

[0032] FIG. 5 is a diagram showing the learning flow. In (3), embedding is performed on the input action sequence data. The action sequence is represented as a vector of the action embedding representation, and a location code based on the date and time is added. (a1) corresponds to [“Purchase”, “Product X”], and (a2) corresponds to [“Purchase”, “Product Y”, “Product Z”, “Payment”, “Code Payment”]. In (4), encoder learning is performed. Encoder learning corresponds to the self-supervised learning disclosed herein. For encoder learning, error calculation is performed using MASK learning. Weights are repeatedly adjusted so that the encoder output results match the input data. In (5), semantic vectors are extracted. The semantic vector is an example of an estimated model parameter. ←? Semantic vectors are expressed as n-dimensional vectors for each element of the action content. Behaviors that occur frequently have a larger weight and a larger norm, and actions with similar tendencies are output with similar orientations.

[0033] For a technique for interpreting a Transformer from a semantic vector, please refer to the following reference 1. [Reference 1] [2209.02535] Analyzing Transformers in Embedding Space (arxiv.org)

[0034] The vector similarity calculation unit 120 receives the model parameters output by the learning unit 114, and calculates vector similarity for each behavior label using the model parameters.

[0035] The counting unit 122 counts the combinations of activity labels using the calculated vector similarity and outputs the counting result. The combinations of activity labels are output as combinations of items.

[0036] The display unit 124 uses the aggregation results to display predetermined items based on the searched item names. The predetermined items are, for example, products that tend to sell together. In addition to products that tend to sell together, the predetermined items may also be products that tend not to sell together or a specific type of product that is specified as the search target.

[0037] Figure 6 shows an image of the behavior of model parameters projected onto two dimensions. Vector similarity can be calculated using any method, such as cosine similarity or Euclidean distance. By setting an arbitrary distance, vectors within that distance can be considered to have similar developmental tendencies.

[0038] FIG. 7 shows an example of an output combination of items. FIG. 7 shows a combination of products that tend to sell together, and is an example of a case where the user wants to know what products sell well together with toothbrushes. When aggregating similar products, a vector representing the purchase of a product is considered to be the vector of that product, and products that are considered to be close to each other (i.e., have similar purchasing trends) are aggregated using any method. Here, a marketer specifies and inputs the product name of the product they want to sell through a search. The display unit 124 uses the calculation results of the vector similarity shown in FIG. 6 above to output and display the specified product and combinations of products that can be sold together in a table format. The specified product name is an example of a searched item name in the present disclosure.

[0039] Next, the operation of the item analysis device 100 will be described. Figure 8 is a flowchart showing the flow of item analysis processing by the item analysis device 100. The item analysis processing is performed by the CPU 11 reading a program from the ROM 12 or storage 14, expanding it into the RAM 13, and executing it. Note that although this is described as a series of processing steps, the learning phase steps S100 to S106 and the calculation phase steps S108 to S112 may be separated and executed independently.

[0040] In step S100, the CPU 11 receives a behavior sequence as the input unit 110.

[0041] In step S102, the CPU 11, functioning as the behavior model unit 112, receives one or more behavior sequences as input and converts them into a sequence of vectors of embedded representations that take time order into consideration using a predetermined function.

[0042] In step S104, the CPU 11 functions as the behavior model unit 112 and outputs a score for each behavior label, taking into consideration the pattern of the vector series.

[0043] In step S106, the CPU 11, as the learning unit 114, uses the behavioral sequences converted into vectors as training data to estimate parameters used in the process of calculating a score for each behavioral label through self-supervised learning using the difference between each behavioral label score and the correct data as an error function, and outputs the estimated model parameters.

[0044] In step S108, the CPU 11 functions as the vector similarity calculation unit 120 to receive the model parameters output in the previous step and calculate vector similarity for each behavior label using the model parameters.

[0045] In step S110, the CPU 11 functions as the counting unit 122 to count combinations of behavior labels using the calculated vector similarity and output the counting result.

[0046] In step S112, the CPU 11 causes the display unit 124 to display predetermined items based on the searched item names using the counting results.

[0047] As described above, the item analysis device 100 of this embodiment makes it possible to analyze trends in item combinations while taking time information into consideration.

[0048] (Variation) Figure 9 shows an example of analysis using data on the behavioral sequences of customer groups from different data sets. (1) shows an image of the behavior of customer group A projected onto two dimensions, and (2) shows an image of the behavior of customer group B projected onto two dimensions. Using each semantic vector, the similarity between the vectors of each customer group is calculated, and combinations of products with similar purchasing tendencies are tallied using any method.

[0049] Fig. 10 shows an example of an item combination that is output by combining multiple data sets. In the example of Fig. 10, for a specified product name, products that can be sold together for each of customer group A and customer group B are output.

[0050] As described above, the behavioral model unit 112 processes a plurality of different data sets, and the learning unit 114 outputs model parameters. The vector similarity calculation unit 120 calculates vector similarities for each of the plurality of data sets using the model parameters for each of the plurality of data sets. The aggregation unit 122 outputs aggregation results for each of the plurality of data sets using the vector similarities calculated for each of the plurality of data sets. The display unit 124 displays predetermined items in the combination of the plurality of data sets based on the searched product names, using the aggregation results for each of the plurality of data sets.

[0051] In the above embodiment, the item analysis process executed by the CPU by reading the software (program) may be executed by various processors other than the CPU. Examples of such processors include programmable logic devices (PLDs) (such as field-programmable gate arrays (FPGAs)) whose circuit configuration can be changed after manufacture, graphics processing units (GPUs), and dedicated electrical circuits, such as application-specific integrated circuits (ASICs), which are processors with circuit configurations designed specifically for executing specific processes. Furthermore, the item analysis process may be executed by one of these processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). Furthermore, the hardware structure of these various processors is, more specifically, an electric circuit that combines circuit elements such as semiconductor elements.

[0052] In the above embodiment, the program is pre-stored (installed) in the storage 14, but the present invention is not limited to this. The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network.

[0053] The following additional notes are provided regarding the above-described embodiments.

[0054] (Supplementary Item 1) An item analysis device including: a memory; and at least one processor connected to the memory, wherein the processor is configured to: accept an action sequence consisting of records including at least three items: a user attribute, a time, and an action label representing an action content; accept one or more action sequences as input, convert the one or more action sequences into a series of vectors of embedded representation taking into account time order using a predetermined function, and output a score for each action label taking into account the pattern of the series of vectors; use the action sequences converted into vectors as training data, to estimate parameters used in the process of calculating the score for each action label by self-supervised learning using the difference between each action label score and ground truth data as an error function, and output the estimated model parameters.

[0055] (Supplementary Item 2) A non-transitory storage medium storing a program executable by a computer to perform item analysis processing, the non-transitory storage medium comprising: an action sequence consisting of records including at least three items: a user attribute, a time, and an action label representing the action content; one or more action sequences are received as input; the program converts the input into a vector sequence of embedded representations that takes into account the time order using a predetermined function; and outputs a score for each action label taking into account the pattern of the vector sequence; the program uses the action sequences converted into vectors as training data to estimate parameters used in the process of calculating the score for each action label by self-supervised learning using the difference between each action label score and ground truth data as an error function; and outputs the estimated model parameters.

Claims

1. An item analysis device comprising: an input unit that accepts an action sequence consisting of a record including at least three items: a user attribute, a time, and an action label representing the action content; a behavior model unit that accepts one or more action sequences as input, converts them into a vector sequence of embedded representations that take time order into account using a predetermined function, and outputs a score for each action label taking into account the pattern of the vector sequence; and a learning unit that uses the action sequences converted into vectors and the scores for each action label as training data, estimates parameters for the action model unit by self-supervised learning using the difference between each action label score and correct data as an error function, and outputs the estimated model parameters.

2. The item analysis device of claim 1, further comprising a vector similarity calculation unit, a counting unit, and a display unit, wherein the vector similarity calculation unit receives model parameters output by the learning unit and calculates vector similarities for each behavior label using the model parameters, the counting unit uses the calculated vector similarities to count combinations of behavior labels and output the counting results, and the display unit uses the counting results to display specified items based on the searched item names.

3. The item analysis device of claim 2, wherein the behavioral model unit processes a plurality of data sets having different behavioral sequence data and the learning unit outputs the model parameters, the vector similarity calculation unit calculates vector similarity for each of the plurality of data sets using the model parameters for each of the plurality of data sets, the aggregation unit outputs the aggregation results for each of the plurality of data sets using the vector similarity calculated for each of the plurality of data sets, and the display unit displays the specified item in the combination of the plurality of data sets based on the searched product name using the aggregation results for each of the plurality of data sets.

4. An item analysis method in which a computer performs the following processes: accepts an action sequence consisting of records containing at least three elements: user attributes, time, and action labels representing action content; accepts one or more action sequences as input, converts them into a series of vectors of embedded representations that take time order into account using a predetermined function, and outputs a score for each action label taking into account the pattern of the series of vectors; uses the action sequences converted into vectors as training data to estimate parameters used in the process of calculating the score for each action label using self-supervised learning that uses the difference between each action label score and correct data as an error function; and outputs the estimated model parameters.

Images