Information processing device, information processing method, and program

Abstract

To be able to recommend more appropriate products to users based on their individual needs. [Solution] An information processing device is provided, comprising: an acquisition unit that acquires input data indicating the characteristics of an item; an item space learning unit that converts the vectorized input data into a Box, which is an N-dimensional (N is an integer of 2 or more) rectangular region; and a personal space learning unit that adjusts the model of the Box based on the user data of a specific user.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] Patent Document 1 discloses a method for training a recommendation model. In Patent Document 1, the collected information is formed into triplet data that reflects the relationships among entities, attribute values, and attribute relationships. The training objective is that the generated feature vector reflects the mutual relationship of each element in the triplet data, and the vector conversion model is subjected to primary training to obtain a primary vector conversion model. The feature vector corresponding to the attribute value is represented as a box vector called an attribute box vector, and the feature vector corresponding to the entity is represented as a point vector called an entity point vector. The recommendation model obtained by training in this processing process has a clearer and more distinct expression of the mutual relationship between each element in the triplet data and can match the characteristics of different elements. Furthermore, the recommendation model can be further trained to obtain a tertiary vector conversion model that can accurately represent the user's interest feature vector, and this model can make accurate recommendations to the user.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the prior art, for example, there is room for improvement in the process of making recommendations to users.

[0005] The purpose of this disclosure is to provide technology that can recommend more appropriate products to users based on their individual needs. [Means for solving the problem]

[0006] In a first aspect of this disclosure, an information processing device is provided, comprising: an acquisition unit that acquires input data indicating the characteristics of an item; an item space learning unit that converts the vectorized input data into a Box, which is an N-dimensional (where N is an integer of 2 or more) rectangular region; and a personal space learning unit that adjusts the model of the Box based on the user data of a specific user.

[0007] Furthermore, a second aspect of this disclosure provides an information processing method that acquires input data indicating the characteristics of an item, converts the vectorized input data into a Box which is an N-dimensional (where N is an integer of 2 or more) rectangular region, and adjusts the model of the Box based on the user data of a specific user.

[0008] Furthermore, a third aspect of this disclosure provides a program that causes a computer to perform a process of acquiring input data that describes the characteristics of an item, converting the vectorized input data into a Box which is an N-dimensional (where N is an integer of 2 or more) rectangular region, and adjusting the model of the Box based on the user data of a specific user. [Effects of the Invention]

[0009] From one perspective, it allows for the recommendation of more appropriate products to users based on their individual needs. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of the configuration of an information processing device according to the embodiment. [Figure 2] A flowchart showing an example of processing performed by the information processing apparatus according to the present invention. [Figure 3] This figure shows an example of a Box for a specific user and a Box for each item according to the embodiment. [Figure 4]This figure shows an example of a specific user's Box and each person's Box according to the embodiment. [Figure 5] This figure shows an example of the hardware configuration of the information processing device according to the embodiment. [Modes for carrying out the invention]

[0011] The principles of this disclosure will be described with reference to several exemplary embodiments. These embodiments are described for illustrative purposes only and should be understood as helping those skilled in the art to understand and implement this disclosure without implying any limitation on the scope of this disclosure. The disclosures described herein may be implemented in various ways other than those described below.

[0012] In the following description and claims, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those generally understood by those skilled in the art to which this disclosure belongs. Embodiments of this disclosure will be described below with reference to the drawings.

[0013] <Structure> Referring to Figure 1, the configuration of the information processing device 10 according to the embodiment will be described. Figure 1 is a diagram showing an example of the configuration of the information processing device 10 according to the embodiment. In the example of Figure 1, the information processing device 10 has an acquisition unit 11, a vectorization processing unit 12, an item space learning unit 13, a personal space learning unit 14, and a determination unit 15. Each of these units may be realized through the cooperation of one or more programs installed in the information processing device 10 and hardware such as the processor and memory of the information processing device 10.

[0014] The acquisition unit 11 acquires input data that indicates the characteristics of the item. The vectorization processing unit 12 converts the input data acquired by the acquisition unit 11 into a vector.

[0015] The item space learning unit 13 converts the input data vectorized by the vectorization processing unit 12 into a Box, which is an N (N is an integer of 2 or more) - dimensional rectangular area. The models of the Boxes generated by the personal space learning unit 14 and the item space learning unit 13 are adjusted based on the user data of a specific user. The determination unit 15 determines the items to recommend to the user using the model of the Box generated by the personal space learning unit 14.

[0016] <Processing> Next, referring to FIGS. 2 to 4, an example of the processing of the information processing apparatus 10 according to the embodiment will be described. FIG. 2 is a flowchart showing an example of the processing of the information processing apparatus 10 according to the embodiment. FIG. 3 is a diagram showing an example of the Box of a specific user and the Boxes of each item according to the embodiment. FIG. 4 is a diagram showing an example of the Box of a specific user and the Boxes of each person in charge according to the embodiment. The processing in FIG. 2 may be executed, for example, in response to a specific operation by the administrator of the information processing apparatus 10. In addition, each processing in FIG. 2 may be executed in a different order as long as there is no contradiction.

[0017] In step S101, the acquisition unit 11 acquires input data indicating the characteristics of each arbitrary item. The input data may be, for example, text data. Also, the input data may be, for example, image data. Examples of items may include, for example, products sold at stores or EC (Electronic Commerce) sites, and persons (for example, users, persons in charge), etc.

[0018] Subsequently, the vectorization processing unit 12 vectorizes the input data using a learned model (step S102). Here, the vectorization processing unit 12 may perform pre - processing or the like on the input data acquired by the acquisition unit 11 and then input it to the learned model.

[0019] The learned model only needs to be a model capable of vectorizing input data. When the input data is text data, the learned model may be, for example, a natural language processing model using deep learning. In this case, the learned model may be, for example, a BERT (Bidirectional Encoder Representations from Transformers) model or a Word2Vec model. Also, when the input data is text data and image data, the learned model may be, for example, a LLaVa (Large Language and Vision Assistant) model.

[0020] Subsequently, the item space learning unit 13 converts (Box embedding) the input data vectorized by the vectorization processing unit 12 into a Box, which is an N-dimensional rectangular (including square) region, using a learned model that converts a vector into a Box (step S103). Through Box embedding, the object included in the input data is embedded in the latent space in the form of a Box, so that an embedded representation of the region (Box) reflecting the ambiguity of the object and the hierarchical relationship of upper and lower levels can be obtained. Here, the item space learning unit 13 may convert the input vector into a Box using, for example, unsupervised learning or supervised learning based on teacher data indicating the relationship between each Box for each vector. Note that in vector representation, data can only be represented as mere "points". By using a Box as the data representation form, the semantic and conceptual hierarchical relationship between data can be represented. Through Box embedding, the Box in the N-dimensional space may be represented by a pair of a "starting point" (for example, the smallest coordinate) and an "ending point" (for example, the largest coordinate).

[0021] Next, the personal space learning unit 14 adjusts (personalizes) the Box embedding space (model) generated by the item space learning unit 13 for each user based on user data (step S104). This adjusts the acquired embedding representations using user data to create an embedding representation space that reflects the user's perception. User data may include, for example, the user's behavioral history (e.g., purchase history, website browsing history, visit history to tourist spots, etc.), attribute information, and survey results. Here, the personal space learning unit 14 may adjust the model for the individual using a specific loss function and a classifier appropriate to the personalization task.

[0022] Next, the decision unit 15 uses the spatial model of the Box embedding generated by the personal space learning unit 14 to determine the items (products, services, staff, etc.) to recommend to the user (step S105). This makes it possible, for example, to estimate the user's preferences using information in the latent space that reflects how each user perceives the subject.

[0023] According to this disclosure, the hierarchical relationships of concepts are learned. Therefore, by embedding subjective words (e.g., pretty, easy to understand) as metadata, the reasons for recommending an item can be expressed linguistically. Furthermore, since the only data that needs to be collected in advance is user data and single item information (e.g., at least one of the description and image), the cost of data collection can be reduced. In addition, since all objects are treated in the form of boxes after conversion to vectors, there is no need to consider an index that expresses the similarity between boxes and points. Therefore, the impact on performance when optimizing the embedding space is reduced.

[0024] Furthermore, since the user's range of interests is determined mechanically through model learning based on user behavior (conversions, etc.), it offers advantages in terms of performance and effort. More specifically, the learned range of interests covers not only the user's explicit preferences but also their latent preferences.

[0025] (An example of training a purchase estimation model using a user's purchase history) The information processing device 10 may, for example, learn a purchase estimation model using the user's purchase history. However, users have different judgments (perceptions, ideas, and values) regarding whether a particular product is similar to other products. Therefore, in order to recommend products that will satisfy users, a purchase estimation model that reflects each user's perception of products is necessary.

[0026] In step S101, the acquisition unit 11 acquires input data that indicates the characteristics of the product (for example, the product description). Then, in step S103, the item space learning unit 13 generates a Vec2Box model (a model that converts vectors to boxes) based on the input data. The Vec2Box model calculates boxes for the input vectors (points). The item space learning unit 13 may, for example, take vector data related to products purchased by any (for example, all) users (for example, descriptions vectorized using a BERT model, etc.) and generate a box embedding model for those products.

[0027] Then, in step S104, the personal space learning unit 14 may retrain the Box model based, for example, on the classification results of a classification task for each item based on the logical OR of each Box for each item according to the user data. In this case, the personal space learning unit 14 may retrain the Box model to minimize a value based on the ratio of the logical OR of each Box to the accuracy of the classification task.

[0028] More specifically, the personal space learning unit 14 may, for example, obtain each Box in the Box embedding model for each item included in the purchase history of a specific user, and determine the logical OR of each Box as the initial purchase range for that user. The personal space learning unit 14 may then have a classifier solve a binary classification task for each product. In this case, the binary classification task may classify an item as "buy" if its Box overlaps with the user's purchase range by a certain percentage or more, and as "do not buy" if the overlap does not exceed a certain percentage. The personal space learning unit 14 may then adjust (retrain) the Vec2Box model for a specific user by backpropagating the results of the calculation using the classification results and a specific loss function.

[0029] The personal space learning unit 14 may use the following equation (1) as a specific loss function. Here, Vol is the volume of the Box of the predicted purchase range (preference region) (for example, the sum of the logical ORs of each Box in each dimension of N dimensions). Acc is the accuracy of the binary classification task. L CE is an arbitrary loss function or penalty term. Note that L CE It is not required. L V = Vol / Acc+L CE ...(1)

[0030] The personal space learning unit 14 uses, for example, the ratio of the volume of the box to the accuracy of the binary classification task to determine L V You can retrain the Vec2Box model to minimize the value of . This allows you to adjust (e.g., optimize) the purchase range to be relatively compact and to efficiently include only the items you will buy.

[0031] Then, in step S105, the determination unit 15 calculates the match rate between the Box of a specific user and the Box of each item. Figure 3 shows the Box 211 of a specific user, the Box 321 of the first item, the Box 322 of the second item, and the Box 323 of the third item according to this embodiment.

[0032] Here, the decision unit 15 may, for example, determine a higher match rate for an item if the overlap area between the specific user's Box and the item's Box is large. Alternatively, for example, the decision unit 15 may determine a higher match rate for an item if the similarity (e.g., cosine similarity, the reciprocal of the Euclidean distance, etc.) between the centroid (center) of the specific user's Box and the centroid of the item's Box is high. The decision unit 15 may, for example, recommend the item with the highest match rate to the specific user.

[0033] (Example of training a model for assigning personnel) The information processing device 10 may, for example, learn a model for assigning staff to sales representatives in a dealership. In situations such as car dealerships, where each user is assigned a dedicated staff member (salesperson) to conduct business negotiations and provide services, the compatibility between the user and the staff member is important from the perspective of improving sales and customer satisfaction for the dealership.

[0034] In step S101, the acquisition unit 11 acquires as input data the descriptions of items (products, etc.) that are of interest or preference to any user, and the descriptions of items that a particular person is good at or specializes in. Then, in step S103, the item space learning unit 13 generates a Vec2Box model based on the input data.

[0035] Then, in step S104, the personal space learning unit 14 may retrain the Box model based, for example, on the classification results of a classification task for each item based on the logical OR of each Box for each item according to the user data. In this case, the personal space learning unit 14 may retrain the Box model to minimize a value based on the ratio of the logical OR of each Box to the accuracy of the classification task.

[0036] More specifically, the personal space learning unit 14 may obtain each box in the box embedding model for each item that a specific user is interested in or has a preference for, and determine the logical OR of each box as the range of interest for that specific user. Note that each item that a specific user is interested in or has a preference for may be obtained, for example, through a user survey.

[0037] The personal space learning unit 14 may also have each person in charge solve a binary classification task using a classifier. In this case, the binary classification task may classify a box as a "match" if it overlaps with the user's range of interest by a certain percentage or more, and as a "does not match" if it does not overlap by a certain percentage or more. The personal space learning unit 14 may then adjust (retrain) the Vec2Box model for a specific user by backpropagating the classification result and the result of calculation using a specific loss function such as equation (1) above into the Vec2Box model. Figure 4 shows the Box 411 of a specific user, the Box 421 of the first person in charge, and the Box 422 of the second person in charge according to the embodiment.

[0038] Then, in step S105, the decision unit 15 calculates the match rate between the specific user's Box and each person in charge's Box. Here, the decision unit 15 may, for example, determine a higher match rate for a person in charge if the overlap area between the specific user's Box and the person in charge's Box is large. Alternatively, the decision unit 15 may, for example, determine a higher match rate for a person in charge if the similarity (e.g., cosine similarity, the reciprocal of the Euclidean distance, etc.) between the centroid (center) of the specific user's Box and the centroid of the person in charge's Box is high. The decision unit 15 may, for example, assign the person in charge with the highest match rate to the specific user.

[0039] (Examples of training various recommendation models) The recommendations described in this disclosure are not limited to the examples mentioned above. It is important that various recommendation systems adequately reflect user preferences in their recommendations. To achieve this, accurate preference estimation is necessary. Therefore, by using the learning method described in this disclosure to estimate and optimize user preference domains, and utilizing the results in the recommendation system, recommendation performance can be improved.

[0040] The information processing device 10 may learn a model for recommending, for example, tourist spots, experiences while out and about, and experiences with other people. In this case, the acquisition unit 11 may acquire, for example, descriptions of items such as tourist spots, experiences while out and about, and experiences with other people that are of interest or preference to any given user as input data. The personal space learning unit 14 may then estimate the interest characteristic areas of a specific user based, for example, the user's visit history, web page browsing history, etc.

[0041] <Hardware Configuration> Figure 5 shows an example of the hardware configuration of each information processing device 10 according to the embodiment. In the example in Figure 5, the information processing device 10 (computer 100) includes a processor 101, memory 102, and a communication interface 103. These parts may be connected by a bus or the like. The memory 102 stores at least a portion of the program 104. The communication interface 103 includes an interface necessary for communication with other network elements.

[0042] When program 104 is executed in cooperation with the processor 101 and memory 102, etc., the computer 100 performs at least some of the processing of embodiments of this disclosure. Memory 102 may be of any type. Memory 102 may, in non-limiting examples, be a non-temporary computer-readable storage medium. Memory 102 may also be implemented using any suitable data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory and removable memory. Although only one memory 102 is shown for computer 100, computer 100 may have several physically different memory modules. Processor 101 may be of any type. Processor 101 may include one or more general-purpose computers, dedicated computers, microprocessors, digital signal processors (DSPs), and, in non-limiting examples, processors based on multicore processor architectures. Computer 100 may have multiple processors, such as application-specific integrated circuit chips that are time-dependent to a clock that synchronizes the main processor.

[0043] Embodiments of the present disclosure may be implemented in hardware or in dedicated circuitry, software, logic, or any combination thereof. Some embodiments may be implemented in hardware, while others may be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device.

[0044] This disclosure also provides at least one computer program product tangibly stored on a non-temporary computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions contained in a program module, and is executed on a device on a target real or virtual processor to perform the processes or methods of this disclosure. The program module includes routines, programs, libraries, objects, classes, components, data structures, etc., that perform a specific task or implement a specific abstract data type. The functionality of the program module may be combined or divided among the program module as desired in various embodiments. The machine-executable instructions of the program module can be executed on a local or distributed device. On a distributed device, the program module can reside on both local and remote storage media.

[0045] Programs can be stored and supplied to a computer using various types of non-temporary computer-readable media. Non-temporary computer-readable media include various types of tangible recording media. Examples of non-temporary computer-readable media include magnetic recording media, magneto-optical recording media, optical disc media, and semiconductor memory. Magnetic recording media include, for example, flexible disks, magnetic tapes, and hard disk drives. Magneto-optical recording media include, for example, magneto-optical disks. Optical disc media include, for example, Blu-ray discs, CD (Compact Disc)-ROM (Read Only Memory), CD-R (Recordable), and CD-RW (ReWritable). Semiconductor memory includes, for example, solid-state drives, mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (random access memory). Programs may also be supplied to a computer using various types of temporary computer-readable media. Examples of temporary computer-readable media include electrical signals, optical signals, and electromagnetic waves. Temporary computer-readable media can supply programs to a computer via wired communication channels such as electric wires and optical fibers, or via wireless communication channels.

[0046] <Other> Services on digital marketplaces offer systems that support user decision-making by providing recommendations, such as product recommendations and person matching. In these systems, the interpretability of the recommendation rationale (the degree to which humans can understand the machine learning algorithm) is crucial. Therefore, methods to improve interpretability based on points obtained through collaborative filtering or embedding are being considered.

[0047] Another approach is to embed elements as regions rather than points (box embedding), which allows for the representation of both higher-level and lower-level concepts of the target. Box embedding can embed the hierarchical relationships of concepts, allowing for the identification of higher-level and lower-level items of a target item through set operations. However, if attributes and features are embedded as boxes and items are embedded as "points," the semantic hierarchy between the boxes and points cannot be evaluated. Therefore, because the hierarchical relationships of items are unclear, it becomes difficult to recommend items with interpretability.

[0048] Furthermore, when using Box embedding to perform knowledge graph-based embedding, the amount of data that needs to be collected increases significantly. This is because, in addition to each user's behavioral data (e.g., purchase / non-purchase), data in the form of triplets such as (item, Tag, Relation) is required. Therefore, preparing the data to be used for training is relatively complicated.

[0049] Furthermore, when using Box embedding, if points and boxes are mapped to separate spaces, performance may degrade when embedding them into a common space. This is because, for example, the target of optimization differs for each pattern, such as product-to-product similarity (IRI), tag-to-tag relationship (TRT), and product-to-tag association (IRT), but it is unclear where in the space the optimization will converge when both are optimized simultaneously. As a result, the performance of the learned space is unstable. On the other hand, according to this disclosure, it is possible to recommend more appropriate items to users depending on their needs.

[0050] <Variation> The information processing device 10 may be a device contained in a single enclosure, but the information processing device 10 of this disclosure is not limited to this. Each part of the information processing device 10 may be implemented by cloud computing, which is composed of, for example, one or more computers.

[0051] It should be noted that the present invention is not limited to the embodiments described above, and can be modified as appropriate without departing from the spirit of the invention. [Explanation of Symbols]

[0052] 10 Information Processing Devices 11 Acquisition Department 12 Vectorization Processing Unit 13 Item Spatial Learning Department 14. Personal Space Learning Department 15. Decision Section

Claims

1. An acquisition unit that acquires input data indicating the characteristics of an item, An item space learning unit that converts the vectorized input data into a Box, which is a rectangular region of N dimensions (where N is an integer of 2 or more), An information processing device having a personal space learning unit that adjusts the Box model based on user data of a specific user.

2. The aforementioned personal space learning unit is Based on the classification results from a classification task for each item, which is based on the logical OR of each Box for each item according to the user data, the model of the Box is retrained. The information processing apparatus according to claim 1.

3. The personal space learning unit retrains the model of the Boxes to minimize a value based on the ratio of the logical OR of each Box to the accuracy rate of the classification task. The information processing apparatus according to claim 2.

4. Obtain input data that describes the characteristics of the item, The vectorized input data is converted into a Box, which is a rectangular region of N dimensions (where N is an integer greater than or equal to 2). The aforementioned Box model is adjusted based on the user data of a specific user. Information processing methods.

5. Obtain input data that describes the characteristics of the item, The vectorized input data is converted into a Box, which is a rectangular region of N dimensions (where N is an integer greater than or equal to 2). The aforementioned Box model is adjusted based on the user data of a specific user. A program that instructs a computer to perform a process.

Images