Recommendation device and recommendation method

Abstract

Not only do we recommend items that are appropriate for the user, but we also recommend items in a way that allows them to experience serendipity. [Solution] The recommendation system 1 includes a learning unit 10 and an input / output unit 20. The learning unit 10 inputs a user and an item history, inputs the user and partially missing item history into a model, encodes the user and partially missing item history into a latent space, learns a model so that the original item history can be reconstructed from the user and partially missing item history encoded in the latent space, learns a model so that recommended items can be predicted from the user and partially missing item history encoded in the latent space, and learns the model so as to minimize an objective function that represents the rarity of the recommended item from the perspective of all users and the unexpectedness of the recommended item from the perspective of the users and item history. The input / output unit 20 inputs the user and item history into the learned model, and outputs recommended items to be recommended to the user.

Description

[Technical field]

[0001] The present disclosure relates to a recommendation device and a recommendation method. [Background technology]

[0002] Recommendation is a technology that predicts each user's preferences and behavior using their personal history and attributes, and recommends information and items that match them. In order to improve the accuracy of predictions and the quality of recommendations, not only actions such as access and purchases and personal attributes are used, but also the time and location information of each action.

[0003] Known recommendation models incorporating time information are described in Non-Patent Documents 1-4. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Jiaxi Tang and Ke Wangy, “Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding,” In WSDM, 2018, pp. 565-573 [Non-Patent Document 2] Lu Yu, Chuxu Zhang, Shangsong Liang, and Xiangliang Zhang, “Multi-Order Attentive Ranking Model for Sequential Recommendation,” In AAAI, 2019, pp. 5709-5716. [Non-Patent Document 3] Jiacheng Li, Yujie Wang, and Julian J. McAuley, “Time Interval Aware Self-Attention for Sequential Recommendation,” In WSDM, 2020, pp. 322-330. [Non-Patent Document 4] Liwei Wu, Shuqing Li, Cho-Jui Hsieh, and James Sharpnack, “SSE-PT: Sequential Recommendation Via Personalized Transformer,” In RecSys, 2020, pp. 328-337 Summary of the Invention [Problem to be solved by the invention]

[0005] However, simply using history to suggest items had the problem of lacking new discoveries and unexpected surprises.

[0006] The present disclosure has been made in consideration of the above, and aims to not only appropriately recommend items that suit a user, but also to recommend items in a way that allows the user to experience serendipity. [Means for solving the problem]

[0007] A recommendation device according to one aspect of the present disclosure inputs a user and an item history, inputs the user and the partially missing item history into a model, encodes the user and the partially missing item history into a latent space, and Maximize the likelihood of the conditional probability of the actual item history given User and item history encoded in the latent space with some missing items Maximize the likelihood of the conditional probability of the actual next item given , so that the ratio of the number of users who recommended the recommended item to the total number of users is small and the distance in the latent space between the recommended item and each item in the user's item history is large. The system includes a learning unit that learns the model, and an input / output unit that inputs a user and an item history to the learned model and outputs recommended items to be recommended to the user. Effect of the Invention

[0008] According to the present disclosure, not only can items that suit a user be appropriately recommended, but items can also be recommended in a way that allows the user to experience serendipity. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration of a recommendation system. [Diagram 2] FIG. 2 is a diagram for explaining learning and prediction of user and item history. [Diagram 3] FIG. 3 is a diagram illustrating an example of the proposed model. [Figure 4] FIG. 4 is a diagram illustrating an example of the implementation of the proposed model. [Diagram 5] FIG. 5 is a flowchart showing an example of the flow of the learning process. [Figure 6] FIG. 6 is a flowchart showing an example of a process for recommending an item. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] [System Configuration] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0011] FIG. 1 is a diagram showing an example of the configuration of a recommendation system according to this embodiment. The recommendation system 1 learns a user's item history, accepts a user identifier from a user terminal 5, inputs the user identifier into a trained deep learning model, and returns recommended items to be recommended to the user. The items are, for example, products purchased by the user, content used by the user (books, music, videos, etc.), and services used by the user. The item history is a history of items purchased or used by the user, and includes time information such as the purchase time, the use time, or the purchase order.

[0012] The recommendation system 1 shown in Fig. 1 includes a learning unit 10, an input / output unit 20, a data storage unit 30, and a calculation result storage unit 40. Each unit included in the recommendation system 1 may be configured with at least one computer including an arithmetic processing unit, a storage device, etc., and the processing of each unit may be executed by a program. This program is stored in a storage device included in the recommendation system 1, and can be recorded on a non-transitory computer-readable recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or can be provided via a network.

[0013] The learning unit 10 embeds different data such as users and items in the same space and learns dynamic user preferences and item trends by focusing on the similarity between recommendations and natural language processing. In natural language processing, the next word to appear is predicted based on words that have appeared so far. Similarly, recommendations can predict the next item that a user is likely to purchase or use based on items that the user has previously purchased or used.

[0014] The learning unit 10 includes a pre-processing unit 11 and a calculation processing unit 12 .

[0015] The pre-processing unit 11 sets various parameters of the model, acquires learning data from the data storage unit 30, extracts user identifiers and item histories (including, for example, item identifiers and time information), and generates a dataset consisting of item histories for each user.

[0016] The calculation processing unit 12 inputs the user identifier and the item history into the model, and learns the parameters of the model so as to minimize the objective function. The calculation processing unit 12 masks some items in the item history, and learns the parameters of the model so as to correctly infer the masked items, thereby making it possible to predict more appropriate recommended items.

[0017] The calculation processing unit 12 maximizes serendipity by learning a model so that the recommended items are rare to all users and highly unexpected from the user's past item history. Serendipity is the chance encounter or discovery of something unexpected.

[0018] The input / output unit 20 inputs a user identifier and an item history from the user terminal 5, inputs them into the deep learning model, and returns recommended items. The input / output unit 20 may accept a user identifier from the user terminal 5 and obtain an item history corresponding to the user identifier from the data storage unit 30.

[0019] The data storage unit 30 stores data for learning. The data for learning includes, for example, information about a user, an item purchased or used by the user, and time information.

[0020] The calculation result storage unit 40 holds the parameters of the model learned by the learning unit 10.

[0021] [Proposed model] An overview of model learning in this embodiment will be described.

[0022] FIG. 2 shows an overview of learning and prediction of user and item history. In this embodiment, u,1:t-1 In addition to user u (user identifier), we use the user's unique influence to predict the recommended item i t To achieve this, we introduce a latent space z that maps the features of users and items to a common representation. u,1:t-1 Item history i with a random item removed from u,1:t-1 (~tilde over i) to display the original full item history i u,1:t-1 We train the latent space z so that we can correctly predict the recommended item i. t It is believed that by predicting this, more appropriate items can be recommended.

[0023] As shown in the following equation, recommendations are decomposed into learning and prediction of user preferences (item history).

[0024]

number

[0025] Here, L FPM (ζ) is the objective function for maximizing serendipity. FPM Details of (ζ) will be given later.

[0026] The configuration of the model of this embodiment is shown in Fig. 3. In this embodiment, a variational autoencoder (VAE) configuration is used, and a model including an encoder, a latent space z, a reconstruction decoder, and a prediction decoder is used.

[0027] The above equation can be transformed using the calculus of variations to obtain the following equation.

[0028]

number

[0029] The encoder records user u and the item history i with a randomly selected item removed. u,1:t-1 (~tilde on i) is encoded into the latent space z, and the item history of user u and item i with a random selection removed is u,1:t-1 (~tilde above i) to original full item history i u,1:t-1 The likelihood logP(i u,1:t-1 We train the model to maximize |z;ξ. We also train the model to maximize the item history i of user u with randomly selected items removed. u,1:t-1 (~tilde above i) to select the recommended item i t The likelihood logP(i t|z;ψ) and train the model to maximize it.

[0030] The objective function L for maximizing serendipity is given in the following equation: FPM An example of (ζ) is shown below.

[0031]

number

[0032] where U is the set of users, e i is the embedding representation of the item.

[0033] novelty(i) calculates the novelty (unusualness) of item i. It is calculated based on the number of users who have recommended item i. Items with high novelty are considered to be more likely to provide a fresh and interesting experience. unexpect(u,i) calculates the unexpectedness of user u and item i. The item history of user u excluding item i u,\i Calculate the distance between item i and item i, and u,\i An item that is highly unexpected is considered to have a high possibility of providing a new experience that the user has not experienced before.

[0034] [Model implementation example] Next, an example of model implementation will be described with reference to Fig. 4. In this embodiment, as shown in Fig. 4, a VAE and a Transformer are linked to add a latent factor layer to the Transformer decoder, and a model for the above task is learned.

[0035] In addition, a mask called AAM shown on the right side of the figure was introduced into the Transformer's attention mechanism, and the reconstruction decoder and predictive decoder were consolidated into a single Transformer decoder. Specifically, when training the reconstruction decoder, an AAM (the lower mask in the figure) that can refer to all tokens in the user and item history is used, and when training the predictive decoder, an AAM (the upper mask in the figure) that can only refer to tokens in past item history among the user and item history is used. In other words, the upper mask controls access to tokens of other items based on the context of the item history.

[0036] The latent factor layer is expressed by the following equation:

[0037]

number

[0038] Here, h L,t is the t-th output of the L-th layer encoder. The function F is a function that reflects the latent space z in the output of the Transformer. Any of the transformation formulas can be arbitrarily selected. W IZ , and h L is a parameter that is updated during model training. g z ,W z , and b z are weights for the latent space and are parameters that are updated during model training.

[0039] During training, we input the user and the item history with randomly selected items removed into the proposed model, and calculate the objective function L MUM ,L MIM The parameters are trained to minimize

[0040]

number

[0041] where B is the batch and e a ,e p ,en are the embedding representations of the masked ID, the inferred ID, and the distinct ID, respectively. p is at least ε e n than e a Learn to get closer to.

[0042] User and item history are input into the proposed model, and parameters are trained to correctly predict recommended items.

[0043] In addition, the objective function L for maximizing the aforementioned serendipity FPM The parameters are trained to minimize (ζ).

[0044] Putting the above tasks together, model training is performed by minimizing the following objective function L:

[0045]

number

[0046] Here, e is the Napier number, t (bar on top) is the time elapsed until the present, and λ is the decay constant. TDCE is the discounted sum of rewards.

[0047] During inference, the user and item history are input to the proposed model to obtain inference items following the item history.

[0048] [Operation] Next, an example of the flow of the learning process will be described with reference to the flowchart in FIG.

[0049] In step S11, the learning unit 10 reads out users and item histories for learning from the data storage unit 30, randomly drops some of the item histories, and inputs them into the model.

[0050] In step S12, the learning unit 10 updates the parameters of the model so as to minimize the objective function shown above, and stores the updated parameters in the calculation result storage unit 40.

[0051] Next, an example of a process for recommending an item will be described with reference to the flowchart of FIG.

[0052] In step S21, the input / output unit 20 inputs a user identifier from the user terminal 5, and acquires an item history corresponding to the user identifier from the data storage unit 30. The input / output unit 20 may input the item history from the user terminal 5.

[0053] In step S22, the input / output unit 20 inputs the user identifier and the item history to the model, obtains recommended items from the model, and returns the recommended items to the user terminal 5.

[0054] As described above, the recommendation system 1 of this embodiment includes a learning unit 10 and an input / output unit 20. The learning unit 10 inputs a user and an item history, inputs the user and a partially missing item history into a model, encodes the user and a partially missing item history into a latent space, learns a model so that the original item history can be reconstructed from the user and a partially missing item history encoded in the latent space, learns a model so that recommended items can be predicted from the user and a partially missing item history encoded in the latent space, and learns a model so that an objective function that represents the rarity of the recommended item from the users as a whole and the unexpectedness of the recommended item from the users and the item history is minimized. The input / output unit 20 inputs the user and the item history into the learned model, and outputs recommended items to be recommended to the user. According to this embodiment, by learning the latent space so that the original item history can be reconstructed from the partially missing item history, recommended items can be predicted more accurately. In addition, by training the model to minimize an objective function that represents the rarity of the recommended item from the perspective of all users and the unexpectedness of the recommended item from the perspective of users and item history, it is possible to recommend items that allow users to experience serendipity. [Explanation of symbols]

[0055] 1 Recommendation Systems 10. Learning Department 11 Pre-processing section 12 Calculation processing unit 20 Input / output section 30 Data Storage Section 40 Calculation result storage section 5. User terminal

Claims

1. Input the user and item history, input the user and partially missing item history into the model, and encode the user and partially missing item history into a latent space. a learning unit that maximizes the likelihood of a conditional probability of an actual item history given the user and a partially missing item history encoded in the latent space, maximizes the likelihood of a conditional probability of an actual next item given the user and a partially missing item history encoded in the latent space, and learns the model so that the ratio of the number of users who recommended a recommended item to the total number of users is small and the distance in the latent space between the recommended item and each item in the user's item history is large; An input / output unit inputs a user and an item history into the trained model and outputs recommended items to be recommended to the user. Recommendation device.

2. The recommendation device according to claim 1 , The model includes an encoder, a reconstruction decoder, and a prediction decoder, the encoder encoding user and item histories into a latent space, the reconstruction decoder reconstructing original items from the user and partially missing item histories encoded in the latent space, and the prediction decoder predicting recommended items from the user and partially missing item histories encoded in the latent space. Recommendation device.

3. The recommendation device according to claim 1 or 2, The model is a Transformer with a latent factor layer added, Input the user and the tokens of the missing item history into the Transformer, and train the model so that the original item history can be reconstructed and the recommended items can be predicted. In the attention mechanism of the Transformer, when learning to reconstruct item history, all input tokens are referenced, and when learning to predict recommended items, access to other tokens is controlled based on the context of the item history. Recommendation device.

4. A computer comprising: Input the user and item history, input the user and partially missing item history into the model, and encode the user and partially missing item history into a latent space. maximizing the likelihood of the conditional probability of the actual item history given the user and the partially missing item history encoded in the latent space; maximizing the likelihood of the conditional probability of the actual next item given the user and the partially missing item history encoded in the latent space; training the model so that the ratio of the number of users who recommended the recommended item to the total number of users is small and the distance in the latent space between the recommended item and each item in the user's item history is large; Input the user and item history into the trained model, and output recommended items to be recommended to the user. How to recommend.

5. The recommendation method according to claim 4, The model includes an encoder, a reconstruction decoder, and a prediction decoder, the encoder encoding user and item histories into a latent space, the reconstruction decoder reconstructing original items from the user and partially missing item histories encoded in the latent space, and the prediction decoder predicting recommended items from the user and partially missing item histories encoded in the latent space. How to recommend.

Images