A deep model sequence recommendation method and system based on user representation

Abstract

The present disclosure provides a method and system for deep model sequence recommendation based on user representation, the method comprising the following steps: respectively obtaining a long-term sequence of the user's overall preference and a short-term sequence of the user's current dynamic preference; obtaining a recommendation result through a deep sequence recommendation model MLUR; wherein the deep sequence recommendation model comprises a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module. Through a gating module based on a multi-layer perceptron MLP, the contribution ratio of long-term and short-term representations is determined by considering the relationship between hot-selling items, recently interacted items and long-term preferences. This module balances long-term and short-term representations by taking into account the information of hot-selling items, so that the dynamic nature of user intentions can be processed simultaneously.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data mining recommendation, and in particular to a method and system for recommending deep model sequences based on user representation. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Information overload refers to the fact that users have limited knowledge and cognitive abilities. When faced with a large amount of complex network information, they cannot quickly and accurately find the information they need, or even cannot understand and use the information. Therefore, in the face of massive amounts of data, how to extract data that is useful to users becomes the focus. The emergence of recommendation systems is to solve the problem of information overload, help consumers find the products they need in the vast sea of ​​products, and provide consumers with a personalized shopping experience. Recommendation systems are increasingly favored by users and are also attracting the attention of more and more scholars and e-commerce websites. In particular, the emerging topic of sequential recommendation systems has attracted more and more attention in recent years, because sequential recommendation systems more accurately describe the characteristics of consumer users, thereby making more accurate, personalized and dynamic recommendations. Although the current sequential recommendation system has achieved certain results, due to the uncertainty of user shopping behavior, the interaction between users and products contains some noise and some irrelevant interactions that will interfere with the prediction of the next step, and further improvement is needed.

[0004] The Sequential Recommender System mainly models the sequential dependency of user-item interactions in a sequence to recommend items that may be of interest to users. From a technical perspective, the various methods of the recommendation system can be divided into three categories: the first category is the traditional sequence model, which includes sequence pattern mining and Markov chain. It is a relatively simple and intuitive solution. This method uses the natural advantages of the sequence recommendation system to establish sequential dependencies between user-item interactions in a sequence. However, sequence pattern mining usually generates a large number of redundant patterns, which increases unnecessary consumption of time and space. The second category is the latent representation model, which first learns the latent representation of a user or item, and then uses this learned latent representation to predict subsequent user-item interactions. Therefore, more implicit and complex dependencies are captured in a latent space, which greatly improves the recommendation function of the recommendation system. The third category is the deep neural network in the recommendation system. This category can be divided into two subcategories: sequence recommendation systems based on basic deep neural networks and sequence recommendation systems based on neural networks with advanced models. Basic neural networks include sequence recommendation systems based on recurrent neural networks, sequence recommendation systems based on convolutional neural networks, and sequence recommendation systems based on graph neural networks. In order to solve the limitations of sequence recommendation systems based on basic neural network structures, some advanced models such as the more commonly used attention models, memory networks and hybrid models are usually combined with a basic deep neural network to build a more authoritative sequence recommendation system. Generally speaking, memory networks have shown their potential in sequence recommendation systems, but have not been fully studied. Sequential recommendation systems based on hybrid models combine different models that are good at capturing different types of dependencies to enhance the ability of the entire model to capture multiple dependencies in order to give better recommendations. However, these models are in the early stages.

[0005] Although deep learning technology has demonstrated its potential in sequential recommendation systems, previous work has rarely considered the impact of uncertainty in user intentions caused by external uncertain environments on recommendations. This aspect has not been fully studied, so there is an urgent need for a deep model sequential recommendation method and system based on user representation. Summary of the invention

[0006] In order to solve the above problems, the present disclosure proposes a deep model sequence recommendation method and system based on user representation. The present disclosure can effectively process long-term sequences with a deep learning-based method, and use the uncertainty of user intentions to balance the user's long-term and short-term preferences.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] A deep model sequence recommendation method based on user representation, comprising the following steps:

[0009] Obtain the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference respectively;

[0010] The recommendation results are obtained through the deep sequence recommendation model MLUR;

[0011] Among them, the deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module.

[0012] Furthermore, in the short-term modeling representation learning module, a self-attention sequence recommendation model is used to capture the item transitions of users in short-term sequences using a hierarchical self-attention network.

[0013] Furthermore, in the long-term modeling representation learning module, recurrent neural networks (RNNs) are used to model the function of capturing sequence patterns.

[0014] Furthermore, the recurrent neural network update sequence process is:

[0015] h t =g(x t W+h t-1 A);

[0016] Among them, g is the activation function, x t is the user's current interaction item in the long-term sequence, h t-1 is the previous hidden state.

[0017] Furthermore, in the long-term modeling representation learning module, a gated recurrent unit GRU is used as a recurrent unit for sequence recommendation.

[0018] Furthermore, the GRU is specifically:

[0019]

[0020]

[0021]

[0022] h t =z t ⊙h t-1 +(1-z t )⊙h′ t ;

[0023] in h t-1 ,h t, h′ t , z t , W z , W r , W h , A z , A r , z t and r t represent the update gate and reset gate respectively, h′ t is the candidate state, h t is the output hidden state, σ and tanh are activation functions, and + and ⊙ represent element-wise addition and element-wise multiplication operations, respectively.

[0024] Furthermore, in the long-term modeling representation learning module, the time interval context is introduced to encode and process the long-term sequence, and the time interval context is modeled by the GRU network. evolution.

[0025] Furthermore, in the gated fusion module, the item similarity gated fusion model is used to calculate the weights of short-term and long-term representations to balance the contribution of long-term and short-term sequences.

[0026] Furthermore, in the gated fusion module, the Adam optimizer is used to minimize the binary cross entropy loss to train the MLUR, and the loss function is:

[0027]

[0028] A deep model sequence recommendation system based on user representation, comprising:

[0029] The sequence acquisition module is configured to respectively acquire the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference;

[0030] The deep sequence recommendation module is configured to obtain recommendation results through the deep sequence recommendation model MLUR;

[0031] Among them, the deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present disclosure can effectively process long-term sequences with a deep learning-based method, and use the uncertainty of user intent to balance the user's long-term and short-term preferences; through a gating module based on a multi-layer perceptron MLP, the contribution ratio of long-term and short-term representations is determined by considering the relationship between hot-selling items, recently interacted items, and long-term preferences. The module balances long-term and short-term representations by considering the information of hot-selling items, thereby simultaneously processing the dynamics of user intent. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0035] Figure 1 is an architectural diagram of this embodiment;

[0036] Figure 2 Graph 1 is a network structure diagram of long-term preference representation learning in this embodiment. DETAILED DESCRIPTION

[0037] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0038] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0040] Example 1.

[0041] The present invention is a deep model based on long-term and short-term user representation for sequence recommendation. The deep model MLUR structure proposed in the present invention includes three parts: short-term modeling representation learning, long-term modeling representation learning and a gate model for balancing the two representations.

[0042] like Figure 1 As shown, a deep model sequence recommendation method based on user representation includes the following steps:

[0043] Obtain the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference respectively;

[0044] The recommendation results are obtained through the deep sequence recommendation model MLUR;

[0045] Among them, the deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module.

[0046] Furthermore, in the short-term modeling representation learning module, a self-attention sequence recommendation model is used to capture the item transitions of users in short-term sequences using a hierarchical self-attention network.

[0047] Furthermore, in the long-term modeling representation learning module, recurrent neural networks (RNNs) are used to model the function of capturing sequence patterns.

[0048] Furthermore, the recurrent neural network update sequence process is:

[0049] h t =g(x t W+h t-1 A);

[0050] Among them, g is the activation function, x t is the user's current interaction item in the long-term sequence, h t-1 is the previous hidden state.

[0051] Furthermore, in the long-term modeling representation learning module, a gated recurrent unit GRU is used as a recurrent unit for sequence recommendation.

[0052] Furthermore, the GRU is specifically:

[0053]

[0054]

[0055]

[0056] h t =z t ⊙h t-1 +(1-z t )⊙h′ t ;

[0057] in h t-1 ,h t , h′ t , z t , W z , W r , W h, A z , a r , z t and r t represent the update gate and reset gate respectively, h′ t is the candidate state, h t is the output hidden state, σ and tanh are activation functions, and + and ⊙ represent element-wise addition and element-wise multiplication operations, respectively.

[0058] Furthermore, in the long-term modeling representation learning module, the time interval context is introduced to encode and process the long-term sequence, and the time interval context is modeled by the GRU network. evolution.

[0059] Furthermore, in the gated fusion module, the item similarity gated fusion model is used to calculate the weights of short-term and long-term representations to balance the contribution of long-term and short-term sequences.

[0060] Furthermore, in the gated fusion module, the Adam optimizer is used to minimize the binary cross entropy loss to train the MLUR, and the loss function is:

[0061]

[0062] Specifically,

[0063] 1. Short-term modeling representation learning module.

[0064] In short-term modeling representation learning, self-attention sequence recommendation SASRec uses a hierarchical self-attention network to capture the user's item transitions in short-term sequences.

[0065] Embedding layer. First, the short-term sequence S u As input sequence, let It is represented as a learnable item embedding matrix with d-dimensional latent space. So the input sequence can be represented as an embedding matrix

[0066] To capture the influence of position, a learnable position embedding matrix is ​​added to the input matrix E, and get the input matrix To the self-attention network:

[0067] Self-attention block. Then, the sequence X (0) Provided to a series of self-attention blocks (SABS), where the output of the b-th block is:

[0068] X (b) =SABS (b)(X (b-1) ),b∈{1,2,…,B} (2)

[0069] Ignoring the normalization layer of the residual connection, each self-attention block can be represented as a self-attention layer SAL(·) with a feed-forward layer FFL(·):

[0070] SABS(X)=FFL(SAL(X)) (3)

[0071]

[0072] FFL(X′)=ReLU(X′W (1) +b (1) )W (2) +b (2) (5)

[0073] X is the position-aware input matrix, Q = XW Q ,K=XW K and V = W V ,W Q ,W K and Represent the projection query, key and value matrices respectively to improve flexibility. Note that W (1) , and b (1) , are the weights and biases of the two convolutional layers. The normalization and dropout layers used in this module are the same as those in the Attention mechanism method.

[0074] In this module, the output vector from the last self-attention block Recorded as It is the representation of short-term behavior sequence, which represents the current needs of users.

[0075] 2. Long-term modeling representation learning module.

[0076] Because RNNs have an outstanding ability to capture sequential patterns, they have achieved great success in user modeling. The sequence update process can be expressed as:

[0077] h t =g(x t W+h t-1 A) (6)

[0078] g is the activation function, x t is the user's current interaction item in the long-term sequence, h t-1is the previous hidden state. Long short-term memory network (LSTM) is a popular RNN variant, and gated recurrent unit GRU is a simplified version of LSTM. It retains the properties of LSTM while simulating the LSTM structure, and the effect is also very good, so it is also a very popular network at present. In the present invention, the GRU network is used as the recurrent unit for sequence recommendation, and GRU can be specifically described as:

[0079]

[0080]

[0081]

[0082] h t =z t ⊙h t-1 +(1-z t )⊙h′ t (10)

[0083] in h t-1 ,h t , h′ t , z t , W z , W r , W h , A z , A r , z t and r t represent the update gate and reset gate respectively, h′ t is the candidate state, h t is the output hidden state, σ and tanh are activation functions, and + and ⊙ represent element-wise addition and element-wise multiplication operations, respectively.

[0084] From a long-term perspective, the relationship between two interactions in a user behavior sequence is not only related to their relative positions, but also highly affected by the information between items in the sequence. Because the short time interval between two items has a greater impact on users than the long time interval, and the time interval between different items is also different. Therefore, the present invention introduces the time interval context to encode and process long-term sequences. Time context It occurs between two adjacent items. The present invention proposes to model the time interval context through the GRU network The evolution is as follows:

[0085] z t =σ(W z c t +A z pt-1 ) (11)

[0086] r t =σ(W r c t +A r p t-1 ) (12)

[0087] p′ t =tanh(W p c t +r t ⊙A p p t-1 ) (13)

[0088] p t =z t ⊙p t-1 +(1-z t )⊙p′ t (14)

[0089] In GRU, the propagation sequence signal of the recurrent connection weight matrix A is represented in the successive hidden layers p t and p t-1 Instead of A, an intermediate feed-forward layer is used to transform p t Convert to M t , to maintain the low-dimensional latent representation p t To a higher dimensional representation M t Then, connect the two GRU networks. Figure 2 As shown. So the dynamic d-dimensional hidden layer user h t It can be expressed as follows:

[0090] M t =σ(p t F) (14)

[0091]

[0092]

[0093]

[0094] h t =z t ⊙h t-1 +(1-z t )⊙h′ t (18)

[0095] The temporal context representation M of the transformation t Used in h t-1 and h t Use M tSubstitute A is mainly used to combine changing dynamic time interval contexts.

[0096] Therefore, the final output of the hidden layer h t As a representation of a long-term series, denoted by k = h t .

[0097] 3. Gated fusion module.

[0098] In order to better combine the short-term representation and the long-term representation, it may be natural to think of connecting these two parts or performing a weighted sum operation. However, these methods are based on an ideal situation that all user preferences are reflected in the historical sequence. But in the real world, the user's intention will be affected by many factors and become uncertain. For example, the SDM model proposed by Alibaba Group, after modeling the long-term and short-term behavior representations of users respectively, also combines long-term and short-term behaviors to more accurately represent user behavior. When combining long-term and short-term behaviors, they designed a gate neural network that not only takes the user's long-term and short-term behavior representations as the input of the gate neural network, but also takes the user configuration as the common input of the gate neural network. The gate vector obtained is then used to determine the contribution percentage of the long-term and short-term at time t. The model of the present invention is the same as the gated fusion module in SDM in that it not only takes the two parts of the user's long-term and short-term preference representations as the input of the gate neural network, but also considers other factors that affect the recommendation. SDM considers the user's own configuration, while the present invention considers the user's intention with uncertainty.

[0099] In order to solve the uncertainty problem of user intention in sequential recommendation, the present invention designs a gated network based on multi-layer perceptron. i ∈I leads to a change in user intention, and inspired by the fusion of item similarity models and self-attention networks (FISSA), this paper proposes an item similarity gating fusion model to calculate the weights of short-term and long-term representations to balance the contribution of long and short-term sequences.

[0100] Specifically, by modeling hot-selling products And the latest interactive product m sl There is also the similarity between historical items k to calculate the weight, using the hot-selling product similarity gating function HISG, and then writing the single-level gating into the MLP as follows:

[0101]

[0102] HISG(·) is the hot selling product gate function, [·] represents the ternary cascade operation, Represent the weights and biases to be learned respectively. Use the sigmoid function σ(ζ)=1 / (1+e-ζ ) as an activation function can limit the value of G to between 0 and 1.

[0103] The final representation of the user behavior sequence in step l is obtained by the weighted sum of the corresponding short-term representation and long-term representation as follows:

[0104] o l =G⊙j+(1-G)⊙k (20)

[0105] Finally, at the next time t+1, the probability of predicting item i to become the preferred item of user u is:

[0106] p i+1 =σ(r i+1 ) (twenty one)

[0107] σ is the sigmoid function, r i+1 is the predicted score of user u for item i at time t+1, defined as

[0108]

[0109] MLUR is trained by minimizing the binary cross entropy loss using the Adam optimizer. The loss function is as follows:

[0110]

[0111] Example 2.

[0112] A deep model sequence recommendation system based on user representation, comprising:

[0113] The sequence acquisition module is configured to respectively acquire the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference;

[0114] The deep sequence recommendation module is configured to obtain recommendation results through the deep sequence recommendation model MLUR;

[0115] Among them, the deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module.

[0116] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0120] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0121] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A deep model sequence recommendation method based on user representation, characterized in that: The following steps are involved: Obtain the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference respectively; The recommendation results are obtained through the deep sequence recommendation model MLUR; The deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module; In the short-term modeling representation learning module, a self-attention sequence recommendation model is used to capture the item transition of users in short-term sequences using a hierarchical self-attention network; In the long-term modeling representation learning module, the recurrent neural network RNNs is used to capture the sequence pattern for modeling, and the gated recurrent unit GRU is used as the recurrent unit for sequence recommendation; In the gated fusion module, the item similarity gated fusion model is used to calculate the weights of short-term and long-term representations to balance the contributions of long-term and short-term sequences; By modeling hot selling products , interactive products and historical projects k The similarity between them is used to calculate the weight, and the hot-selling product similarity gating function HISG is used to write the single-level gating into the MLP: G = ( HISG ( )) = ( [ ] + )； Among them, HISG(·) is the hot-selling product gate function, [·] represents the ternary cascade operation, , Respectively represent the weights and biases to be learned; use the sigmoid function ( ) = 1 / (1+ ) as an activation function to limit the value of G to between 0 and 1; User behavior sequence is in l The final representation of the step is obtained by the weighted sum of the corresponding short-term representation and long-term representation: = G + (1-G) k’ ; At the next time t+1, the probability of predicting item i to become the preferred item of user u is: = （ ）； in, is the sigmoid function, is the predicted score of user u for item i at time t+1, Defined as: = score( ) = ; In the gated fusion module, the Adam optimizer is used to minimize the binary cross entropy loss to train the MLUR, and the loss function is: = (log ( )) + log (1- )) 。 2. A method for recommending a sequence of deep models based on user representation as claimed in claim 1, characterized in that: The recurrent neural network update sequence process is: = g ( W+ A )； Among them, g is the activation function, is the user’s current interaction item in the long-term sequence, is the previous hidden state.

3. The method for recommending a sequence of deep models based on user representation according to claim 1, characterized in that: The GRU is specifically: = ； = ； = fishy ); = + (1 - ) ； in , , , , , , , , , , , ; and represent the update gate and reset gate respectively, For candidate status, is the output hidden state, and tanh are activation functions, + and Represent element-wise addition and element-wise multiplication operations, respectively.

4. The method for recommending a sequence of deep models based on user representation according to claim 1, characterized in that: In the long-term modeling representation learning module, the time interval context is introduced to encode and process long-term sequences, and the time interval context is modeled by the GRU network. evolution.

5. A user representation based deep model sequence recommendation system, based on a user representation based deep model sequence recommendation method according to any one of claims 1 to 4, characterized in that: include: The sequence acquisition module is configured to respectively acquire the long-term sequence of the user's overall preference and the short-term sequence of the user's current dynamic preference; The deep sequence recommendation module is configured to obtain recommendation results through the deep sequence recommendation model MLUR; The deep sequence recommendation model includes a short-term modeling representation learning module, a long-term modeling representation learning module and a gated fusion module, and the gated fusion module is combined with the short-term modeling representation learning module and the long-term modeling representation learning module; In the short-term modeling representation learning module, a self-attention sequence recommendation model is used to capture the item transition of users in short-term sequences using a hierarchical self-attention network; In the long-term modeling representation learning module, the recurrent neural network RNNs is used to capture the sequence pattern for modeling, and the gated recurrent unit GRU is used as the recurrent unit for sequence recommendation; In the gated fusion module, the item similarity gated fusion model is used to calculate the weights of short-term and long-term representations to balance the contributions of long-term and short-term sequences; By modeling hot selling products , interactive products and historical projects k The similarity between them is used to calculate the weight, and the hot-selling product similarity gating function HISG is used to write the single-level gating into the MLP: G = ( HISG ( )) = ( [ ] + )； Among them, HISG(·) is the hot-selling product gate function, [·] represents the ternary cascade operation, , Respectively represent the weights and biases to be learned; use the sigmoid function ( ) = 1 / (1+ ) as an activation function to limit the value of G to between 0 and 1; User behavior sequence is in l The final representation of the step is obtained by the weighted sum of the corresponding short-term representation and long-term representation: = G + (1-G) k’ ; At the next time t+1, the probability of predicting item i to become the preferred item of user u is: = （ ）； in, is the sigmoid function, is the predicted score of user u for item i at time t+1, Defined as: = score( ) = ; In the gated fusion module, the Adam optimizer is used to minimize the binary cross entropy loss to train the MLUR, and the loss function is: = (log ( )) + log (1- )) 。

Images