Item Recommendation Method Based on Consensus Preferences and Personalized Preferences

Abstract

The present invention proposes an item recommendation method based on consensus preference and personalized preference. The implementation steps are as follows: constructing a training sample set for the representation construction model CF, a training sample set for the recommendation model UPN, and a test sample set; constructing the representation construction model CF; initializing parameters; training the representation construction model CF; obtaining the trained representation construction model; constructing the consensus preference and personalized preference fitting recommendation model UPN; initializing parameters; training the consensus preference and personalized preference fitting recommendation model UPN; obtaining the trained consensus preference and personalized preference fitting recommendation model UPN; and obtaining the item recommendation result. The fully connected network in the consensus preference and personalized preference fitting recommendation model of the present invention has strong fitting ability, can simultaneously fit the consensus preference and personalized preference of users, capture the personalized preference of users more accurately at a finer level, and effectively improve the accuracy of item recommendation.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and relates to an item recommendation method, in particular to an item recommendation method based on consensus preference and personalized preference recommendation. Background Art

[0002] The Internet has a vast amount of information. However, the upper limit of the amount of information that users can obtain on the Internet is not the upper limit of the amount of information provided by information providers, but the physiological limit of the amount of information that users can obtain. For example, in the field of music recommendation, to increase user stickiness and have more users on a certain music recommendation platform, it is necessary to recommend music that users are interested in from a large number of music. In this way, it is necessary to screen a vast amount of information. This task is the main task of the recommendation system, that is, to screen a vast amount of information to alleviate the problem of information overload.

[0003] In item recommendation methods, recall rate and hit rate are usually used as evaluation indicators. The recall rate mainly measures the proportion of correctly predicted samples among all true positive samples. The closer this indicator is to 1, the higher the accuracy of the model recommendation. The hit rate mainly measures the proportion of items successfully recommended by the model among the recommended items. The closer this indicator is to 1, the higher the accuracy of the model recommendation.

[0004] In the field of item recommendation, the item recommendation method based on machine learning models is the current mainstream method. In recent years, with the rapid development of graph neural networks, and since the interactions between users and items can naturally be used to construct graph-structured data, applying graph neural networks to item recommendation has become a new recommendation method. For example, He et al. proposed a graph convolutional network-based item recommendation method GraphDA in their paper "Graph Collaborative Signal Denoising and Augmentation for Recommendation" (Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2023, pp. 2037–2041). The core of this method's recommendation is the collaborative filtering idea, that is, taking advantage of the similarity of the preferences of users who have interacted with the same item, that is, using the consensus preferences among users to recommend items for users. GraphDA first pre-trains the classic recommendation model LightGCN to obtain the representation vectors of users and items; then, based on the representation vectors obtained in the first step, by constructing a user-item augmented bipartite graph composed of a user-user augmented adjacency matrix, an item-item augmented adjacency matrix, and a user-item augmented adjacency matrix, and using this bipartite graph to construct a new graph neural network to further extract features from the representation vectors obtained in the first step, so as to obtain the final representation vectors of users and items; finally, taking the inner product of the final representation vectors of users and items to obtain the interaction prediction probability of users for items, and recommending the top Topk items with the highest interaction prediction probability to users. The graph neural network in this method utilizes the high-order connectivity of the user-item bipartite graph and improves the accuracy of item recommendation to a certain extent. However, its shortcoming is that GraphDA only fits the consensus preferences of users by using the inner product method of vectors, without considering the phenomenon of preference contradictions among users, that is, the phenomenon that users who have interacted with the same item may have different preferences, and the personalized preference information of users it contains, resulting in still low accuracy of item recommendation. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects existing in the above-mentioned prior art, and propose an item recommendation method based on consensus preferences and personalized preferences, which is used to solve the problem in the prior art that the influence of user personalized preferences on the recommendation accuracy is not considered.

[0006] To achieve the above purpose, the technical solutions adopted by the present invention include the following steps:

[0007] (1) Construct a training sample set L to represent the construction model CF CF , training sample set L of the recommended model UPN UPN And the test sample set L test ;

[0008] (2) Constructing the representation structure model CF:

[0009] Construct a representation construction model CF including a sequentially connected embedding layer, a layer combination module, a prediction layer, and a graph convolution module. The graph convolution module includes H sequentially connected graph convolution neural networks GCN. The output of the embedding layer is also connected to the input of the first graph convolution neural network GCN1. The input of the layer combination module is also connected to the output of each graph convolution neural network, where H ≥ 1.

[0010] (3) Initialization parameters:

[0011] The number of initial iterations is t, the maximum number of iterations is T, T>1000, and the parameters of the CF embedding layer of the constructed model are characterized The value of is a random number that obeys the standard normal distribution, and let t = 0;

[0012] (4) Training the representation construction model CF:

[0013] The training sample set L CF As the input of the representation construction model CF, forward propagation is performed to obtain each user u m With two items i n1 、i n2 Probability of interaction

[0014] (5) Obtain the trained representation construction model:

[0015] Use the BPR loss function and pass the probability value Calculate the loss value L that represents the constructed model CF BPR , then through L BPR For the embedding layer parameters The partial derivative of right Update to get the recommended model CF for this iteration t Finally, determine whether t>T is true. If so, get the trained representation construction model CF * Otherwise, let t = t + 1, CF = CF t , and execute step (4);

[0016] (6) Constructing consensus preference and personalized preference fitting recommendation model UPN:

[0017] Construct a model CF including the trained representation *The embedding layer in, and the fully connected module including K fully connected networks connected in sequence thereto, the loss function It is a consensus preference and personalized preference fitting recommendation model UPN with a positive example fitting loss function, where:

[0018]

[0019] Among them, is the m-th element of the interaction vector of item i n and is the m-th element of the item interaction prediction vector output by UPN, is the negative example mask of item i n , K≥2;

[0020] (7) Initialize parameters:

[0021] Initialize the number of iterations as t, the maximum number of iterations as T, T>1000, the weight of the k-th fully connected network of the consensus preference and personalized preference fitting recommendation model UPN bias is a random number obeying the standard normal distribution, the recommended preference degree γ of UPN, the learning rate is lr, and let t=0, UPN t =UPN;

[0022] (8) Train the consensus preference and personalized preference fitting recommendation model UPN:

[0023] Use the training sample set L UPN as the input of the consensus preference and personalized preference fitting recommendation model UPN for forward propagation, and obtain the interaction prediction vector of a single item i n with all users

[0024] (9) Obtain the trained consensus preference and personalized preference fitting recommendation model UPN:

[0025] Adopt the positive example fitting loss function and through the interaction prediction vector and the training sample set L of UPN UPN in the sample calculate the loss value of the recommendation model UPN Then through the partial derivative of the UPN model parameters for update to obtain the recommendation model CF of this iteration t , finally judge whether t>T holds. If so, obtain the trained recommendation model UPN * , otherwise, let t=t + 1, UPN t+1 =UPNt and execute step (8);

[0026] (10) Obtain the item recommendation result:

[0027] Take the test sample set L test as the input of the trained consensus preference and personalized preference fitting recommendation model UPN * to perform forward propagation, and obtain the interaction prediction vector of item i n with all users where the m-th element of is the interaction prediction probability of user u m and item i n The interaction prediction probability of user u

[0028] Sort the interaction probabilities of user u m with all un-interacted items from large to small. Among them, the top Topk items are the items recommended by user u m recommended items.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] Based on the trained representation construction model, the present invention constructs a consensus preference and personalized preference fitting recommendation model. In the process of training the model and obtaining the recommendation result, the fully connected network has strong fitting ability, and can simultaneously fit the consensus preference and personalized preference of users, and capture the personalized preference of users ignored by the prior art more accurately at a finer level, effectively improving the accuracy of item recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is the implementation flowchart of the present invention;

[0032] Figure 2 is the structural schematic diagram of the representation construction model and the consensus preference and personalized preference fitting recommendation model in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] The following further describes the present invention in detail with reference to the drawings and specific embodiments. It should be noted that the specific embodiments are only used to explain the present invention and are not used to limit the present invention.

[0034] Refer to Figure 1 for the following steps of the present invention.

[0035] Step 1) Construct the training sample set L of the representation construction model CF CF , the training sample set L of the recommendation model UPN UPN and the test sample set L test :

[0036] Step 1a) Get M users U = {u1, ..., u m ,...,u M} for each user u m Among N items, I={i1,...,i n ,...,i N Some items are interactive User interaction information And the interaction information of all users is combined into a total interaction information set with interaction number |L| Among them, M>0, N>0, u m represents the mth user, i n represents the nth item, l md Represents user u m With items The interactive information In this embodiment, M=1892, N=17632, |L|=92834;

[0037] Step 1b) Construct a user-item bipartite graph in matrix form using the interactive information set L1 consisting of more than half of the elements in the total interactive information set L And will collect As the training sample set for characterizing the constructed model CF, R is the mth row and nth column element r m,n The interaction matrix, R T Represents the transposed result of R, when (u m ,i n )∈L1, r m,n =1, when When m,n =0; represents the normalized result of A, is a degree diagonal matrix, D -1 / 2 is the inverse of the arithmetic square root of D, the diagonal element d in the kth row of D kk is the number of non-zero elements in the kth row of A;

[0038] Step 1c) The set constructed by L1 As the training sample set of the recommended model UPN, and the interaction information except L1 is used as the test sample set L of the recommended model UPN test ,in Indicates item i n The interaction vector.

[0039] Step 2) Construct the representation structure model CF:

[0040] Reference Figure 2, the representation construction model includes an embedding layer, a layer combination module, and a prediction layer connected in sequence, as well as a representation construction model CF of a graph convolution module. The graph convolution module includes H graph convolutional neural networks GCN connected in sequence. The output of the embedding layer is also connected to the input of the first graph convolutional neural network GCN1, and the input of the layer combination module is also connected to the output of each graph convolutional neural network, where H≥1; in this embodiment, H = 3.

[0041] Step 3) Initialize parameters:

[0042] Initialize the number of iterations as t and the maximum number of iterations as T, where T>1000. The parameters of the embedding layer of the representation construction model CF are random numbers obeying the standard normal distribution, and let t = 0; in this embodiment, T = 1100.

[0043] Step 4) Train the representation construction model CF:

[0044] Step 4a) The embedding layer encodes all users and items to obtain the original representation vector m of user u and the original representation vector n of item i where S is the dimension of the representation vector; in this embodiment, S = 64;

[0045] Step 4b) The first graph convolutional neural network GCN1 in the graph convolution module performs convolution on the and to form the original representation matrix E (0) and the normalized user-item bipartite graph in the training sample to obtain the representation matrix E (1) , and each graph convolutional neural network GCN h performs convolution on E (h-1) and to obtain the representation matrix E (h) , where the representation vector of user u (h) in E m is and the representation vector of item i n is where:

[0046]

[0047]

[0048] Step 4c) The layer combination module performs layer combination on and the output of each graph neural network to obtain user u m and item i nThe final representation vector

[0049]

[0050]

[0051] Among them, α h is the weight of the representation vector of the user or item in E (h) when the layer combination module performs weighted summation; in this embodiment, α h = 1 / 4;

[0052] Step 4d) The prediction layer performs inner products on and respectively, and obtains the probability values of the interaction between each user u m and two items i n1 , i n2

[0053]

[0054] Among them, is 's transpose;

[0055] Step 5) Obtain the trained representation construction model:

[0056] Adopt the BPR loss function, and calculate the loss value L of the representation construction model CF through the probability value BPR , and the formula is as follows:

[0057]

[0058]

[0059] Among them, is the set of items interacted by user u m , λ is the weight coefficient, and ∥∥· 2 represents the 2-norm operation; in this embodiment, λ = 10 -5 ;

[0060] Then, through the partial derivative of L BPR with respect to the embedding layer parameter with respect to is updated to obtain the recommendation model CF t of this iteration, and the update formula is as follows:

[0061]

[0062] Among them, ​​The result of updating the CF model parameters, where lr is the learning rate; in this embodiment, lr = 0.001.

[0063] Judge whether t > T holds. If so, obtain the trained representation construction model CF * , otherwise, set t = t + 1, CF = CF t , and execute step 4);

[0064] Step 6) Construct the consensus preference and personalized preference fitting recommendation model UPN:

[0065] Refer to Figure 2 , the recommendation model UPN includes the trained representation construction model CF * in the embedding layer, and a fully connected module including K fully connected networks connected in sequence to it. The loss function is the consensus preference and personalized preference fitting recommendation model UPN of the positive example fitting loss function. Among them, the reason for using the fully connected module to fit the user preference is that the traditional method of using the inner product to fit the user preference is too simple and can only fit the consensus preference of the user. However, the fully connected networks in the fully connected module have the ability to approximate any function. Therefore, it has a stronger fitting ability than the inner product operation and can fit both the consensus preference and the personalized preference of the user. The expression is:

[0066]

[0067] Among them, is the m-th element of the interaction vector of item i n , is the m-th element of the item interaction prediction vector output by UPN, is item i n 's negative example mask, K ≥ 2. Among them, the negative example mask The calculation formula is:

[0068]

[0069] where rand is a random number in the interval [0, 1], γ represents the recommended preference degree of UPN, represents the exposure of item i n , that is, the proportion of the number of users interacting with i n in the total number of users. Among them, the larger γ is, the greater the fitting strength of the fully connected module of UPN to the personalized preference of the user. On the contrary, the greater the fitting strength of the fully connected module of UPN to the consensus preference of the user; in this embodiment, K = 4, γ = 12.

[0070] Step 7) Initialize the parameters:

[0071] Initialize the number of iterations as t, the maximum number of iterations as T, where T > 1000, and the weights of the k-th fully connected network of the consensus preference and personalized preference fitting recommendation model UPN bias is a random number obeying the standard normal distribution, the recommended preference degree γ of UPN, the learning rate is lr, and let t = 0, UPN t = UPN; in this embodiment, T = 1100 and lr = 0.001.

[0072] Step 8) Train the consensus preference and personalized preference fitting recommendation model UPN:

[0073] Take the training sample set L UPN as the input of the consensus preference and personalized preference fitting recommendation model UPN for forward propagation, and obtain the interaction prediction vector of a single item i n with all users The forward propagation process is as follows: First, the embedding layer maps i n in the training sample to obtain the item representation vector Subsequently, in the fully connected module perform feature mapping in sequence on K fully connected networks, and finally obtain the interaction prediction vector of a single item i n with all users The expression of the K-th fully connected network is:

[0074] x (k+1) = Sigmoid(W (k) x (k) + b (k) )

[0075] where Sigmoid(·) is the Sigmoid non-linear activation function, when k = 0, when k = K, and for the other x (k) are all the outputs of the k-th fully connected network.

[0076] Step 9) Obtain the trained consensus preference and personalized preference fitting recommendation model UPN:

[0077] Adopt the positive example fitting loss function defined in Step 6 and calculate the loss value of the recommendation model UPN through the interaction prediction vector and the training sample set L of UPN UPN in the sample Then, through the partial derivative of the UPN model parameters with respect to is updated, and the update formula is: ​

[0078]

[0079]

[0080] wherein are respectively the update results;

[0081] After the update is completed, the recommended model CF for this iteration is obtained t , and finally it is judged whether t>T holds. If so, the trained recommended model UPN * is obtained. Otherwise, let t=t + 1, and UPN t+1 = UPN t , and step 8) is executed.

[0082] Step 10) Obtain the item recommendation result:

[0083] Use the test sample set L test as the input of the trained consensus preference and personalized preference fitting recommendation model UPN * to perform forward propagation, and obtain the interaction prediction vector of item i n with all users wherein the m-th element of is the interaction prediction probability of user u m and item i n

[0084] Sort all the items that user u m has not interacted with in descending order according to the interaction probability to obtain an item recommendation list, where the top Topk items in the item recommendation list are the items recommended by user u m . In this embodiment, Topk = 10.

[0085] The technical effects of the present invention will be further described below in combination with simulation experiments.

[0086] 1. Simulation conditions and content:

[0087] The operating environment of the simulation experiment is: Windows 10 operating system, CPU: 11 th Gen Core TM i7-11800H@2.30GHz, with 16GB of memory, GPU is NVIDIA GeForce RTX 3050Ti Laptop GPU, programming language Python, compiler version 3.7.0, and machine learning library is Pytorch.

[0088] In the simulation experiment, three datasets commonly used in the field of real-world recommendation systems were adopted: the Lastfm dataset collected from the Lastfm website on the interactions between users and songs; a dataset Citeulike and Citeulike-t that records users' self-evaluations of certain artworks and their mutual visits. The statistical results of the three datasets are shown in Table 1 below.

[0089] Table 1

[0090]

[0091] A comparative simulation was conducted on the recall rate (Recall) and hit rate (Hit Rate) of the present invention and the existing item recommendation method GraphDA based on graph convolutional network. The results are shown in Table 2.

[0092] 2. Analysis of simulation results:

[0093] To better evaluate the technical effects of the present invention, the Recall and Hit Rate of the recommended items obtained by the present invention and the prior art on the three datasets were calculated respectively with respect to their corresponding true recommended items.

[0094] Table 2

[0095]

[0096] As can be seen from Table 2, the recall rate and hit rate of the present invention on the three datasets have increased by varying degrees compared with the prior art, indicating that the present invention effectively improves the accuracy of recommendation.

[0097] The above simulation experiments show that: the present invention can use the personalized preference information contained in the preference contradiction phenomenon to recommend for users, thus achieving further performance improvement on the basis of traditional recommendation models. Under the action of more effective consensus preference and personalized preference recommendation methods, the item recommendation model proposed by the present invention has significantly improved performance compared with traditional item recommendation methods.

Claims

1. An item recommendation method based on consensus preference and personalized preference, characterized in that including the following steps: (1) Construct a training sample set \(L\) for representing the construction model \(CF\). CF and the training sample set \(L\) of the recommendation model \(UPN\). UPN and the test sample set \(L\). test ; (2)Construct the representation construction model CF: Construct a representation construction model CF including a sequentially connected embedding layer, a layer combination module, a prediction layer, and a graph convolution module. The graph convolution module includes H graph convolutional neural networks GCN connected in sequence. The output of the embedding layer is also connected to the input of the first graph convolutional neural network GCN1, and the input of the layer combination module is also connected to the output of each graph convolutional neural network, where H≥1; (3)Initialize the parameters: Initialize the number of iterations as t, the maximum number of iterations as T, where T > 1000, representing the parameters of the embedding layer of the constructed model CF The value is a random number following the standard normal distribution, and let t = 0; (4)Train the representation construction model CF: Use the training sample set L CF as the input for forward propagation of the representation construction model CF, obtaining, for each user u m the probability values of interacting with two items i n1 and i n2 ​ (5)Obtain the trained representation construction model: Adopt the BPR loss function and calculate the loss value L representing the constructed model CF through the probability value Then, update the partial derivative of the embedding layer parameter BPR with respect to L BPR to obtain the recommended model CF for this iteration Finally, determine whether t > T holds. If so, obtain the trained constructed model CF Otherwise, set t = t + 1, CF = CF and execute step (4); t * t t t t (6)Construct a consensus preference and personalized preference fitting recommendation model UPN: Construct a consensus preference and personalized preference fitting recommendation model UPN including an embedding layer in the trained representation construction model CF * and a fully connected module including K fully connected networks sequentially connected thereto, and a loss function is a positive example fitting loss function. wherein, is the m-th element of the interaction vector of item i n , is the m-th element of the item interaction prediction vector output by the UPN is the negative example mask of item i n , K≥2, rand is a random number in the range of [0,1], and γ represents the recommendation preference degree of the UPN represents the exposure of item i n , that is, the proportion of the number of users interacting with i n to the total number of users (7)Initialize the parameters: Initialize the number of iterations as t, the maximum number of iterations as T, where T > 1000, the consensus preference and the weights of the k-th fully connected network of the personalized preference fitting recommendation model UPN bias is a random number obeying the standard normal distribution, the recommended preference degree γ of UPN, the learning rate is lr, and let t = 0, UPN t = UPN; (8)Train the consensus preference and personalized preference fitting recommendation model UPN: Use the training sample set L UPN as the input of the consensus preference and personalized preference fitting recommendation model UPN for forward propagation, and obtain the interaction prediction vector of a single item i n with all users (9)Obtain the trained consensus preference and personalized preference fitting recommendation model UPN: Adopt a positive example fitting loss function And through the interactive prediction vector And the training sample set L of UPN UPN In the sample Calculate the loss value of the recommendation model UPN Then through The partial derivative of the UPN model parameters For Perform an update to obtain the recommendation model CF for this iteration t , finally judge whether t > T holds. If so, obtain the trained recommendation model UPN * , otherwise, set t = t + 1, UPN t+1 = UPN t , and execute step (8); (10)Obtain the item recommendation result: Take the test sample set L test As the input of the trained consensus preference and personalized preference fitting recommendation model UPN * Perform forward propagation to obtain the interaction prediction vector of item i n And all users Where The m-th element of Is the user u m And item i n The interaction prediction probability of Rank the interaction probabilities of user u m with all items that have not been interacted with from high to low. The top Topk items are the items recommended for user u m ​ 2. The method according to claim 1, characterized in that, Constructing the training sample set \(L\) for characterizing the construction model \(CF\) described in step (1) CF , the training sample set \(L\) of the recommendation model \(UPN\) UPN and the test sample set \(L\) test , the implementation steps are as follows: (1a) Obtain the user interaction information of each user \(u\) in \(M\) users \(U = \{u_1,\ldots,u_m,\ldots,u_M\}\) with a part of \(N\) items \(I=\{i_1,\ldots,i_n,\ldots,i_N\}\), and form the total interaction information set with the interaction information of all users. m ,\ldots,u M \} in which \(M>0\), \(N>0\), \(u_m\) m represents the \(m\)-th user, \(i_n\) n ,\ldots,i N \} represents the \(n\)-th item, and \(l_{mn}\) represents the interaction information between user \(u_m\) and item \(i_n\). Here, \(M > 0\), \(N > 0\), \(u_m\) m represents the \(m\)-th user, \(i_n\) n represents the \(n\)-th item, \(l_{mn}\) md represents the interaction information between user \(u_m\) m and item \(i_n\). (1b) Construct a user-item bipartite graph in matrix form using the interaction information set L1 composed of more than half of the elements in the total interaction information set L. And use the set as the training sample set for characterizing the construction model CF, where R is the interaction matrix with the element in the m-th row and n-th column being r m,n . R T represents the transpose result of R. When (u m , i n ) ∈ L1, r m,n = 1. When , r m,n = 0; represents the normalization result of A. is the degree diagonal matrix, D -1 / 2 is the inverse of the arithmetic square root of D. The diagonal element d kk in the k-th row of D is the number of non-zero elements in the k-th row of A; (1c) The set constructed through L1 is used as the training sample set of the recommendation model UPN, and the interaction information other than L1 is used as the test sample set L of the recommendation model UPN test , where represents the interaction vector of item i n .

3. The method according to claim 1, wherein The implementation steps of training the representation construction model CF in step (4) are: (4a) The embedding layer encodes all users and items to obtain the original representation vectors of user u m and item i where S is the dimension of the representation vector; n and the original representation vectors of item i where S is the dimension of the representation vector; (4b) The first graph convolutional neural network GCN1 in the graph convolutional module convolves the with to form the original representation matrix E (0) and the normalized user-item bipartite graph in the training samples to obtain the representation matrix E (1) , and each graph convolutional neural network GCN h convolves E (h-1) with to obtain the representation matrix E (h) , where the representation vector of user u (h) in E m is and the representation vector of item i n is Where: (4c) layer combination module pairs with the output of each graph neural network to perform layer combination and obtain the final representation vector of user u m and item i n ​ Among them, α h is the weight of the representation vector of the user or item in E (h) when the layer combination module performs weighted summation; (4d) The prediction layer performs inner products on and respectively, to obtain the probability values of the interaction between each user u m and two items i n1 、i n2 Among them, is the transpose of.

4. The method according to claim 1, characterized in that, The loss value L of the characterization construction model CF described in step (5) BPR , The update formulas are respectively as follows: where is the set of items interacted by user u, λ is the weight coefficient, and ‖·‖ m represents the 2-norm operation, 2 is the updated result of the CF model parameters, and lr is the learning rate.​ 5. The method according to claim 1, characterized in that, The implementation steps of training the consensus preference and personalized preference fitting recommendation model UPN in step (8) are: The embedding layer maps the i in the training samples n to obtain the item representation vector In the fully-connected module Feature mapping is sequentially performed in K fully-connected networks, and finally a single item i n and the interaction prediction vectors with all users The expression of the K-th fully-connected network is as follows: x (k+1) = Sigmoid(W (k) x (k) + b (k) ) Among them, Sigmoid(·) is the Sigmoid non-linear activation function. When k = 0, when k = K, For other x (k) are all the outputs of the k-th fully connected network.

6. The method according to claim 1, wherein The update described in step (9) for is performed, and the update formula is: Among them are respectively update results of

Images