An item recommendation method, system and device based on implicit feedback

Abstract

The present invention discloses an item recommendation method, system and device based on implicit feedback, the method comprising: constructing an information matrix and an information network according to information between users and items and information between items; learning user latent factor vectors, item latent factor vectors, item auxiliary feature vectors and user auxiliary feature vectors according to the information matrix and the information network; constructing a prediction vector according to the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector to make predictions and generate recommended items. The system comprises: a construction module, a learning module and a prediction module. The device comprises a memory and a processor for executing the above-mentioned item recommendation method based on implicit feedback. By using the present invention, item recommendation can be achieved simply and accurately. As an item recommendation method, system and device based on implicit feedback, the present invention can be widely used in the field of item recommendation.

Description

Technical Field

[0001] The present invention relates to the field of item recommendation, and in particular to an item recommendation method, system and device based on implicit feedback. Background Art

[0002] With the rapid development of information technology, the amount of information on the Internet is growing exponentially. Internet users are gradually facing a serious "information overload" problem, that is, users often find it difficult to effectively find the content they are interested in from the massive amount of information. As an effective way to solve the "information overload" problem, the recommendation system has received widespread attention from academia and industry. It has not only become a hot topic in academic research, but has also been widely used in various Internet platforms.

[0003] Commonly used methods of recommendation systems include collaborative filtering recommendation method and content-based recommendation method. Collaborative filtering recommendation method is implemented by using users' historical interaction information. However, in real life, users rarely take the initiative to rate or comment on items. Therefore, collaborative filtering recommendation method faces serious "cold start" and "data sparsity" problems. Content-based recommendation method needs to extract effective features from the original content, which is more difficult. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide an item recommendation method, system and device based on implicit feedback, which can realize item recommendation simply and accurately.

[0005] The first technical solution adopted by the present invention is: an item recommendation method based on implicit feedback, comprising the following steps:

[0006] Construct information matrix and information network based on the information between users and items and the information between items;

[0007] Learn the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector according to the information matrix and information network;

[0008] A prediction vector is constructed based on the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector, and the prediction matching score is calculated.

[0009] Furthermore, it also includes:

[0010] Construct the loss function and perform parameter updates.

[0011] Furthermore, the step of constructing an information matrix and an information network based on the information between users and items and the information between items specifically includes:

[0012] Construct an interactive information matrix based on the interactive information between users and items;

[0013] Construct a generation relationship matrix based on the generation information of users and items;

[0014] Construct a co-generation network based on the generation information of items.

[0015] Furthermore, the step of learning the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector according to the information matrix and the information network specifically includes:

[0016] The interactive information matrix is ​​processed based on a deep neural network multi-layer perceptron to learn the user latent factor vector and the item latent factor vector;

[0017] The co-generation network is processed based on the graph attention network to learn the auxiliary feature vector of the item;

[0018] The generated relationship matrix is ​​combined with the item auxiliary feature vector to generate the user auxiliary feature vector.

[0019] Further, the latent factor vector of the object is learned, and the formula is expressed as:

[0020]

[0021]

[0022]

[0023] In the above formula, represents the interaction matrix of m users to n items, represents the interaction vector of item i, L is the number of layers of the deep neural network multilayer perceptron, are L weight matrices, are L bias vectors, is the L-1 hidden layer representation vector, f(·) represents the activation function, Represents the latent factor vector of item i learned.

[0024] Further, the user latent factor vector is learned, and the formula is expressed as:

[0025]

[0026]

[0027]

[0028] In the above formula, represents the interaction vector of user u, f(·) represents the activation function, Represents the learned latent factor vector of user u.

[0029] Furthermore, the formula for constructing the prediction vector and calculating the prediction matching score is as follows:

[0030]

[0031]

[0032]

[0033] In the above formula, Represents the item prediction vector, which is composed of the item latent factor vector p i and item auxiliary feature vector x i Stitched together, Represents the user prediction vector, which is composed of the user latent factor vector q u and the user auxiliary feature vector y u spliced ​​together, ⊙ represents the multiplication of the corresponding elements of the vector, represents the weight vector, σ(·) represents the activation function Sigmoid, represents the predicted matching score of user u for item i.

[0034] Furthermore, the formula of the loss function is as follows:

[0035]

[0036] In the above formula, Θ represents the parameter to be adjusted, represents a positive sample, Represents a randomly selected negative sample.

[0037] The second technical solution adopted by the present invention is: an item recommendation system based on implicit feedback, comprising:

[0038] A construction module, used for constructing an information matrix and an information network based on information between users and items and information between items;

[0039] A learning module, used for learning a user latent factor vector, an item latent factor vector, an item auxiliary feature vector, and a user auxiliary feature vector according to an information matrix and an information network;

[0040] The prediction module is used to construct a prediction vector based on the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector and calculate the prediction matching score.

[0041] The third technical solution adopted by the present invention is: an item recommendation device based on implicit feedback, comprising:

[0042] at least one processor;

[0043] at least one memory for storing at least one program;

[0044] When the at least one program is executed by the at least one processor, the at least one processor implements the item recommendation method based on implicit feedback as described above.

[0045] The beneficial effects of the method, system and device of the present invention are as follows: the present invention obtains latent factor vectors and feature vectors from information matrices and information networks, uses implicit feedback as interactive information, and has a huge amount of data that is easy to obtain. Finally, the final prediction vector is obtained by concatenating the latent factor vectors and auxiliary feature vectors, and item recommendations are implemented simply and accurately. The item prediction vector and the user prediction vector are then matched to obtain a prediction matching score, and the item recommendation information is scored to further ensure the accuracy of the item recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the steps of an item recommendation method based on implicit feedback of the present invention;

[0047] Figure 2 It is a schematic diagram of the framework of a specific embodiment of the present invention;

[0048] Figure 3 It is a schematic diagram of a deep neural network multi-layer perceptron learning an object latent factor vector in a specific embodiment of the present invention;

[0049] Figure 4 It is a schematic diagram of the principle of the attention network of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] Reference Figure 1 and Figure 2 The present invention provides an item recommendation method based on implicit feedback, which comprises the following steps:

[0052] Construct information matrix and information network based on the information between users and items and the information between items;

[0053] Learn the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector according to the information matrix and information network;

[0054] Construct a prediction vector based on the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector and calculate the prediction matching score;

[0055] Finally, recommended items are generated based on the matching scores.

[0056] As a further preferred embodiment of the method, it also includes:

[0057] Construct the loss function and perform parameter updates.

[0058] As a further preferred embodiment of the method, the step of constructing an information matrix and an information network based on the information between the user and the item and the information between the items specifically includes:

[0059] Construct an interactive information matrix based on the interactive information between users and items;

[0060] Specifically, in the recommendation scenario, assuming there are m users and n items, the interaction information between users and items is generally represented as a user-item interaction matrix

[0061]

[0062] The present invention uses implicit feedback information as the interaction information between users and items. Implicit feedback refers to the user's browsing, clicking, and collecting of items. It is generally Boolean in type, has a huge amount of data, avoids the user's operating burden, and is easy to obtain.

[0063] Construct a generation relationship matrix based on the generation information of users and items;

[0064] Specifically, for m users and n items, the “user-generated item” information is represented as a user-item generation relationship matrix

[0065]

[0066] In the "user-generated item" platform, an item is often generated by only one user, and one user can generate multiple items. Therefore, each column of the user-item generation relationship matrix G has only one non-zero element, which is very sparse.

[0067] Construct a co-generation network based on the generation information of items.

[0068] Specifically, it is known that the user-item generation matrix is very sparse, and it is difficult to effectively use the value of "user-generated items" information directly using this matrix. However, because multiple items may be generated by the same user, the co-generation relationship between items is relatively dense. Therefore, the item-item co-generation network is defined as in represents a set of nodes, each node represents an item, then ε represents a set of edges. If item i and item j are generated by the same user, there will be an edge between node i and node j.

[0069] As a further preferred embodiment of the method, the step of learning the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector according to the information matrix and the information network specifically includes:

[0070] The interactive information matrix is ​​processed based on a deep neural network multi-layer perceptron to learn the user latent factor vector and the item latent factor vector;

[0071] Specifically, the schematic diagram of learning the latent factor vector of an item is shown in Figure 3 , and the user latent factor vector can be obtained in the same way.

[0072] The co-generation network is processed based on the graph attention network to learn the auxiliary feature vector of the item;

[0073] Specifically, refer to Figure 4 , using the Graph Attention Network (GAT) to generate a network from items The auxiliary feature vector of the item is obtained by learning, that is, the auxiliary attribute vector of the item is converted into a higher-level auxiliary feature vector using the graph attention layer. The input of the graph attention layer is the auxiliary attribute vector of all items F = {F 1,: ,F 2,: ,...,F n,:}, represents the auxiliary attribute vector of item i; the output is the required auxiliary feature vector set X = {x1, x2, ..., x n}, Auxiliary feature vector representing item i.

[0074] Taking item i as an example, the specific calculation steps are as follows:

[0075] First, we generate a network for each item. The nodes in use the self-attention mechanism att: Calculate the attention value between two adjacent items:

[0076] e ij =att(F i,:W fx ,F j,: W fx )

[0077] in, Represents the weight matrix shared by all items, used for linear transformation; e ij is the attention coefficient, which indicates the importance of item j relative to item i. Specifically, the self-attention mechanism used in the present invention is a single-layer feedforward neural network, which contains a projection vector And an activation function LeakyReLU (λ = 0.2), that is:

[0078] e ij =LeakyReLU([F i,: W fx ||F j,: W fx ]a T )

[0079] Among them, || represents the concatenation operation.

[0080] Next, use the softmax function on the neighboring items of item i to obtain the normalized attention coefficient:

[0081]

[0082] in, Represents item i in the item-item co-generation network The set of neighbors in (including item i) is the set of all items generated by the generator u of item i.

[0083] Finally, the normalized attention coefficient is used to perform weighted averaging on the auxiliary attribute vectors of the neighboring items of item i, and the auxiliary feature vector of item i is obtained as follows:

[0084]

[0085] The generated relationship matrix is ​​combined with the item auxiliary feature vector to generate the user auxiliary feature vector.

[0086] Specifically, according to the user-item generation matrix G, the user auxiliary feature vector is obtained by combining the item auxiliary feature vectors:

[0087]

[0088] Among them, k u represents the number of items generated by user u, represents the subscript set of items generated by user u, Indicates item g u,k popularity, i.e., the popularity of item gu,k The number of users interacting, y u Represents the auxiliary feature vector of user u.

[0089] As a further preferred embodiment of the method, the latent factor vector of the object is learned, and the formula is expressed as:

[0090]

[0091]

[0092]

[0093] In the above formula, represents the interaction matrix of m users to n items, represents the interaction vector of item i, L is the number of layers of the deep neural network multilayer perceptron, are L weight matrices, are L bias vectors, is the L-1 hidden layer representation vector, f(·) represents the activation function, Represents the latent factor vector of item i learned.

[0094] As a further preferred embodiment of the method, the user latent factor vector is learned, and the formula is expressed as:

[0095]

[0096]

[0097]

[0098] In the above formula, represents the interaction vector of user u, f(·) represents the activation function, Represents the learned latent factor vector of user u.

[0099] As a further preferred embodiment of the method, the formula for constructing the prediction vector and calculating the prediction matching score is as follows:

[0100]

[0101]

[0102]

[0103] In the above formula, Represents the item prediction vector, which is composed of the item latent factor vector p i and item auxiliary feature vector x i Stitched together, Represents the user prediction vector, which is composed of the user latent factor vector q u and the user auxiliary feature vector y u spliced ​​together, ⊙ represents the multiplication of the corresponding elements of the vector, represents the weight vector, σ(·) represents the activation function Sigmoid, represents the predicted matching score of user u for item i.

[0104] The implicit feedback recommendation scenario is often defined as an interaction prediction problem, which suffers from the problem of lack of negative feedback. There are two common solutions: one is negative sampling; the other is to treat all non-interacted items as negative samples, which is obviously unreasonable, because the user's lack of interaction with the item does not mean that the user does not like the item, it is very likely that the user has never seen the item. Therefore, the present invention adopts a negative sampling mechanism to alleviate this problem. For each positive sample, ρ negative samples are randomly selected for training. Assume that the user-item interaction sample R ui Following the Bernoulli distribution, the likelihood function can be defined as follows:

[0105]

[0106] in, represents a positive sample, represents a randomly selected negative sample.

[0107] Finally, take the negative logarithm of the likelihood function and get the loss function as follows:

[0108]

[0109] In the above formula, Θ represents the parameter to be adjusted.

[0110] An item recommendation system based on implicit feedback, comprising:

[0111] A construction module, used for constructing an information matrix and an information network based on information between users and items and information between items;

[0112] The learning module is used to learn the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector based on the information matrix and information network.

[0113] The prediction module is used to construct a prediction vector based on the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector and calculate the prediction matching score.

[0114] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0115] An item recommendation device based on implicit feedback:

[0116] at least one processor;

[0117] at least one memory for storing at least one program;

[0118] When the at least one program is executed by the at least one processor, the at least one processor implements the item recommendation method based on implicit feedback as described above.

[0119] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0120] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. An item recommendation method based on implicit feedback, characterized in that: The following steps are involved: Construct information matrix and information network based on the information between users and items and the information between items; Learn the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector according to the information matrix and information network; Construct a prediction vector based on the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector and calculate the prediction matching score; The step of constructing an information matrix and an information network based on the information between users and items and the information between items specifically includes: Construct an interactive information matrix based on the interactive information between users and items; Construct a generation relationship matrix based on the generation information of users and items; Construct a co-generation network based on the generation information of items; The step of learning the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector according to the information matrix and the information network specifically includes: The interactive information matrix is ​​processed based on a deep neural network multi-layer perceptron to learn the user latent factor vector and the item latent factor vector; The co-generation network is processed based on the graph attention network to learn the auxiliary feature vector of the item; Combine the generated relationship matrix with the item auxiliary feature vector to generate the user auxiliary feature vector; The formula for learning the latent factor vector of an item is: In the above formula, represents the interaction matrix of m users to n items, represents the interaction vector of item i, L is the number of layers of MLP, are L weight matrices, are L bias vectors, is the L-1 hidden layer representation vector, f(·) represents the activation function, Represents the latent factor vector of the learned item i; Learning user latent factor vector, the formula is expressed as: In the above formula, represents the interaction vector of user u, f(·) represents the activation function, Represents the learned latent factor vector of user u.

2. The item recommendation method based on implicit feedback according to claim 1, characterized in that: Also includes: Construct the loss function and perform parameter updates.

3. The item recommendation method based on implicit feedback according to claim 2, characterized in that: The formula for constructing the prediction vector and calculating the prediction matching score is as follows: In the above formula, Represents the item prediction vector, which is composed of the item latent factor vector p i and item auxiliary feature vector x i Stitched together, Represents the user prediction vector, which is composed of the user latent factor vector q u and the user auxiliary feature vector y u spliced ​​together, ⊙ represents the multiplication of the corresponding elements of the vector, represents the weight vector, σ(·) represents the activation function Sigmoid, represents the predicted matching score of user u for item i.

4. The item recommendation method based on implicit feedback according to claim 3, characterized in that: The formula of the loss function is as follows: In the above formula, Θ represents the parameter to be adjusted, represents a positive sample, represents a randomly selected negative sample.

5. An item recommendation system based on implicit feedback, characterized in that: The method for recommending items based on implicit feedback as claimed in claim 1 comprises: A construction module, used for constructing an information matrix and an information network based on information between users and items and information between items; The learning module is used to learn the user latent factor vector, item latent factor vector, item auxiliary feature vector and user auxiliary feature vector based on the information matrix and information network. The prediction module is used to construct a prediction vector based on the user latent factor vector, the item latent factor vector, the item auxiliary feature vector and the user auxiliary feature vector and calculate the prediction matching score.

6. An item recommendation device based on implicit feedback, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the item recommendation method based on implicit feedback as described in any one of claims 1 to 4.

Images