A short video recommendation method based on relationship-aware multi-layer graph convolutional network

Abstract

The present invention discloses a short video recommendation method based on a relation-aware multi-layer graph convolutional network, which predicts the probability of a user clicking on a target short video based on the user's short video behavior sequence on a short video platform. The present invention is divided into six parts: the first part is to define the user's click behavior, and divide the user's historical exposure sequence into a click behavior sequence and a non-click behavior sequence; the second part is to use a graph structure learning method to build a multi-layer graph network; the third part is to use a relation-aware multi-layer graph convolutional network to update the item vector; the fourth part is to use an attention mechanism to obtain the user's positive interest vector and negative interest vector; the fifth part combines the user's positive interest vector and negative interest vector to obtain the user's final interest vector; the sixth part is to predict the user's click rate on short videos, design a loss function, and train model parameters.

Description

Technical Field

[0001] The present invention belongs to the field of Internet service technology, and in particular to a short video recommendation method based on a relationship-aware multi-layer graph convolutional network. Background Art

[0002] With the popularization of 5G technology, personal smartphones have gradually become the main battlefield for the general public to obtain information, and short video software represented by Douyin and Kuaishou has gradually become one of the essential software for most people's smartphones. The present invention aims to design a short video recommendation method to recommend short videos to users. Generally, short video platforms provide users with an immersive short video viewing experience, only playing one short video in full screen at a time, and the short video will be played automatically and repeatedly. The user can swipe to the next short video to stop the playback of the current short video. A large number of stock short videos have accumulated in the short video platform, and it is difficult for users to find their favorite short videos by searching. Therefore, the short video platform relies on the recommendation system to recommend short videos that users are interested in to users. The recommendation system can not only help users find content of interest, but also reduce the difficulty of using the APP and improve user-friendliness.

[0003] Existing short video recommendation systems usually model the user's historical positive feedback sequence, extract the user's positive interests from it, and then recommend short videos of interest to the user. This approach ignores the user's negative feedback behavior data. The negative feedback behavior here refers to the platform showing the user a short video, but the user finds that he is not interested in the content of the short video after watching it for a very short time and skips it. A small number of short video recommendation systems propose that user negative feedback data reflects the user's negative interests, that is, content that is not interesting. This part of the method uses the same network structure to model the user's negative feedback behavior and positive feedback behavior respectively, and obtains the user's negative interests and positive interests respectively, and then combines the two interests to recommend the next short video of interest to the user. They treat the user's positive feedback behavior and negative feedback behavior as two separate parts, ignoring the integrity of the user's behavior sequence and the connection between the user's positive feedback and negative feedback behaviors. According to data analysis, the short videos clicked by users and the short videos not clicked by users have certain similarities in content. In addition, users spend a long time on the short video platform and the user's behavior sequence is long, so the long-range dependency in the sequence is very important.

[0004] Therefore, the present invention proposes a short video recommendation method based on a relationship-aware multi-layer graph convolutional network. First, a graph structure learning method is adopted to construct a multi-layer graph network based on the user's historical positive feedback and negative feedback behavior sequences. Then a relationship-aware multi-layer graph convolutional network is proposed to update the short video vectors in the graph so that the short video vectors learn the complex dependencies of the short videos in the sequence. Finally, according to the updated short video vectors, an attention mechanism based on the target short video is adopted to obtain the user's positive and negative interests, and recommend short videos of interest to the user. The present invention conducted experiments on real data, and the experimental results show that the effect of this method is better than the existing methods. Summary of the invention

[0005] The problem of this method is defined as predicting the probability of a user clicking on a target short video based on the user's short video behavior sequence on the short video platform. The mathematical symbols involved are: the user set is U, the short video set is V, and the exposure sequence of all short videos displayed to the user by the platform is S = {(v 1 ,b 1 ),(v 2 ,b 2 ),…,(v j ,b j ),…,(v |S| ,b |S| )}, where (v j ,b j ) represents the user and the short video v j The interaction behavior is b j The user's interactive behavior includes click behavior and non-click behavior. When the user observes and clicks on the short video v j When b j = +, otherwise, when the user watches but does not click on the short video v j When b j =-. According to the user's interactive behavior, the user's short video exposure sequence can be further divided into a click behavior sequence and non-click behavior sequences Among them, m 1 and m 2are the lengths of click and non-click behavior sequences, respectively. Existing short video recommendation methods regard users’ multi-behavior sequences as fragmented parts, and use the same network structure to model users’ multi-behavior sequences separately, ignoring the associations between different behaviors. In addition, users spend a lot of time on short video platforms every day, so the behavior sequences are long, and there are long-range dependencies between short videos in the sequences. Therefore, the present invention constructs a multi-layer graph network from users’ multi-behavior sequences, and the multi-layer graph consists of several layer graphs and a coupling graph. The purpose of the layer graph is to capture the associations between the same behaviors, and the purpose of the coupling graph is to learn the associations between different behaviors. Based on the constructed multi-layer graph structure, an adaptive relationship-aware multi-layer graph convolutional network is designed. This method adopts the following technical solutions:

[0006] Define user click behavior and divide the user's historical exposure sequence into click behavior sequence and non-click behavior sequence. The user set is U, and the short video set is V. In order to introduce the present invention in a more general language, the professional term "item" is sometimes used to replace short video in this article. The exposure sequence of all short videos displayed to users by the platform is S = {(v 1 ,b 1 ),(v 2 ,b 2 ),…,(v j ,b j ),…,(v |S| ,b |S| )}, where (v j ,b j ) represents the user and the short video v j The interaction behavior is b j The user's interactive behavior includes click behavior and non-click behavior. When the user observes and clicks on the short video v j When b j =+, otherwise when the user watches but does not click on the short video v j When b j =-. To determine whether a user clicks on a short video, the following formula can be used:

[0007]

[0008] Among them, T watch Indicates the user’s viewing time of the short video, T video is the total duration of the short video, and c is a time threshold. click Indicates whether the user's interactive behavior on the short video is a click behavior. If it is a click behavior, then y click =1, otherwise y click = 0. According to this formula, the exposure sequence A can be divided into the click behavior sequence and non-click behavior sequences Among them, m1 and m 2 are the lengths of click and non-click behavior sequences, respectively. Click behavior is a positive feedback behavior of the user, and non-click behavior is a negative feedback behavior of the user.

[0009] Based on the user's historical click and non-click behavior sequence data, a multi-layer graph network is constructed using graph structure learning methods. The multi-layer graph network is represented as It is composed of several layer graphs and a coupling graph. Represents a set of nodes in a multi-layer graph, Presentation layer collection, is a super-adjacency matrix. The multi-layer graph network in the present invention includes two layers, the lower layer is the exposure layer, and the upper layer is the click layer. The exposure layer contains short videos in the exposure sequence of a given user, representing the user's observation behavior, and the click layer contains short videos in the user's click behavior sequence. For the convenience of representation, all short video nodes are included in the exposure layer and the click layer, and the unclicked short videos in the click layer are called virtual nodes, and there is no edge connection between the virtual node and any other node. The learning of a multi-layer graph network requires three steps: the first step is to construct a layer graph; the second step is to construct a coupling graph; the third step is to combine the layer graph and the coupling graph to form a multi-layer graph.

[0010] Each layer of the multi-layer graph network is an <item-item> graph network. in is a node in the layer graph, α represents the layer number, A α is the adjacency matrix that the model needs to learn. α Each element of Reflects the items v in layer α i and itemsv j The semantic relevance of . The adjacency matrix A α The learning process of is divided into two steps: the first step is to use multi-channel weighted cosine similarity to obtain a dense graph; the second step is to sparse the dense graph. The specific process is as follows:

[0011] Using multi-channel weighted cosine similarity to build an <item-item> graph network as follows:

[0012]

[0013] Among them, x i and x j Item v i and itemsv j The vector representation of w α,φ is a vector parameter, Indicates that in the α-layer φ channel, item v i and itemsv jThe correlation between them, n is the number of channels. Indicates that the average pooling is used to convert the n channels under Fusion to middle, Represents the final item v at layer α i and itemsv j The correlation between α Is a dense adjacency matrix, the elements in the matrix represent the graph network The edge weights of , need to be further sparse. α Set the negative values ​​to 0, and then transform the matrix M α The values ​​in are sorted in descending order, and the top K edges are retained to form the set ε top-K The value of K here is K=∈·(||M α || 0 ) 2 ,||·|| 0 represents the L0 norm, which is the matrix M α The number of non-zero values ​​in , and the parameter ∈ controls the density of the graph. The sparse adjacency matrix is ​​expressed as The specific mathematical calculation formula is:

[0014]

[0015] Further add self-connection to the sparse adjacency matrix In the adjacency matrix I N represents the identity matrix. Perform normalization:

[0016]

[0017] Among them, D α express The degree matrix of With A α The calculation process finally obtains the adjacency matrix of the layer graph network of the exposure layer and the click layer, which are represented as A 1 and A 2 .

[0018] Next, we construct the coupling graph network structure in the multi-layer graph. The coupling graph is represented by mathematical symbols as is the node, and C is the coupling matrix. Specifically, C αβ Indicates the connection relationship between the α layer and the β layer. The short video nodes in the exposure layer and the click layer in the multi-layer graph network are arranged in the same order, and the clicked nodes in the two layers are connected by edges. i Clicked, otherwise,

[0019] Each layer of the multi-layer graph network and coupling diagram Multi-layer graph network structure Expressed as in, is the set of all layer graphs, The symbol ∪ represents the graph connection operation. α Indicates that all layer graphs Connect them. Accordingly, the multi-layer graph network The super-adjacency matrix (supra-adjacencymatrix) It is expressed as:

[0020]

[0021] in, is a matrix concatenation operation.

[0022] Based on the multi-layer graph network, the relationship-aware multi-layer graph convolutional network is used to update the item vector. There are four relationships in the multi-layer graph constructed in the present invention: watch→click, click→watch, watch→watch, and click→click. Therefore, a relationship-aware multi-layer graph convolutional network is proposed to model these four relationships and update the item vector in the multi-layer graph. For node i in layer α, the vector update formula is as follows:

[0023]

[0024] in, and It is the set of neighbor nodes of node i within and between layers. and is the matrix parameter. is the vector representation of node i updated l times, initialized is the vector representation x of the corresponding short video i , φ is the activation function. After L times of convolution operation, we finally get the vector representation of the click sequence and the non-click sequence, which are expressed as and

[0025] Based on the user's click behavior sequence, non-click behavior sequence and target short video, the user's positive interest vector and negative interest vector are obtained using the attention mechanism. Given a target short video v new and the vector representation of the user's click sequence The user’s positive interest vector is obtained by using the attention mechanism based on the target short video. The formula is as follows:

[0026]

[0027] Among them, π i represents the importance of the i-th short video in the user's click sequence, W u is the matrix parameter, x new Is the target short video v new The vector representation of is the user’s positive interest vector. Similarly, the attention mechanism based on the target short video is used to represent the vector of the user’s unclicked sequence. Extract the user's negative interest vector

[0028] The user's positive interest vector and negative interest vector are combined in a weighted sum manner to obtain the user's final interest vector. The user's final interest vector is represented by e u , the calculation formula is as follows:

[0029]

[0030] Among them, λ p and λ n Control and Hyperparameters of importance.

[0031] Predict the user's click rate on the short video, design the loss function, and train the model parameters. new The vector representation of is connected and input into a two-layer MLP network to obtain the user's click probability on the short video.

[0032]

[0033] Among them, W 1 and W 2 is the matrix parameter, b 1 is a vector parameter, b 2 is a scalar parameter, σ is the sigmoid activation function, and φ is the activation function. Represents the probability of users clicking on short videos.

[0034] According to the characteristics of the model, the loss function is designed. The predicted value of the click rate of the target short video is calculated by the user. Calculate predicted values The error between the true value y and the error is then used to update the model parameters. The cross entropy loss function is used to guide the update process of the model parameters:

[0035]

[0036] Among them, y∈{0,1} is the real value, representing whether the user clicks on the target short video. The Adam optimizer is used to update the model parameters.

[0037] The beneficial technical effects of the present invention are as follows:

[0038] (1) This method innovatively proposes to construct a multi-layer graph network to represent the user's multi-behavior sequence, thereby capturing the long-range dependencies within the same behavior and between different behaviors from the user's multi-behavior sequence.

[0039] (2) This method innovatively proposes a relationship-aware multi-layer graph convolutional network to learn the topological structure of the constructed multi-layer graph network and distinguish the impact between different behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of a process of a short video recommendation method based on a relationship-aware multi-layer graph convolutional network according to the present invention;

[0041] Figure 2 A multi-layer graph network schematic diagram of a short video recommendation method based on a relationship-aware multi-layer graph convolutional network according to the present invention;

[0042] Figure 3 A schematic diagram of a hyper-adjacency matrix of a multi-layer graph network of a short video recommendation method based on a relationship-aware multi-layer graph convolutional network according to the present invention. DETAILED DESCRIPTION

[0043] In order to further understand the present invention, a short video recommendation method based on a relationship-aware multi-layer graph convolutional network provided by the present invention is described in detail below in combination with a specific implementation method. However, the present invention is not limited to this. Non-essential improvements and adjustments made by technical personnel in this field under the core guiding idea of ​​the present invention still fall within the scope of protection of the present invention.

[0044] The problem of this method is defined as predicting the probability of a user clicking on a target short video based on the user's short video behavior sequence on the short video platform. The mathematical symbols involved are: the user set is U, the short video set is V, and the exposure sequence of all short videos displayed to the user by the platform is S = {(v 1 ,b 1 ),(v 2 ,b 2 ),…,(v j ,b j ),…,(v |S| ,b |S| )}, where (v j ,b j ) represents the user and the short video v j The interaction behavior is b jThe user's interactive behavior includes click behavior and non-click behavior. When the user observes and clicks on the short video v j When b j = +, otherwise, when the user watches but does not click on the short video v j When b j =-. According to the user's interactive behavior, the user's short video exposure sequence can be further divided into a click behavior sequence and non-click behavior sequences Among them, m 1 and m 2 are the lengths of click and non-click behavior sequences, respectively. Click behavior is a positive feedback behavior of the user, and non-click behavior is a negative feedback behavior of the user. In order to introduce the present invention in a more general language, the professional term item is sometimes used in this article to replace micro-video.

[0045] A short video recommendation method based on a relation-aware multi-layer graph convolutional network consists of six parts. The first part is to define user click behavior and divide the user's historical exposure sequence into click behavior sequence and non-click behavior sequence; the second part is to build a multi-layer graph network based on the user's historical click and non-click behavior sequence data using a graph structure learning method; the third part is to update the item vector based on the multi-layer graph network using a relation-aware multi-layer graph convolutional network; the fourth part is to obtain the user's positive interest vector and negative interest vector based on the user's click behavior sequence, non-click behavior sequence and target short video using an attention mechanism; the fifth part is to combine the user's positive interest vector and negative interest vector using a weighted sum method to obtain the user's final interest vector; the sixth part is to predict the user's click rate on short videos, design a loss function, and train model parameters.

[0046] like Figure 1 As shown, according to one embodiment of the present invention, the method includes the following steps:

[0047] S100, define user click behavior, and divide the user's historical exposure sequence into click behavior sequence and non-click behavior sequence. The user set is U, the short video set is V, and the exposure sequence of all short videos displayed to the user by the platform is S = {(v 1 ,b 1 ),(v 2 ,b 2 ),…,(v j ,b j ),…,(v |S| ,b |S| )}, where (v j , b j ) represents the user and the short video v j The interaction behavior is bj The user's interactive behavior includes click behavior and non-click behavior. When the user observes and clicks on the short video v j When b j =+, otherwise when the user watches but does not click on the short video v j When b j =-. The user's short video exposure sequence refers to the short videos recommended by the platform to the user. After watching the short video for a short time and roughly understanding the content of the short video, the user can choose to continue watching or skip the short video and watch the next one. Therefore, the exposure sequence S can be divided into a click behavior sequence according to the user's interactive behavior and non-click behavior sequences Among them, m 1 and m 2 are the lengths of click and non-click behavior sequences respectively. Click behavior belongs to the user's positive feedback behavior, and non-click behavior belongs to the user's negative feedback behavior. Whether the user clicks on the short video needs to be calculated based on the user's viewing time data of the short video and the total duration data of the short video. The short video recommendation field is different from other traditional recommendation fields, such as product recommendations in e-commerce. There is no explicit click behavior in the short video recommendation field, because short videos are played automatically and do not require users to click explicitly. Therefore, it is necessary to define the click behavior based on the user's viewing time of the short video and the total duration of the short video. Given a threshold c, when the total duration of the short video is greater than or equal to the threshold c, it is defined as a click behavior when the user watches the short video for a duration greater than the threshold c. When the total duration of the short video is less than the threshold c, it is defined as a click behavior when the user watches the short video for a duration greater than the total duration of the short video. To determine whether a user clicks on a short video, the following formula can be used to summarize it:

[0048]

[0049] Among them, T watch Indicates the user’s viewing time of the short video, T video is the total duration of the short video, c is a time threshold, in this embodiment, c = 7 seconds. click Indicates whether the user's interactive behavior on the short video is a click behavior. If it is a click behavior, then y click =1, otherwise y click = 0. According to this formula, the user's historical click and non-click behavior sequence data can be obtained.

[0050] S200, based on the user's historical click and non-click behavior sequence data, a multi-layer graph network is constructed using a graph structure learning method. In order to capture the non-sequential dependencies within the same behavior and between different behaviors, a multi-layer graph network is constructed based on a graph structure learning method. The multi-layer graph network is represented as It is composed of several layer graphs and a coupling graph. Represents a set of nodes in a multi-layer graph, Presentation layer collection, is the supra-adjacency matrix. The multi-layer graph network in the present invention includes two layers, the lower layer is the exposure layer, and the upper layer is the click layer. The exposure layer contains the short videos in the exposure sequence of a given user, representing the user's observation behavior, and the click layer contains the short videos in the user's click behavior sequence. For convenience of representation, all short video nodes are included in the exposure layer and the click layer, and the unclicked short videos in the click layer are called virtual nodes. There is no edge connection between the virtual node and any other node. i ={(v 1 ,+),(v 2 ,-),(v ^ ,-),(v _ ,+),(v ` ,+)} as an example, the multi-layer graph network structure that may be learned is as follows Figure 2 As shown in Figure 2, learning a multi-layer graph network requires three steps: the first step is to build a layer graph; the second step is to build a coupling graph; the third step is to combine the layer graph and the coupling graph to form a multi-layer graph.

[0051] Each layer of the multi-layer graph network is an <item-item> graph network. in is a node in the layer graph, α represents the layer number, A α is the adjacency matrix that the model needs to learn. α Each element of Reflects the items v in layer α i and itemsv j The semantic relevance of the items is determined by multi-channel weight cosine similarity to construct the <item-item> graph network. as follows:

[0052]

[0053] Among them, x i and x j Item v i and itemsv j The vector representation of w α,φ is a vector parameter, Indicates that in the α-layer φ channel, item v iand itemsv j m is the number of channels, and in this embodiment, the number of channels n=4. Indicates that the average pooling is used to pool the n channels. Fusion to middle, Represents the final item v at layer α i and itemsv j The correlation between α Is a dense adjacency matrix, the elements in the matrix represent the graph network The weights of the edges of M need to be further sparsed. The reason is that the dense adjacency matrix has high computational complexity and contains a lot of noise. α Perform sparse processing. It is divided into two steps: the first step is to transform the matrix M α The negative values ​​are set to 0. This step is to retain only positive correlations. The second step is to transform the matrix M α The values ​​in are sorted in descending order, and the top K edges are retained to form the set ε top-K Here, the value of K is K=∈·(||M α || 0 ) 2 ,||·|| 0 represents the L0 norm, which is the matrix M α The number of non-zero values ​​in , the parameter ∈ controls the density of the graph. In this embodiment, ∈ = 0.6. The sparse adjacency matrix is ​​expressed as The specific mathematical calculation formula is:

[0054]

[0055] Further add self-connection to the sparse adjacency matrix In the adjacency matrix I N Represents the identity matrix. In order to avoid gradient explosion or gradient disappearance, the adjacency matrix Perform normalization:

[0056]

[0057] Among them, D α express The degree matrix of With A α The calculation process finally obtains the adjacency matrix of the layer graph network of the exposure layer and the click layer, which are represented as A 1 and A2 .

[0058] Next, we construct a coupling graph network structure in the multi-layer graph. The coupling graph enables information to be transferred between different layers. The coupling graph is represented by mathematical symbols as in, is a node, and C is a coupling matrix. Specifically, C αβ Indicates the connection relationship between the α layer and the β layer. The short video nodes in the exposure layer and the click layer in the multi-layer graph network are arranged in the same order, and the clicked nodes in the two layers are connected by edges. i Clicked, otherwise,

[0059] Each layer of the multi-layer graph network and coupling diagram Multi-layer graph network structure Expressed as in, is the set of all layer graphs, The symbol ∪ represents the graph connection operation. α Indicates that all layer graphs Connect them. Accordingly, the multi-layer graph network The super-adjacency matrix (supra-adjacencymatrix) It is expressed as:

[0060]

[0061] in, is a matrix concatenation operation. Figure 3 Shown Figure 2 The corresponding hyper-adjacency matrix of the multi-layer graph network

[0062] S300, based on a multi-layer graph network, uses a relationship-aware multi-layer graph convolutional network to update the item vector. The graph convolutional network is a process of information transmission, which uses the neighbor information of the nodes in the graph to update the node information. Common graph convolutional networks include GCN and LightGCN. However, these graph convolutional networks are only suitable for homogeneous ordinary single-layer graph structures, while the present invention constructs a multi-layer graph network (multiplex graph). Directly applying GCN or LightGCN in a multi-layer graph does not produce good experimental results. There are four types of relationships in the multi-layer graph constructed in the present invention: watch→click, click→watch, watch→watch, and click→click. Therefore, a relationship-aware multi-layer graph convolutional network is proposed to model these four relationships and update the item vectors in the multi-layer graph. For node i in the α layer, the vector update formula is as follows:

[0063]

[0064] in, and It is the set of neighbor nodes of node i within and between layers. and are matrix parameters. is the vector representation of node i updated l times, initialized is the vector representation x of the corresponding short video i φ is an activation function. In this embodiment, φ is a LeakyReLU activation function. After L convolution operations, the vector representations of the click sequence and the non-click sequence are finally obtained, which are expressed in mathematical formulas as and

[0065] S400, based on the user's click behavior sequence, non-click behavior sequence and target short video, the user's positive interest vector and negative interest vector are obtained using the attention mechanism. Given a target short video v new and the vector representation of the user's click sequence The user’s positive interest vector is obtained by using the attention mechanism based on the target short video. The formula is as follows:

[0066]

[0067] Among them, π i represents the importance of the i-th short video in the user's click sequence, W u is the matrix parameter, x new Is the target short video v new The vector representation of is the user’s positive interest vector. Similarly, the attention mechanism based on the target short video is used to represent the vector of the user’s unclicked sequence. Extract the user's negative interest vector

[0068] S500, using a weighted sum method, the user's positive interest vector and negative interest vector are combined to obtain the user's final interest vector. The user's final interest vector is represented by e u , the calculation formula is as follows:

[0069]

[0070] Among them, λ p and λ n Control and In this embodiment, λ p +λ n =1 and λ p =0.5.

[0071] S600: predict the click rate of the user on the short video, design the loss function, and train the model parameters. new The vector representation of is connected and input into a two-layer MLP network to obtain the user's click probability on the short video.

[0072]

[0073] Among them, W 1 and W 2 is the matrix parameter, b 1 is a vector parameter, b 2 is a scalar parameter, σ is a sigmoid activation function, and φ is an activation function. In this embodiment, φ is set to a RELU activation function. Represents the probability of users clicking on short videos.

[0074] According to the characteristics of the model, the loss function is designed. The predicted value of the click rate of the target short video is calculated by the user. Calculate predicted values The error between the true value y and the error is then used to update the model parameters. The cross entropy loss function is used to guide the update process of the model parameters:

[0075]

[0076] Among them, y∈{0,1} is the real value, representing whether the user clicks on the target short video. The Adam optimizer is used to update the model parameters.

[0077] The technical effect of the present invention is verified based on specific experimental data. For fairness, the experimental data is taken from the real public short video data MicroVideo-Anon. The sequence data of this data set is divided into a training set and a test set at a ratio of 9:1, and the training set is further divided into a training set and a validation set at a ratio of 9:1. In order to verify the performance effectiveness of this method, it is compared with the existing advanced model methods, including: ATRank, THACIL, SR-GNN, LSTM, DIEN, SURGE and CLSR. In order to evaluate the performance of each model, five widely used indicators are used: the area under the PR curve (AUC), normalized discounted cumulative gain (NDCG), precision, recall and F1 value. The model calculates the recommendation list, intercepts the TOP-N short videos, recommends them to users, and calculates the corresponding NDCG, precision, recall and F1 value. Among them, according to the scale of this embodiment, N takes a value of 20.

[0078] Table 1

[0079]

[0080] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made to the present invention by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. A short video recommendation method based on relationship-aware multi-layer graph convolutional network, characterized in that The following steps are involved: S1. Define user click behavior and divide the user's historical exposure sequence into click behavior sequence and non-click behavior sequence; S2. Based on the user's historical click and non-click behavior sequence data, a multi-layer graph network is constructed using graph structure learning methods; The step S2 further comprises: S21. Construct a layer graph in a multi-layer graph network. The multi-layer graph network consists of two layers, the lower layer is the exposure layer, and the upper layer is the click layer. The exposure layer contains short videos in the exposure sequence of a given user, representing the user's observation behavior, and the click layer contains short videos in the user's click behavior sequence. The layer graph of each layer of the multi-layer graph network is an <item-item> graph network. in is a node in the layer graph, α represents the layer number, A α is the adjacency matrix that the model needs to learn; the adjacency matrix A α The learning process is as follows: S211, use multi-channel weighted cosine similarity to obtain a dense graph; use multi-channel weighted cosine similarity to construct an <item-item> graph network as follows: Among them, x i and x j Item v i and itemsv j The vector representation of w α,φ is a vector parameter, Indicates the item v in the α-layer φ channel i and itemsv j The correlation between them, n is the number of channels; Represents the final item v at layer α i and itemsv j The correlation between S212, for dense graphs Sparse α is a dense adjacency matrix, which needs to be further sparsely processed; first, the matrix M α Set the negative values ​​to 0, and then transform the matrix M α The values ​​in are sorted in descending order, and the top K edges are retained to form the set ε top-K ; Here K is K=∈·(||M α ||1) 2 , ||·||1 represents the L0 norm, which is the matrix M α The number of non-zero values ​​in , the parameter ∈ controls the density of the graph; the sparse adjacency matrix is ​​expressed as The specific mathematical calculation formula is: S213, further adding self-connection to the sparse adjacency matrix In the adjacency matrix I N represents the identity matrix; for the adjacency matrix Perform normalization: Among them, D α express The degree matrix of According to A α The calculation process finally obtains the adjacency matrix of the layer graph network of the exposure layer and the click layer, which are represented as A 1 and A 2 ; S22. Construct a coupling graph network structure in a multi-layer graph; the coupling graph is represented by mathematical symbols as is a node, and C is a coupling matrix; specifically, C αβ Indicates the connection relationship between the α layer and the β layer; the short video nodes in the exposure layer and the click layer in the multi-layer graph network are arranged in the same order, and the clicked nodes in the two layers are connected by edges; if the short video v i Clicked, otherwise, S23. Each layer of the graph based on the multi-layer graph network and coupling diagram Multi-layer graph network structure Expressed as in, is the set of all layer graphs, The symbol ∪ represents the graph connection operation. α Indicates that all layer graphs Connecting it up; multi-layer graph networks The super-adjacency matrix of It is expressed as: in, is a matrix concatenation operation; S3, based on the multi-layer graph network, using the relationship-aware multi-layer graph convolutional network to update the item vector; The step S3 further comprises: S31. Use the relationship-aware multi-layer graph convolutional network to update the item vectors in the multi-layer graph network. For node i in the α layer, the vector update formula is as follows: in, and is the set of neighbor nodes of node i within and between layers; and is the matrix parameter; is the vector representation of node i updated l times, initialized is the vector representation x of the corresponding short video i , φ is the activation function; through L times of convolution operation, we finally get the vector representation of the click sequence and the non-click sequence, which are expressed in mathematical formulas as and S4. Based on the user's click behavior sequence, non-click behavior sequence and target short video, the user's positive interest vector and negative interest vector are obtained using the attention mechanism; S5, combining the user's positive interest vector and negative interest vector to obtain the user's final interest vector; S6. Predict the click rate of users on short videos, design loss functions, and train model parameters.

2. According to claim 1, a short video recommendation method based on a relationship-aware multi-layer graph convolutional network is characterized in that The calculation method of step S1 is as follows: S11. Determine whether the user clicks on a short video: Among them, T watch Indicates the user's viewing time of the short video, T video is the total duration of the short video, c is a time threshold; y click Indicates whether the user's interactive behavior on the short video is a click behavior. If it is a click behavior, then y click =1, otherwise y click =0; According to this formula, the exposure sequence is divided into a click behavior sequence and non-click behavior sequences Among them, m1 and m2 are the lengths of click and non-click behavior sequences respectively.

3. According to claim 1, a short video recommendation method based on a relationship-aware multi-layer graph convolutional network is characterized in that The calculation method of step S4 is as follows: S41. Use the attention mechanism based on the target short video to extract the user's positive interest vector from the user's click sequence: Among them, π i represents the importance of the i-th short video in the user's click sequence, W u is the matrix parameter, x new Is the target short video v new The vector representation of is the user’s positive interest vector; S42. Using the same method, extract the user's negative interest vector from the user's non-click sequence 4. According to claim 1, a short video recommendation method based on a relationship-aware multi-layer graph convolutional network is characterized in that The calculation method of step S5 is as follows: S51, using a weighted sum method, combining the user's positive interest vector and negative interest vector to obtain the user's final interest vector e u : Among them, λ p and λ n Control and Hyperparameters of importance.

5. According to claim 1, a short video recommendation method based on a relationship-aware multi-layer graph convolutional network is characterized in that The calculation method of step S6 is as follows: S61, the user's final interest vector and the target short video v new The vector representation of the data is connected and input into a two-layer MLP network to obtain the probability of users clicking on the short video. Among them, W1 and W2 are matrix parameters, b1 is a vector parameter, b2 is a scalar parameter, σ is the sigmoid activation function, and φ is the activation function; Represents the probability of users clicking on short videos; S62. Design a loss function based on the model characteristics; predict the click rate of the target short video by the user Calculate predicted values The error between y and the true value y is used to update the model parameters. The cross entropy loss function is used to guide the update process of the model parameters: Among them, y∈{0,1} is the true value, indicating whether the user clicks on the target short video; the Adam optimizer is used to update the model parameters.

Images