A Recommendation Method Based on Click Sequence Time Enhancement for Anonymous Login Scenarios

Abstract

The present invention discloses a recommendation method based on click sequence time enhancement for anonymous login scenarios, belonging to the field of recommendation, including data preprocessing; constructing a session graph with enhanced time information; generating hidden vectors of nodes on the session graph; generating user local interest representations; arranging the hidden vectors generated by the graph neural network according to the clicks during the session as the input of the BiLSTM, using a variable-length sequence processing method to process unequal-length sequences, and finally obtaining a vector containing order-dependent information. The node features are randomly masked through the Dropout algorithm to reduce the impact of noisy data, and finally the global interest of the anonymous user in the session sequence is generated through the Attention layer; generating the final interest representation of the user and performing regression prediction. The present invention integrates the user's interaction time information, enhances the long-term and short-term dependencies between nodes in the sequence, further explores the potential interests of users, and can effectively improve the recommendation effect in anonymous scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a recommendation method based on click sequence time enhancement applied to the anonymous login scenario. Background Art

[0002] With the advent of the big data era, it has become easy for people to obtain information on the Internet, but "information overload" has also followed. Recommendation systems can process a large amount of interaction data between users and items to help people filter information for personalized recommendations. Session recommendation is a subfield of recommendation systems. A session refers to an interaction process between a user and an item. When a user logs in anonymously on a website, the recommendation system cannot capture the user's personal information and browsing history information, and only predicts the user's next click operation based on the user's current session interaction information. Because there are more and more studies on graph neural networks and good results have been achieved in many tasks, many researchers have incorporated graph neural networks into the construction of session recommendation models to represent the complex dependencies within a session or between various sessions using a graph structure, and have also obtained good results. Session-based Graph Neural Network Recommendation (SR-GNN) is a pioneering work that introduces graph neural network research into session-based recommendation for predicting the behavioral interest preferences of anonymous users. The session list is modeled into graph-structured data, and the transition relationship of session items is captured through a gated graph neural network, where an attention network is used to combine the user's global interest preference and the current interest preference into a session sequence. Although SR-GNN can well express the high-order and low-order dependencies between entities in the sequence by representing sequence data through a graph structure, it weakens the long-term and short-term dependencies between sequences. In addition, SR-GNN does not consider the time interval information of node transitions in the sequence, which does not conform to reality because the user's preference can be reflected by the length of the time interval in the current session.

[0003] Therefore, it is urgent to develop a recommendation method based on click sequence time enhancement applied to the anonymous login scenario to solve the above problems. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a recommendation method based on click sequence time enhancement applied to the anonymous login scenario, which integrates the user's interaction time information, enhances the long-term and short-term dependencies between nodes in the sequence, further explores the user's potential interests, and can effectively improve the recommendation effect in the anonymous scenario.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A recommendation method based on click sequence time enhancement applied to the anonymous login scenario, comprising the following steps:

[0007] S1. Data preprocessing: Extract the click sequence information in the historical interactions of anonymous logged-in users and the time information of the interactions between users and items, and construct a directed session graph that reflects the order of nodes. The session graph is represented by out-degree and in-degree matrices A = [A out , A in , and the values in the matrix are the corresponding time interval information;

[0008] S2. Construct a session graph with enhanced time information: Consider the out-edges of each node on the graph. The initial weight on the edge is the time interval of the transfer between adjacent nodes in the click sequence. Then, divide each time weight on the out-edge by the sum of the out-edge weights of the node for normalization. The same processing method is applied to the in-edge weights of each node, so as to obtain a session graph with enhanced time intervals of the user click sequence;

[0009] S3. Generate the hidden vectors of the nodes on the session graph: Use the gated graph neural network algorithm to propagate and aggregate the information on the session graph to generate the hidden vectors of the nodes;

[0010] S4. Generate the local interest representation of the user: The hidden vector of the last node in the user click sequence is used as the current local interest of the user;

[0011] S5. Generate the global interest representation of the user: Arrange the hidden vectors generated by the graph neural network according to the clicks during the session as the input of the BiLSTM. Use the variable-length sequence processing method to process the unequal-length sequences, and finally obtain a vector containing order-dependent information. Randomly mask the node features through the Dropout algorithm to reduce the influence of noisy data. Finally, generate the global interest of the anonymous user in this session sequence through the Attention layer;

[0012] S6. Generate the final interest representation of the user: Concatenate the local interest and global interest of the user, and use a fully connected neural network to obtain the final interest representation of the user;

[0013] S7. Regression prediction: Calculate the similarity between the final interest representation of the user and the hidden vectors of all nodes, and select the node with the highest similarity degree for recommendation, which is used as the prediction of the user's next click.

[0014] A further improvement of the technical solution of the present invention lies in: In S2, the following steps are specifically included:

[0015] 2.1. In the session graph G constructed according to the session sequence s, there is only one type of node v i ∈V s and one type of directed edge E(v i , v i+1 )∈E s , and the directed edges on the graph reflect the order relationship between adjacent nodes in the sequence. The initial weight on the edge It reflects the user's stay time, which is specifically expressed as follows:

[0016] s = {v1, v2, …, v t} (1)

[0017] G s = (V s , E s ) (2)

[0018] v i ∈V s , E(v i , v i+1 ) ∈ E s (3)

[0019]

[0020] 2.2. For each node v on the graph i respectively count the weight information of all its incoming edges outgoing edges and sum them up. Then divide the initial weights of all the edges of this node by the corresponding incoming edge weight sum outgoing edge weight sum for normalization to obtain the final weight representation. This process is described as:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] Thus, a session graph with enhanced user click sequence time intervals after normalization is obtained.

[0028] A further improvement of the technical solution of the present invention lies in: In S3, it specifically includes the following steps:

[0029] 3.1. Use a gated neural network to model the nodes in the session graph, thereby generating the feature vectors of the nodes. Take the session sequence s = {v s,1 , v s,2 ,..., v s,n} as the input. The gated neural network calculates the item node v in the session graph G s ins,i The information transfer formula between them is as follows:

[0030]

[0031] In the formula, refers to the result of obtaining the information of the adjacent neighbor nodes of the item node v at time t in the session sequence s, which includes the mutual information propagation relationship between adjacent nodes. s,i H ∈ R is the weight matrix calculated according to the session graph, A d×2d ∈ R s The connection matrix of the session graph s represents the dependency conversion relationship between each node in the session graph, A n×2n ∈ R s,i: represents the two-column block matrix corresponding to the item node v in the connection matrix A 1×2n s s in the item node v s,i corresponding two-column block matrix; represents the feature vector of the i-th item node in the session sequence s at time t - 1, and b is the bias value;

[0032] 3.2. Generate the feature vector representation of the nodes in the session graph:

[0033] For each session graph, all the nodes in the session graph are input into the GNN for training and learning. After the learning calculation of formula (6), all the item nodes in the session graph have propagated information with their adjacent neighbor nodes; the information update and iteration formula of the item nodes using GRU is as follows:

[0034]

[0035]

[0036]

[0037]

[0038] In formulas (7) to (10), and respectively represent the update gate and reset gate in GRU, σ represents the sigmoid function, and ⊙ is the matrix multiplication operation. refers to the newly obtained intermediate layer information state after learning and training. represents the latest potential feature information of the item node generated at the next time t.

[0039] A further improvement of the technical solution of the present invention lies in: In S5, it specifically includes the following steps:

[0040] 5.1. Construct the input of the bidirectional long short-term memory network (BiLSTM) in the order of the original sequence

[0041]

[0042] 5.2. For Use BiLSTM to learn the order information of the evolution of nodes in the sequence and generate new vectors The specific formula is as follows:

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] In the forget gate f t A σ∈(0,1) will be generated based on the information obtained at the previous time and the current input information to determine whether the information at the previous moment is retained and passed; i t Represents the input gate. First, it controls which information needs to be updated, and then a new value obtained through the tanh layer is used, combined with f t and i t To obtain a new memory cell C t ; Finally, the output gate O t Determines the final output of the model and outputs the hidden node

[0050] 5.3. Use dropout to randomly mask node features, and finally use the attention mechanism to aggregate the nodes in the sequence, so that the important features in the sequence are retained, the interference of unimportant features is reduced, and finally the global interest representation S of the user is generated g , the specific formula is as follows:

[0051]

[0052]

[0053]

[0054] All item nodes in it have passed through the BiLSTM layer to obtain rich context information and order information in the sequence, and then input into the attention mechanism layer for aggregation to obtain a more important item feature representation;

[0055] In the formula, W1, W2, q, and c are learning parameters of the feature vector weights, and W1, W2 ∈ R d×d , q, c ∈ R d .

[0056] A further improvement of the technical solution of the present invention lies in: In S6, it specifically includes the following steps:

[0057] 6.1. Concatenate the global feature S g of the anonymous logged-in user in the sequence and the hidden vector S l of the last clicked node, and then obtain the mixed embedding input S s of the MLP layer:

[0058] S s = concat(S l , S g ) (26)

[0059] 6.2. Input S s into the MLP layer to generate the final interest representation S f of this anonymous user:

[0060] S f = ReLU(W′S s + b′) (27)

[0061] In the formula, W′ ∈ R 2d*d is a learnable parameter, b′ is a bias term, and ReLU is used as the activation layer.

[0062] A further improvement of the technical solution of the present invention lies in: In S7, it specifically includes the following steps:

[0063] 7.1. Calculate the similarity between the final interest representation S f of the anonymous user and the item node and normalize it using softmax. The specific formula is as follows:

[0064]

[0065]

[0066] where V i is the node feature vector generated by the graph neural network;

[0067] 7.2. Calculate the similarity between the true value and the predicted value using cross - entropy. The calculation formula is as follows:

[0068]

[0069] In the formula, y i represents the true value label, which is the probability that the item will be clicked next under the actual click result of the user, and is represented in the form of a one - hot encoded vector; is the probability that the item is predicted. When the loss value is smaller, it means that the predicted value is closer to the true value. Conversely the larger it is.

[0070] Due to the adoption of the above - mentioned technical solution, the technical progress achieved by the present invention is as follows:

[0071] 1. In the implementation process of the present invention, the weights on the graph are adjusted through the time intervals in the sequence, and the gated graph neural network is used to propagate and aggregate the information on the graph to generate node hidden vectors that integrate time information.

[0072] 2. The present invention inputs the hidden vectors of the sequence nodes into a bidirectional long - short - term memory network for sequence learning, so that the generated hidden vectors integrate the long - term and short - term dependence information of the sequence.

[0073] 3. The present invention uses dropout and attention mechanisms to obtain the global interest representation of the user. By concatenating it with the local interest representation as the input of the fully - connected neural network, the final representation of the user is obtained, and the similarity is calculated with the node hidden vector as the metric for recommendation. Compared with a variety of the latest methods, this method has a competitive advantage. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings;

[0075] Figure 1 is the flowchart of the recommendation method based on click - sequence time enhancement applied to the anonymous login scenario provided by the present invention;

[0076] Figure 2 is the algorithm model framework diagram in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] It should be noted that the terms "including" and "having" in the description and claims of the present invention and any variations thereof in the above-mentioned drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0078] An embodiment of the present invention provides a recommendation method based on click sequence time enhancement for anonymous login scenarios, which solves the problem in the prior art that "SR-GNN weakens the long-term and short-term dependencies between sequences, and the time interval information of node transitions in SR-GNN does not conform to reality". The general idea is as follows: By extracting the click sequence information in the historical interactions of anonymous login users and the time information of the interaction between users and items, a directed session graph reflecting the order of nodes is constructed. The session graph is represented by out-degree and in-degree matrices A = [A out , A in . The values in the matrix are the corresponding time interval information. Considering the out-edges of each node on the graph, the initial weight on the edge is the time interval of the adjacent node transition in the click sequence. Normalize each time weight on the out-edge by dividing it by the sum of the out-edge weights of the node. The same method is used for processing the in-edge weights of each node, so as to obtain a session graph with enhanced click sequence time of the user, and update the out-degree and in-degree matrices A. Then, use the gated graph neural network algorithm to propagate and aggregate the information on the session graph to generate the hidden vector of the node. Arrange the hidden vectors generated by the graph neural network according to the clicks during the session as the input of BiLSTM, and use the variable-length sequence processing method to process unequal-length sequences. Finally, a vector containing the order long-term and short-term dependence information is obtained. Randomly mask the node features through the Dropout algorithm to reduce the influence of noise data. Finally, generate the global interest of the anonymous user in the session sequence through the Attention layer. Then, use the hidden vector of the last node in the user click sequence as the user's current local interest and splice it with the global interest as the input of the fully connected neural network. Finally, obtain the interest representation of the user. Calculate the similarity between the final interest representation of the user and the hidden vectors of all nodes, and select the node with the highest similarity degree for recommendation, which is used as the prediction of the user's next click.

[0079] The following further describes the present invention in detail with reference to the accompanying drawings and embodiments:

[0080] As Figure 1 , 2 shown, the recommendation method based on click sequence time enhancement for anonymous login scenarios provided in this embodiment includes the following steps:

[0081] S1. Data preprocessing: Extract the click sequence information in the historical interactions of anonymous logged-in users and the time information of the interactions between users and items, and construct a directed session graph that reflects the order of nodes. The session graph is represented by out-degree and in-degree matrices A = [A out , A in , and the values in the matrix represent the corresponding time interval information;

[0082] S2. Construct a session graph with enhanced time information: Consider the out-edges of each node on the graph. The initial weight on the edge is the time interval of the transfer between adjacent nodes in the click sequence. Then, divide each time weight on the out-edge by the sum of the out-edge weights of the node for normalization. The same method is used for the in-edge weights of each node, so as to obtain a session graph with enhanced time information of the user click sequence; specifically, it includes the following steps:

[0083] 2.1. In the session graph G constructed according to the session sequence s, there is only one type of node v i ∈V s and one type of directed edge E(v i , v i+1 )∈E s . The directed edges on the graph reflect the order relationship between adjacent nodes in the sequence, and the initial weight on the edge reflects the user's stay time, which is specifically represented as follows:

[0084] s = {v1, v2,..., v t} (1)

[0085] G s = (V s , E s ) (2)

[0086] v i ∈V s , E(v i , v i+1 )∈E s (3)

[0087]

[0088] 2.2. For each node v i on the graph, respectively count the weight information of all its in-edges and out-edges and sum them up. Then, divide the initial weights of all the edges of the node by the corresponding in-edge weight sum and out-edge weight sum for normalization to obtain the final weight representation. This process can be described as:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] Thus, a normalized session graph with enhanced user click sequence time is obtained;

[0096] S3. Generate the hidden vectors of the nodes on the session graph: Use the gated graph neural network algorithm to propagate and aggregate the information on the session graph to generate the hidden vectors of the nodes; specifically, it includes the following steps:

[0097] 3.1. Use a gated neural network (GGNN) to model the nodes in the session graph, thereby generating the feature vectors of the nodes. Take the session sequence s = {v s,1 , v s,2 ,..., v s,n} as the input. GGNN calculates the information transfer formula between the item nodes v s in the session graph G s,i as follows:

[0098]

[0099] In the formula, refers to the result of obtaining the information of the neighboring nodes adjacent to the item node v s,i at the t-th moment in the session sequence s, which includes the mutual information propagation relationship between adjacent nodes. H ∈ R d×2d is the weight matrix calculated according to the session graph, and A s ∈ R n×2n The connection matrix of the session graph s represents the dependency conversion relationship between the nodes in the session graph. A s,i: ∈ R 1×2n represents the two-column block matrix corresponding to the item node v s in the connection matrix A s,i ; represents the feature vector of the i-th item node in the session sequence s at the (t - 1)-th moment, and b is the bias value.

[0100] 3.2. Generate the feature vector representation of the nodes in the session graph:

[0101] For each session graph, all the nodes in the session graph are input into the GNN for training and learning. After the learning calculation of formula (6), information propagation occurs between all the item nodes in the session graph and their adjacent neighbor nodes. The information update and iteration formula for the item nodes using GRU is as follows:

[0102]

[0103]

[0104]

[0105]

[0106] In formulas (7) to (10), and represent the update gate and the reset gate in GRU respectively, σ represents the sigmoid function, ⊙ represents the matrix multiplication operation, refers to the newly obtained intermediate layer information state after learning and training, represents the latest latent feature information of the item node generated at the next moment t.

[0107] S4. Generate the user's local interest representation: The hidden vector of the last node in the user click sequence is used as the user's current local interest;

[0108] S5. Generate the user's global interest representation: The hidden vectors generated by the graph neural network are arranged according to the clicks during the session as the input of the BiLSTM. The variable-length sequence processing method is used to process the unequal-length sequences. Finally, a vector containing order-dependent information is obtained. The node features are randomly masked through the Dropout algorithm to reduce the influence of noisy data. Finally, the global interest of the anonymous user in this session sequence is generated through the Attention layer. Specifically, it includes the following steps:

[0109] 5.1. Construct the input of the bidirectional long short-term memory network (BiLSTM) according to the order of the original sequence

[0110]

[0111] 5.2. For use BiLSTM to learn the evolving order information of the nodes in the sequence and generate a new vector The specific formula is as follows:

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] In the forget gate f t a σ ∈ (0, 1) is generated based on the information obtained at the previous time and the current input information to determine whether the information at the previous moment is retained and transmitted; i t represents the input gate. First, it controls which information needs to be updated, and then a new value is obtained through the tanh layer. Combining f t and i t a new memory cell C is obtained t ; finally, the output gate O t determines the final output of the model, outputting the hidden node

[0119] 5.3. Randomly mask the node features using dropout, and finally use the attention mechanism to aggregate the nodes in the sequence, so that the important features in the sequence are retained, reducing the interference of unimportant features, and finally generating the global interest representation S of the user g , and the specific formula is as follows:

[0120]

[0121]

[0122]

[0123] All item nodes in obtain rich context information and order information in the sequence through the BiLSTM layer, and then d×d input into the attention mechanism layer for aggregation to obtain a more important item feature representation. Among them, W1, W2, q, and c are learning parameters of the feature vector weights, where W1, W2 ∈ R d .

[0124] S6. Generate the final interest representation of the user: Concatenate the local interest and global interest of the user, and use a fully connected neural network to obtain the final interest representation of the user; specifically, it includes the following steps:

[0125] 6.1. The global feature S of the anonymous logged-in user in the sequence g and the hidden vector S of the last clicked node lConcatenate them to obtain the mixed embedding input S of the MLP layer s :

[0126] S s = concat(S l , S g ) (26)

[0127] 6.2. Input S s into the MLP layer to generate the final interest representation S of the anonymous user f :

[0128] S f = ReLU(W′S s + b′) (27)

[0129] In the formula, W′ ∈ R 2d*d is a learnable parameter, b′ is a bias term, and ReLU is used as the activation layer;

[0130] S7. Regression prediction: Calculate the similarity between the final interest representation of the user and the hidden vectors of all nodes, and select the node with the highest similarity degree for recommendation, which is used as the prediction of the user's next click.

[0131] 7.1. Calculate the similarity between the final interest representation S f of the anonymous user and the item nodes and normalize it using softmax. The specific formula is as follows:

[0132]

[0133]

[0134] In the formula, V i is the node feature vector generated by the graph neural network;

[0135] 7.2. Use cross-entropy to calculate the similarity between the true value and the predicted value. The calculation formula is as follows:

[0136]

[0137] In the formula, y i represents the true value label, which is the probability that the item will be clicked next under the actual click result of the user, and is expressed in the form of a one-hot encoded vector; is the probability that the item is predicted. When the loss value is smaller, it means that the predicted value is closer to the true value, and vice versa is larger.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A recommendation method based on click sequence time enhancement for anonymous login scenarios, characterized in that Including the following steps: S1. Data preprocessing: Extract the click sequence information in the historical interaction of anonymous logged-in users and the time information of the interaction between users and items, and construct a directed conversation graph that reflects the order of nodes. The conversation graph is composed of the out-degree and in-degree matrices A = [A out ,A in ] indicates that the value in the matrix is the corresponding time interval information; S2. Construct a session graph with enhanced time information: Consider the out-edges of each node in the graph. The initial weight on the edge is the time interval between the transfer of adjacent nodes in the click sequence. Then, divide each time weight on the out-edge by the sum of the out-edge weights of the node for normalization. The same method is used for the in-edge weights of each node, so as to obtain a session graph with enhanced time intervals of the user click sequence; S3. Generate the hidden vectors of the nodes on the session graph: Use the gated graph neural network algorithm to propagate and aggregate the information on the session graph to generate the hidden vectors of the nodes; S4. Generate the local interest representation of the user: The hidden vector of the last node in the user click sequence is used as the current local interest of the user; S5. Generate the global interest representation of the user: Arrange the hidden vectors generated by the graph neural network according to the clicks during the session as the input of the BiLSTM. Use the variable-length sequence processing method to process the unequal-length sequences. Finally, obtain a vector containing order-dependent information. Randomly Mask the node features through the Dropout algorithm to reduce the influence of noisy data. Finally, generate the global interest of the anonymous user in the session sequence through the Attention layer; S6. Generate the final interest representation of the user: Concatenate the local interest and the global interest of the user, and use a fully connected neural network to obtain the final interest representation of the user; S7. Regression prediction: Calculate the similarity between the final interest representation of the user and the hidden vectors of all nodes, and select the node with the highest similarity degree for recommendation, which is used as the prediction of the user's next click.

2. The recommended method based on click sequence time enhancement applied to the anonymous login scenario according to claim 1, characterized in that In S2, it specifically includes the following steps: 2.

1. In the session graph G constructed according to the session sequence s, there is only one type of node v i ∈V s and one type of directed edge E(v i ,v i+1 )∈E s . The directed edges on the graph reflect the order relationship between adjacent nodes in the sequence, and the initial weights on the edges reflect the user's stay time, which is specifically represented as follows: s = {v1, v2, ..., v t} (1) G s = (V s , E s ) (2) v i ∈ V s such that, for (v i , v i+1 ) ∈ E s (3) 2.

2. For each node v on the graph i Respectively count all its incoming edges Outgoing edges The weight information of and sum them up, and then the initial weights on all the edges of this node Are respectively divided by the corresponding incoming edge weight sum Outgoing edge weight sum For normalization to obtain the final weight representation. This process is described as: Thus, a normalized session graph with enhanced time intervals of the user click sequence is obtained.

3. The recommendation method based on click sequence time enhancement applied to the anonymous login scenario according to claim 1, wherein In S3, it specifically includes the following steps: 3.

1. Model the nodes in the session graph using a gated neural network to generate the feature vectors of the nodes. Take the session sequence s = {v s，1 , v s，2 ,..., v s,n} as the input. The gated neural network calculates the information transfer formula between item nodes v s in the session graph G s,i as follows: In the formula, refers to the result of obtaining the information of adjacent neighbor nodes of the item node v s,i at the t-th moment in the session sequence s, which includes the mutual information propagation relationship between adjacent nodes. H ∈ R d×2d is the weight matrix calculated according to the session graph, and A s ∈ R n×2n The connection matrix of the session graph s represents the dependency conversion relationship between each node in the session graph, and A s,i : ∈ R 1×2n means the two-column block matrix corresponding to the item node v s in the connection matrix A s,i ; V i t-1 represents the feature vector of the i-th item node in the session sequence s at the (t - 1)-th moment, and b is the bias value. 3.

2. Generate the feature vector representation of the nodes in the session graph: For each session graph, input all the nodes in the session graph into the GNN for training and learning. After the learning calculation of formula (6), information propagation has occurred between all the item nodes in the session graph and their adjacent neighbor nodes; The information update iteration formula of the item nodes using GRU is as follows: In formulas (7) to (10), and represent the update gate and the reset gate in the GRU respectively, σ represents the sigmoid function, and ⊙ represents the matrix multiplication operation. refers to the newly obtained intermediate layer information state after learning and training, and V i t represents the latest latent feature information of the item node generated at the next moment t.

4. A recommendation method based on click sequence time enhancement for anonymous login scenarios according to claim 1, characterized in that In S5, specifically including the following steps: 5.

1. Construct the input of the bidirectional long short-term memory network (BiLSTM) in the order of the original sequence 5.

2. For Use BiLSTM to learn the order information of the evolution of nodes in the sequence and generate new vectors The specific formula is as follows: In the forget gate f t a σ ∈ (0, 1) is generated based on the information obtained at the previous time and the current input information to determine whether the information at the previous moment is retained and passed on; f t represents the input gate. First, it controls which information needs to be updated, and then uses the values obtained through the tanh layer to combine with f t and i t to obtain the new memory cell C t ; finally, the output gate O t determines the final output of the model and outputs the hidden nodes 5.

3. Randomly mask node features using dropout, and finally use the attention mechanism to aggregate nodes in the sequence, so that important features in the sequence are retained, the interference of unimportant features is reduced, and finally the global interest representation S of the user is generated g , and the specific formula is as follows: All item nodes in obtain rich context information and order information in the sequence through the BiLSTM layer, and then input them into the attention mechanism layer for aggregation to obtain a more important item feature representation; where W1, W2, q, and c are learning parameters of the eigenvector weights, and W1, W2 ∈ R d×d , q, c ∈ R d .

5. The recommendation method based on click sequence time enhancement for an anonymous login scenario according to claim 1, wherein In S6, it specifically includes the following steps: 6.

1. Concatenate the global features S of the anonymous logged-in user in the sequence g and the hidden vector S of the last clicked node l to obtain the mixed embedding input S of the MLP layer S : S s = concat(S l , S g ) (26) 6.

2. Input S s into the MLP layer to generate the final interest representation S of this anonymous user f : S f = ReLU(W′S s + b′) (27) where \(W'\in R\) 2d*d is a learnable parameter, \(b'\) is a bias term, and ReLU is used as the activation layer.

6. The recommended method based on click sequence time enhancement applied to the anonymous login scenario according to claim 1, wherein, In S7, it specifically includes the following steps: 7.

1. Calculate the final interest representation S of anonymous users f The similarity with the item node Normalize using softmax, and the specific formula is as follows: Among which V i is the node feature vector generated by the graph neural network; 7.

2. Use cross-entropy to calculate the similarity between the true value and the predicted value. The calculation formula is as follows: Where y i represents the true value label, which is the probability that the item will be clicked next under the result of the user's actual click, and is expressed in the form of a one-hot encoded vector; is the probability that the item is predicted. When the loss value is smaller, it means that the predicted value is closer to the true value, and vice versa is larger.

Images