CN114238765A - Block chain-based position attention recommendation method - Google Patents

Abstract

The invention discloses a position attention recommendation method based on a block chain, which comprises the following steps: firstly, deploying a recommendation system on a block chain platform; secondly, uploading user data; thirdly, constructing a graph; fourthly, establishing an attention network; fifthly, determining position information; sixthly, establishing a self-attention layer; and step seven, establishing a prediction layer. Has the advantages that: the conversion relation between adjacent items in the conversation can be fully trained, and the sequential significance of the data item relation can be obtained while the data item relation is obtained. Reducing the loss of the training process. The block chain technology is used, the behavior data are uploaded to the block chain through the behavior recording contract, and accurate recommendation is performed according to user preference on the premise that the safety of the user data is guaranteed.

Description

Block chain-based position attention recommendation method

Technical Field

The invention relates to a position attention recommending method, in particular to a block chain-based position attention recommending method.

Background

At present, with the continuous progress of science and technology, the problems of information explosion, overload and the like become more and more serious, and the recommendation system can well solve the problems. However, the traditional recommendation model cannot acquire the information of the short-term transactions of the user, so that the short-term preference prediction of the user is greatly reduced. The recommendation model based on the conversation effectively solves the problems, and the conversation is the recent affairs of the user, such as the purchase of articles by the user, the click sequence of the user and the like.

The author combines the session with a K-neighborhood algorithm and makes recommendations according to K pieces of session information similar to the current session, and the SKNN takes context information into consideration. The FPMC model proposed by Rendle et al combines a user-project matrix decomposition technique with a Markov chain, firstly constructs a personalized transfer matrix based on the Markov chain, and then solves the matrix sparsity problem by utilizing a matrix decomposition model. GRU4Rec by Hidasi et al first applies a recurrent neural network to the recommendations, the model having multiple levels of gating cells to control the amount of information passed to the next level in the neural network. Wu et al propose an SR-GNN model to represent sessions as graphs, which can enhance the transformation information of adjacent items, then obtain long-term session embedding through a gated neural network, regard the last clicked item as short-term session embedding, and combine the two session embedding to form final session embedding. Although these models provide a great improvement in recommendation accuracy, they do not take into account the impact on recommendations of the relevance of the order of the session items to the higher-order items in the session, and the privacy of the user data. In the era of easy data acquisition, all large e-commerce can acquire behavior data of users through some technologies, but the user data is usually stored in a centralized database for each large module to call and access. This is prone to data leakage problems.

It has been found through research that many of the problems of the conventional recommendation system are not solved. The first problem is that the order information of items in the user historical data is not integrated, the second problem is that the high-order correlation information of the items in the data cannot be obtained, and the third problem is that the privacy problem is not solved. To solve the above problem, the GPAN recommendation system is proposed herein, which applies session item location representation and self-attention neural network in a blockchain platform to obtain user preferences.

Disclosure of Invention

The method mainly aims to solve the problem that the conventional recommendation model is not integrated into item sequence information in data;

the second purpose of the invention is to solve the problem that the existing recommendation model can not obtain the high-order correlation information of the items in the data;

the third purpose of the invention is to solve the problem of how to improve the recommendation precision on the premise of ensuring the privacy of the user data;

the present invention provides a location attention recommendation method based on a block chain to achieve the above object and solve the above problems.

The invention provides a block chain-based position attention recommendation method, which comprises the following steps:

firstly, deploying a recommendation system on a block chain platform: the deployment recording intelligent contract compiling module is used for distributing an area to record user information for each block in the block chain while a user registers as a user in an e-commerce webpage;

and secondly, uploading user data: uploading the history records of the users to a system, and then compiling a behavior record intelligent contract through a behavior record intelligent contract compiling module, wherein each block in a block chain records the encrypted user history records;

thirdly, constructing a graph: in the construction of the graph, the method comprises an item set, a session composed of user behaviors ordered according to time stamps, an item to be clicked next by a user, a generated probability sequence and the like, wherein the item set forms nodes in the network graph of the knowledge graph, and the session is shown by edges linking the nodes;

the existing item set Let V ═ V₁,v₂,v₃,...,v_|n|Denotes a set of all unique items involved in a session, n is the total number of items, and each session is denoted by a time stamp as S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]In which V is_s,iE.g. V, m is the session length, and the purpose is to predict the next item to be clicked by the user, namely V_s,m+1Session S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]Can be regarded as a node, wherein (V)_s,i-1,V_s,i) Indicating that the user is clicking V_s,i-1After the node is connected with V_s,iInteractions occur so that each session can be represented as a graph;

the interaction between the nodes is divided into four types, the first type is self-connection, each node is self-connected firstly, the purpose is to strengthen the influence of self information in the following network, and the self-connection is represented by 1; the second type is out-degree connection, which represents that the current node is out and is represented by 2; the third type is an inbound connection, which indicates that the current node is inbound and is represented by 3; the fourth type is interconnection, which means that the current node and other nodes are mutually interacted and is represented by 4; if no interaction is represented by 0, the session representation space and the node vector V epsilon R are set^dD is the dimension, and the session representation consists of node vectors;

step four, establishing an attention network: acquiring conversion information between adjacent items by using an attention network, multiplying adjacent nodes first, and then multiplying corresponding weights, wherein a specific formula is as follows:

where σ is the LeakyRuLU activation function,

r_ijis a node v_iAnd v_jThe relationship between the nodes is trained by four different weight matrixes, and is A_self,A_in,A_out,A_inoutRespectively corresponding to the self-connection, in-degree connection, out-degree connection and two-way connection relations; w is a linear transformation of the shared parameter applied to all nodes,

the specific formula is as follows:

then the weights are normalized to facilitate calculation,

is a node v_jAfter the weights of the node neighbors are obtained, weighted multiplication is carried out, the purpose is to aggregate each neighbor node information, and the calculation formula is as follows:

through the establishment of the attention network, a local session representation only focusing on the characteristics of the local session representation and the neighbor is obtained;

fifthly, determining the position information: the position information is the reverse learning item position embedding, and a position matrix R is set_l＝[Ρ₁,Ρ₂,Ρ₃,...,Ρ_m]M is the length of the current session, l is the number of sessions, and a session indicates that S ═ h_s,1,h_s,2,h_s,3,...,h_s,m]Then, the formula for embedding the ith item fusion position is as follows:

σ is the tan h activation function,

W₁the position weight matrix is a position weight matrix, b is a bias term and is a trainable parameter, the weight matrix is fused in the embedding matrix of the position information, so that the position information can obtain a proper weight, because the position information is only auxiliary information, the real key point is conversation representation, and the specific formula is as follows:

W₂,W₃,W₄is a weight matrix, c is an offset term, H^*Is the average value of all the items in the current session, and the specific formula is as follows:

for average embedding of t-th article in conversation, multiplying the matrix with adjusted position information weight to corresponding conversation representation to obtain S^*Embedding for short-term session, then embedding S^*Afferent from the attention layer;

and a sixth step of establishing a self-attention layer: the long-term preference is obtained by aggregating all item embeddings in a conversation with self-attention, which has resulted in a conversation representation S incorporating location embedding^*And aggregating the conversation expressions into a final conversation expression through a self-attention layer, wherein the specific formula is as follows:

WQ，WK，

are all self-attention linear transformation matrixes, all derived from S^*Self-attention is divided into two processes, the first process being based on W^QAnd W^KPerforming weight coefficient calculation, and performing the second process based on the weight coefficient pair W^VSumming is carried out;

and finally, a brand-new residual error connection is used to enable the attention network to be more stable, ReLU activation is firstly carried out during training, then linear connection is carried out, two layers of residual error connection are used, more information of the previous layer can be kept, and the training loss is reduced, wherein the specific formula is as follows:

C′＝σ(σ(F)W₅+b₂)W₆+b₃

C^*＝W₇C′+(1-W₇)F

where σ is the ReLU activation function, W₅，W₆，

b₃And b₂For d-dimension bias phase, each linear transformation adds a "Dropout" layer to prevent overfitting, and finally long-term embedding of the session into C^*Short-term embedding with session S^*And performing linear combination to obtain the final session embedding, wherein the specific formula is as follows:

C＝W₈C^*+(E-W₈)S^*

e is a unit matrix of dxd, and finally, the final session embedding including short-term and long-term embedding is obtained;

step seven, establishing a prediction layer: and predicting the item clicked next by the user by using the final session embedding, wherein the specific formula is as follows:

using the cross entropy as a loss function for the predicted probability of the next user clicked on the item, the formula is as follows:

y is the representation of the real item, we get the minimum loss through Adam's optimization, and finally recommend the item with higher score to the user.

The invention has the beneficial effects that:

compared with the traditional recommendation model, the block chain technology is added in the block chain-based position attention recommendation method provided by the invention, the reconstructed neural network structure is combined, the session position information is fused, the conversion relation between adjacent items in the session can be fully trained, and the sequential significance can be obtained while the data item relation is obtained. After the representation fused with the data position information is obtained, the long-term preference representation of the user is obtained through a self-attention layer. The invention improves the self-attention network, and provides a new residual error network after self-attention training to make the network more stable and reduce the loss of the training process. The long-term preferences of the user are linearly linked with the short-term preferences to form end-user preferences. According to the invention, a redundant hypergraph is not required to be established, and a graph used for both long-term preference and short-term preference is obtained, so that the time complexity and the space complexity are reduced while the accuracy is ensured. The invention divides the relation between projects into self-linking, unidirectional linking and bidirectional linking, wherein the unidirectional linking is divided into a node a to a node b and a node b to a node a. The algorithm is trained by using different weight matrixes according to the relation among different projects, is more suitable for the scene of the user historical behavior in reality, and has a more closed structure with the user historical behavior and more accurate prediction result. The invention uses the block chain technology, the behavior data is uploaded to the block chain through the behavior recording contract, and accurate recommendation is carried out according to the user preference on the premise of ensuring the safety of the user data.

Drawings

FIG. 1 is a schematic diagram of the overall operation of the method of the present invention.

Detailed Description

Please refer to fig. 1:

the invention provides a block chain-based position attention recommendation method, which comprises the following steps:

the first step is to deploy the system on a block chain platform: and the deployment recording intelligent contract compiling module is used for distributing an area to record user information for each block in the block chain while a user registers as a user in the e-commerce webpage.

And secondly, uploading user data: uploading the history records of the users to a system, then compiling the behavior record intelligent contract through a behavior record intelligent contract compiling module, and recording the encrypted user history records in each block in the block chain.

Thirdly, constructing a graph: in the construction of the graph, the method comprises an item set, a session composed of user behaviors ordered according to time stamps, an item to be clicked next by a user, a generated probability sequence and the like, wherein the item set forms nodes in the network graph of the knowledge graph, and the session is shown by edges linking the nodes;

the existing item set Let V ═ V₁,v₂,v₃,...,v_|n|Denotes a set of all unique items involved in a session, n is the total number of items, and each session is denoted by a time stamp as S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]In which V is_s,iE.g. V, m is the session length, and the purpose is to predict the next item to be clicked by the user, namely V_s,m+1Session S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]Each item in (1) canTo be seen as a node, where (V)_s,i-1,V_s,i) Indicating that the user is clicking V_s,i-1After the node is connected with V_s,iInteractions occur so that each session can be represented as a graph;

the interaction between the nodes is divided into four types, the first type is self-connection, each node is self-connected firstly, the purpose is to strengthen the influence of self information in the following network, and the self-connection is represented by 1; the second type is out-degree connection, which represents that the current node is out and is represented by 2; the third type is an inbound connection, which indicates that the current node is inbound and is represented by 3; the fourth type is interconnection, which means that the current node and other nodes are mutually interacted and is represented by 4; if no interaction is represented by 0, the session representation space and the node vector V epsilon R are set^dD is the dimension, and the session representation consists of node vectors;

step four, establishing an attention network: acquiring conversion information between adjacent items by using an attention network, multiplying adjacent nodes first, and then multiplying corresponding weights, wherein a specific formula is as follows:

where σ is the LeakyRuLU activation function,

r_ijis a node v_iAnd v_jThe relationship between the nodes is trained by four different weight matrixes, and is A_self,A_in,A_out,A_inoutRespectively corresponding to the self-connection, in-degree connection, out-degree connection and two-way connection relations; w is a linear transformation of the shared parameter applied to all nodes,

the specific formula is as follows:

then the weights are normalized to facilitate calculation,

is a node v_jAfter the weights of the node neighbors are obtained, weighted multiplication is carried out, the purpose is to aggregate each neighbor node information, and the calculation formula is as follows:

through the establishment of the attention network, a local session representation only focusing on the characteristics of the local session representation and the neighbor is obtained;

fifthly, determining the position information: the position information is the reverse learning item position embedding, and a position matrix R is set_l＝[Ρ₁,Ρ₂,Ρ₃,...,Ρ_m]M is the length of the current session, l is the number of sessions, and a session indicates that S ═ h_s,1,h_s,2,h_s,3,...,h_s,m]Then, the formula for embedding the ith item fusion position is as follows:

σ is the tan h activation function,

W₁the position weight matrix is a position weight matrix, b is a bias term and is a trainable parameter, the weight matrix is fused in the embedding matrix of the position information, so that the position information can obtain a proper weight, because the position information is only auxiliary information, the real key point is conversation representation, and the specific formula is as follows:

W₂,W₂,W₄is a weight matrix, c is an offset term, H^*Is the average value of all the items in the current session, and the specific formula is as follows:

for average embedding of t-th article in conversation, multiplying the matrix with adjusted position information weight to corresponding conversation representation to obtain S^*Embedding for short-term session, then embedding S^*Afferent from the attention layer;

and a sixth step of establishing a self-attention layer: the long-term preference is obtained by aggregating all item embeddings in a conversation with self-attention, which has resulted in a conversation representation S incorporating location embedding^*And aggregating the conversation expressions into a final conversation expression through a self-attention layer, wherein the specific formula is as follows:

W^Q,W^K,

are all self-attention linear transformation matrixes, all derived from S^*Self-attention is divided into two processes, the first process being based on W^QAnd W^KPerforming weight coefficient calculation, and performing the second process based on the weight coefficient pair W^VSumming is carried out;

and finally, a brand-new residual error connection is used to enable the attention network to be more stable, ReLU activation is firstly carried out during training, then linear connection is carried out, two layers of residual error connection are used, more information of the previous layer can be kept, and the training loss is reduced, wherein the specific formula is as follows:

C′＝σ(σ(F)W₅+b₂)W₆+b₃

C^*＝W₇C′+(1-W₇)F

where σ is the ReLU activation function, W₅，W₆，

b₃And b₂For d-dimension bias phase, each linear transformation adds a "Dropout" layer to prevent overfitting, and finally long-term embedding of the session into C^*Short-term embedding with session S^*And performing linear combination to obtain the final session embedding, wherein the specific formula is as follows:

C＝W₈C^*+(E-W₈)S^*

e is a unit matrix of dxd, and finally, the final session embedding including short-term and long-term embedding is obtained;

step seven, establishing a prediction layer: and predicting the item clicked next by the user by using the final session embedding, wherein the specific formula is as follows:

using the cross entropy as a loss function for the predicted probability of the next user clicked on the item, the formula is as follows:

y is a representation of the real term, and we get the least loss through Adam's optimization. And finally recommending the items with higher scores to the user.

The specific embodiment is as follows:

example 1:

in the digenetica dataset, the whole process of operation is carried out:

the first step is to deploy a system on a block chain platform and deploy an intelligent contract compiling module for behavior record.

And secondly, transmitting Diginetica data into the system, wherein each block in the block chain records the encrypted user history.

And thirdly, initializing a Digimetia data set, constructing the data into a graph structure, and dividing the relationships among the projects into four types, namely self-linking, out-degree linking, in-degree linking and bidirectional linking.

And fourthly, inputting the items in the conversation into the attention network for training to obtain an initial conversation representation.

And fifthly, training the reverse conversation item sequence to obtain a conversation position representation. The session location representation is fused with the initial session representation to obtain a user short-term preference representation.

And sixthly, inputting the short-term preference representation of the user into the self-attention layer to obtain the long-term preference of the user.

And seventhly, linearly linking the long-term preference of the user with the short-term preference of the user to obtain the final user preference, multiplying the final preference by the item representation to obtain the final score of each item, and selecting the top 20 items with the highest scores to recommend to the user.

And step eight, comparing eight reference models such as SR-GNN, NARM, STAMP and the like respectively, and obtaining an optimal result by the algorithm.

Example 2:

in the nowcasting dataset, the whole process of operation is carried out:

the first step is to deploy a system on a block chain platform and deploy an intelligent contract compiling module for behavior record.

And secondly, transmitting the Nowplaying data into the system, wherein each block in the block chain records the encrypted user history.

And thirdly, initializing a Digimetia data set, constructing the data into a graph structure, and dividing the relationships among the projects into four types, namely self-linking, out-degree linking, in-degree linking and bidirectional linking.

And fourthly, inputting the items in the conversation into the attention network for training to obtain an initial conversation representation.

And fifthly, training the reverse conversation item sequence to obtain a conversation position representation. The session location representation is fused with the initial session representation to obtain a user short-term preference representation.

And sixthly, inputting the short-term preference representation of the user into the self-attention layer to obtain the long-term preference of the user.

And seventhly, linearly linking the long-term preference of the user with the short-term preference of the user to obtain the final user preference, multiplying the final preference by the item representation to obtain the final score of each item, and selecting the top 20 items with the highest scores to recommend to the user.

And step eight, comparing eight reference models such as SR-GNN, NARM, STAMP and the like respectively, and obtaining an optimal result by the algorithm.

Example 3:

in the Tll data set, the whole process of operation is as follows:

the first step is to deploy a system on a block chain platform and deploy an intelligent contract compiling module for behavior record.

And secondly, transmitting the Tmax data into the system, wherein each block in the block chain records the encrypted user history.

And thirdly, initializing a Digimetia data set, constructing the data into a graph structure, and dividing the relationships among the projects into four types, namely self-linking, out-degree linking, in-degree linking and bidirectional linking.

And fourthly, inputting the items in the conversation into the attention network for training to obtain an initial conversation representation.

And fifthly, training the reverse conversation item sequence to obtain a conversation position representation. The session location representation is fused with the initial session representation to obtain a user short-term preference representation.

And sixthly, inputting the short-term preference representation of the user into the self-attention layer to obtain the long-term preference of the user.

And seventhly, linearly linking the long-term preference of the user with the short-term preference of the user to obtain the final user preference, multiplying the final preference by the item representation to obtain the final score of each item, and selecting the top 20 items with the highest scores to recommend to the user.

And step eight, comparing eight reference models such as SR-GNN, NARM, STAMP and the like respectively, and obtaining an optimal result by the algorithm.

Claims (1)

1. A location attention recommendation method based on a block chain is characterized in that: the method comprises the following steps:

firstly, deploying a recommendation system on a block chain platform: in the block chain platform deployment recommendation system, an intelligent contract compiling module is deployed and recorded, and when a user registers as a user in an e-commerce webpage, each block in a block chain is allocated with a block of area to record user information;

and secondly, uploading user data: uploading user data, uploading a history record of a user to a system, and then compiling a behavior record intelligent contract through a behavior record intelligent contract compiling module, wherein each block in a block chain records the encrypted user history record;

thirdly, constructing a graph: in the construction of the graph, the method comprises an item set, a session composed of user behaviors ordered according to time stamps, an item to be clicked next by a user, a generated probability sequence and the like, wherein the item set forms nodes in the network graph of the knowledge graph, and the session is shown by edges linking the nodes;

the existing item set Let V ═ V₁,v₂,v₃,...,v_|n|Represents a set of all unique items involved in a session, n is the total number of items, each session is represented by a timestampIs S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]In which V is_s,iE.g. V, m is the session length, and the purpose is to predict the next item to be clicked by the user, namely V_s,m+1Session S ═ V_s,1,V_s,2,V_s,3,...,V_s,m]Can be regarded as a node, wherein (V)_s,i-1,V_s,i) Indicating that the user is clicking V_s,i-1After the node is connected with V_s,iInteractions occur so that each session can be represented as a graph;

the interaction between the nodes is divided into four types, the first type is self-connection, each node is self-connected firstly, the purpose is to strengthen the influence of self information in the following network, and the self-connection is represented by 1; the second type is out-degree connection, which represents that the current node is out and is represented by 2; the third type is an inbound connection, which indicates that the current node is inbound and is represented by 3; the fourth type is interconnection, which means that the current node and other nodes are mutually interacted and is represented by 4; if no interaction is represented by 0, the session representation space and the node vector V epsilon R are set^dD is the dimension, and the session representation consists of node vectors;

step four, establishing an attention network: acquiring conversion information between adjacent items by using an attention network, multiplying adjacent nodes first, and then multiplying corresponding weights, wherein a specific formula is as follows:

where σ is the LeakyRuLU activation function,

r_ijis a node v_iAnd v_jThe relationship between the nodes is trained by four different weight matrixes, and is A_self,A_in,A_out,A_inoutRespectively corresponding to the self-connection, in-degree connection, out-degree connection and two-way connection relations; w is an applicationA linear transformation of the shared parameters of all nodes,

the specific formula is as follows:

then the weights are normalized to facilitate calculation,

is a node v_jAfter the weights of the node neighbors are obtained, weighted multiplication is carried out, the purpose is to aggregate each neighbor node information, and the calculation formula is as follows:

through the establishment of the attention network, a local session representation only focusing on the characteristics of the local session representation and the neighbor is obtained;

fifthly, determining the position information: the position information is the reverse learning item position embedding, and a position matrix R is set_l＝[Ρ₁,Ρ₂,Ρ₃,...,Ρ_m]M is the length of the current session, l is the number of sessions, and a session indicates that S ═ h_s,1,h_s,2,h_s,3,...,h_s,m]Then, the formula for embedding the ith item fusion position is as follows:

σ is the tan h activation function,

W₁the position weight matrix and b is a bias term, which are trainable parameters, and the position isThe weight matrix is fused in the embedded matrix of the information, so that the position information can obtain a proper weight, because the position information is only auxiliary information, the real key point is conversation representation, and the specific formula is as follows:

W₂,W₃,W₄is a weight matrix, c is an offset term, H^*Is the average value of all the items in the current session, and the specific formula is as follows:

for average embedding of t-th article in conversation, multiplying the matrix with adjusted position information weight to corresponding conversation representation to obtain S^*Embedding for short-term session, then embedding S^*Afferent from the attention layer;

and a sixth step of establishing a self-attention layer: the long-term preference is obtained by aggregating all item embeddings in a conversation with self-attention, which has resulted in a conversation representation S incorporating location embedding^*And aggregating the conversation expressions into a final conversation expression through a self-attention layer, wherein the specific formula is as follows:

are all self-attention linear transformation matrixes, all derived from S^*Self-attention is divided into two processes, the first process being based on W^QAnd W^KPerforming weight coefficient calculation, and performing the second process based on the weight coefficient pair W^VSumming is carried out;

and finally, a brand-new residual error connection is used to enable the attention network to be more stable, ReLU activation is firstly carried out during training, then linear connection is carried out, two layers of residual error connection are used, more information of the previous layer can be kept, and the training loss is reduced, wherein the specific formula is as follows:

C′＝σ(σ(F)W₅+b₂)W₆+b₃

C^*＝W₇C′+(1-W₇)F

where σ is the ReLU activation function,

b₃and b₂For d-dimension bias phase, each linear transformation adds a "Dropout" layer to prevent overfitting, and finally long-term embedding of the session into C^*Short-term embedding with session S^*And performing linear combination to obtain the final session embedding, wherein the specific formula is as follows:

C＝W₈C^*+(E-W₈)S^*

e is a unit matrix of dxd, and finally, the final session embedding including short-term and long-term embedding is obtained;

step seven, establishing a prediction layer: and predicting the item clicked next by the user by using the final session embedding, wherein the specific formula is as follows:

using the cross entropy as a loss function for the predicted probability of the next user clicked on the item, the formula is as follows:

y is the representation of the real item, we get the minimum loss through Adam's optimization, and finally recommend the item with higher score to the user.

Images