CN115310004A - Graph nerve collaborative filtering recommendation method fusing project time sequence relation - Google Patents

Abstract

The invention discloses a graph nerve collaborative filtering recommendation method fusing a project time sequence relation. The invention considers the dynamic change of user interest, and uses the item-item time sequence relation to dynamically depict the interest change of the user; the recommendation performance is improved by using the characteristic that the graph structure displays and expresses high-order information of the user-item. The method comprises the following steps: preprocessing project data interacted by a user; dividing a data set into a training set and a testing set; acquiring user-project interaction data, including constructing an interaction matrix; acquiring project-project time sequence relation data, wherein the project-project time sequence relation data comprises sub-sequences divided in a sliding window mode, a project-project time sequence relation matrix constructed and matrix normalization processing; initializing various hyper-parameters in the graph nerve collaborative filtering recommendation method fusing the project time sequence relation, and testing the effect of the model through prediction on the test set. The method can improve the performance of the recommendation algorithm.

Description

Graph nerve collaborative filtering recommendation method fusing project time sequence relation

Technical Field

The invention relates to a graph nerve collaborative filtering recommendation method fusing a project time sequence relation, and belongs to the field of recommendation.

Background

In recent years, with the rapid development of society, information technology and information resources are expanding, and the whole internet becomes a powerful information base and information exchange place, resulting in the problem of information overload. The great popularity of mobile devices has made internet applications more frequent for users and left a great deal of interactive information during use. By storing and mining the potential rules of the data, the user preference can be well learned and the following preference can be predicted. In recent years, the recommendation algorithm can capture the user interest by mining the historical interaction information of the user, predict and recommend the items interacted by the user next, help the user to save time, and be widely used in the internet. The recommendation system is widely applied to the fields of online shopping malls, video websites, social platforms and the like, and the Taobao commodity recommendation brings convenience to users and simultaneously promotes the user consumption and improves the consumption amount; the news which the user is interested in is recommended to the user through the recommendation system, so that the user is facilitated. Therefore, the recommendation algorithm not only saves the time of the user, but also improves the profits of related companies, so that the recommendation algorithm is an important branch in the field of machine learning regardless of the close attention on the development of the recommendation algorithm in academic circles or industries.

The collaborative filtering recommendation algorithm is the most classical recommendation algorithm, and the core of the collaborative filtering recommendation algorithm is to utilize the historical behaviors of the user to mine the interests of the user. The traditional collaborative filtering recommendation algorithm is widely concerned by the industry and the academia because it only uses scoring information, and its core idea is to recommend to a user the same items that the user prefers as the user prefers, such as: a user-based collaborative filtering recommendation algorithm [9] and an item-based collaborative filtering recommendation algorithm. However, these collaborative filtering-based recommendation algorithms tend to depend largely on how rich the user leaves effective traces in the system, i.e., how much the user has recorded the interaction with the item. Thus, the performance of these algorithms is poor for both new users without interaction records and users with rare interaction records with items, which often leads to cold start and data sparseness problems in the recommendation system. To solve these problems, most collaborative filtering-based recommendation algorithms map user and item information into a potential embedding space, such as matrix decomposition-based and neural network-based methods. In order to obtain high-quality embedded representation, most of current researches utilize a graph neural network to carry out graph embedding on user-item interaction information, so that the overall recommendation performance of the algorithm is remarkably improved. Ying et al propose a graph convolution neural network-based recommendation model PinSage, which captures features of a graph structure and features of nodes by using methods such as random walk and graph convolution to generate embedded representation of the nodes, thereby greatly improving the quality of the embedded representation. Wang et al propose a graph-based neural collaborative filtering recommendation algorithm NGCF, which further improves the embedded representation capability by embedding and coding the historical interaction information of user-items by using a bipartite neural network and explicitly considering the high-order connectivity between the user-items. He et al propose a lightweight volume network model LightGCN for recommendation systems that abandons the feature transformation and nonlinear activation of traditional volume networks, learns user and item embeddings by linear propagation on the user-item interaction matrix, and finally sums the embedding weights learned by all layers as the final embedded representation. Zhang et al propose a hybrid recommendation algorithm based on the combination of paragraph embedding and neural networks. The method adopts the neural network to carry out collaborative filtering on the user item scores, and has higher nonlinearity on the complex structure for capturing the user interaction scores. Meanwhile, the cold start problem is solved to a certain extent by utilizing the content characteristics of the auxiliary information acquired by project embedding. Wang et al propose KGCN, apply knowledge graph with recommending the field to excavate the relation on the knowledge graph attribute of the commodity in combination with the neural network of the figure at the same time, effectively catch the inherent contact of the commodity, relieve the data sparseness.

However, most of the current research based on the graph neural network recommendation algorithm only focuses on the binary relation of user-item, and the intrinsic logic between the user-item interaction information and the item time series dependency relation is less considered. These methods generally assume that the user's interests are static, and less consider characterizing the user's dynamic interest changes through the time series dependencies of the items. In practical applications, as the scale of user-item interaction is continuously enlarged, the extraction of the collaboration signal between users and items becomes more and more difficult, and the accuracy of recommendation is greatly reduced. On the other hand, most of the existing methods adopt descriptive characteristics (such as ID and attributes) of users/projects to construct embedding functions, when the relationship between the users and the projects is more and more complex, overfitting often occurs in the methods, data sparsity becomes more and more obvious, and the generalization capability of the model is influenced to a certain extent. Therefore, it becomes more and more important how to effectively extract the dependency relationship between the user-item interaction information and the item time sequence, accurately depict the dynamic interest change of the user, and effectively prevent the occurrence of the overfitting phenomenon caused by data sparseness.

Disclosure of Invention

In order to overcome the defects of the models, the invention provides a graph neural collaborative filtering recommendation method fusing the project time sequence relation. Dividing the project time sequence into a plurality of groups of subsequences by adopting a sliding window strategy, constructing a project time sequence dependency relationship diagram, and deeply depicting the dynamic change of user interest by aggregating project time information characteristics; secondly, high-dimensional information of user-project and project-project is mapped to a low-dimensional space by utilizing a bipartite neural network, so that mixed embedding of historical interaction information of the user-project, project time sequence dependency information and the like is realized, and expression of user-project interaction sequence information is enhanced. The experiments were validated on three authentic data sets (LastFM, douban, ciao).

The invention discloses a graph nerve collaborative filtering recommendation method fusing a project time sequence relation, which comprises the following specific steps:

step 1: acquiring project data of each user interaction;

step 2: preprocessing the project data of user interaction, including removing abnormal values and removing users with the number of interaction projects less than 5;

and 3, step 3: converting a data format, dividing a data set, and dividing the data set into a training set and a test set;

and 4, step 4: acquiring user-project interaction data and project-project time sequence relation data, wherein the user-project interaction data and the project-project time sequence relation data comprise a user-project interaction matrix, a project-project time sequence relation matrix and normalization processing of the matrix;

and 5: initializing various parameters of a hybrid recommendation model (NGCF-ITS) fusing user-Item interaction information (explicit information) and Item time sequence dependency (implicit information), inputting data of a training set into a network for training until the network converges, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

and 6: and (4) verifying the model effect through the evaluation indexes of the test set data to obtain the performance improvement of the recommendation algorithm.

The invention adopts Bayesian Personalized Ranking (BPR for short) as a loss function, and the calculation formula is as follows:

where (u, v, z) ∈ O, (v, z) is a training sample pair, where v represents an item interacted with by user u, z represents an item not interacted with by user u, and σ () is a sigmoid activation function. Theta denotes all trainable parameters in the model, and L is controlled by the parameter lambda ₂ The regular strength prevents overfitting.

The recommended performance metrics include Precision (Precision), recall (Recall), and Normalized cumulative loss Gain (NDCG)

Where N is the number of items recommended to the user,

to test a set of pooled users, R _u Representing a list of items recommended to the user, I _u The method comprises the following steps that a user interacts with an item set, and precision @ N measures the ratio of correctly recommended items to total recommended items; recall @ N is the ratio of recommended correct items to user-interacted items; DCG @ N is a consideration of the ranking factor, making higher the gain for top ranked items, rel when item i is taken _i =1, otherwise equal to 0; IDCG @ N is ideally DCG @ N, NDCG @ N to measure and evaluate the search result algorithm. The larger the values of precision @ N, recall @ N, NDCG @ N, the better the recommendation effect.

The method and the system not only consider the project-project time sequence relation to dynamically depict the interest change of the user, but also relieve the problems of data sparseness and cold start, can accurately represent the embedded vectors of the user and the project, provide more accurate recommendation, and improve the recommendation performance. Therefore, the method provided by the invention realizes a relatively accurate prediction effect.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

fig. 2 is a diagram of the main steps of the present invention.

Detailed Description

The invention will be further explained with reference to the drawings.

As shown in fig. 1 and 2, the neural collaborative filtering recommendation method for a graph fusing a project time sequence relationship of the present invention specifically includes the following steps:

step 1: acquiring project data of each user interaction;

step 1.1: acquiring project data of each user interaction, wherein the project data comprises four data characteristics of a user id, a project id, a timestamp and a score;

step 2: preprocessing the project data of user interaction, including removing abnormal values and removing users with the number of interaction projects less than 5;

step 2.1: removing abnormal values, and directly deleting records with obvious abnormal values in the data;

step 2.2: removing users with the project data volume of user interaction smaller than 5;

and step 3: converting a data format, dividing a data set, and dividing the data set into a training set and a test set;

step 3.1: converting a data format, converting an original data sequence into a format of a supervised learning sequence, and sequencing according to ascending time stamps of interaction between a user and a project;

step 3.2: dividing the data set, and enabling the historical time-series data set to be divided into 8: the ratio of 2 is divided into a training set and a test set.

And 4, step 4: acquiring user-project interaction data and project-project time sequence data, wherein the user-project interaction data and the project-project time sequence data comprise a user-project interaction matrix, a project-project time sequence relation matrix and normalization processing of the matrix;

step 4.1: constructing user-project interaction diagram and user-project interaction matrix

And constructing a user-item interaction graph G < u, r, i >, wherein u represents a user, i represents an item, r represents interaction between the user and the item, if r =1, successful interaction between the user and the item is represented, and otherwise, r =0, no interaction is represented. Constructing a user-item interaction matrix R ₁ If user u has successfully interacted with item i once, i.e. R =1, then R _1u,i =1, otherwise 0.

And 4.2: constructing an item-item time sequence relation matrix and an item-item time sequence relation matrix

Constructing a graph G < V, E > of the dependency relationship between the project and the time series of the project, effectively sequencing information related to project interaction in all historical behavior data of a user to form a historical project interaction sequence set V = { V = ₁ ,V ₂ ,V ₃ ,...,V _M }; then, a sliding window is usedThe oral strategy effectively divides the user historical item interaction sequence set obtained in the last step, forming several subsequences T = { T = { (T) ₁ ,T ₂ ,T ₃ ,...,T _H }; and finally, on the basis of the subsequences, performing undirected connection on the items in each subsequence and other items, thereby constructing a time series dependency graph of the items. Considering the dynamic variation of user interests, random division will lead to the simplification and inaccuracy of user interest description, and also lose the diversity characteristics of many items. Thus, to avoid loss of generality, uniform partitioning is achieved herein by setting the sliding window size, i.e., setting a parameter β to identify the size of a particular sliding window. Constructing an item-item time sequence relation matrix R according to the item-item time sequence relation graph G < V, E > ₂ If item i _x And item i _y If the connection is made, the item time sequence relation exists between the two items, so R _2ix,iy =1. Then normalization processing is carried out.

And 5: initializing various parameters of a graph nerve collaborative filtering recommendation method (NGCF-ITS) fusing a project time sequence relation, inputting data of a training set into a network for training until the network converges, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

step 5.1: building an NGCF-ITS structure, which comprises an embedding layer, an aggregation layer, a propagation layer and a prediction layer; setting an embedding dimension and initializing a weight matrix;

step 5.2: initializing a user embedded matrix and a project embedded matrix;

P＝[p ₁ ,p ₂ ,p ₃ ,...,p _M ]

Q＝[q ₁ ,q ₂ ,q ₃ ,...,q _N ]

P∈R ^d×M 、Q∈R ^d×N respectively representing a user embedding matrix and an item embedding matrix, wherein M represents the number of users, and N represents itemsThe number of the cells. p is a radical of _u ∈R ^d (q _i ∈R ^d ) An embedding vector representing user u (item i), and d represents the dimension of the embedding vector.

Step 5.3: calculating a project-project time sequence relation, performing time sequence fusion, embedding a learning project into a vector and transmitting the vector to the next layer;

and aggregating the item-item time sequence relation for multiple times, wherein the item embedding formula after the aggregation for the first time is as follows:

wherein

The embedded vector representing item v of the i-th aggregation level, when l =0,

embedding a vector, m, for initialization of item v _v←v′ Representing the time-series dependency of the aggregation of a term v from its neighbor node v ', where v' e N _v ，N _v A set of neighbor nodes denoted as item v. f (-) is a coding function, where m _v←v′ The formula is as follows:

LeakyRelu (. Circle.) is an activation function,

is a normalized matrix, i.e.

A is an item connection matrix, E is an identity matrix,

is a function of the normalization process function,

is the connecting coefficient of the item v and the item v' in the normalized matrix. The item embedding vectors for the final aggregation layer are:

wherein, the first and the second end of the pipe are connected with each other,

and

the weight parameter representing the number l not only considers the item v from its neighbor node set N _v The time relationship of aggregation also takes the item characteristics of the item into consideration.

Step 5.4: calculating the high-order connectivity of user-project interaction, continuously learning a user embedded vector and a user embedded vector, embedding and propagating at multiple levels, and connecting results; the method and the system utilize the connection between the user and the project to embed and propagate, capture the cooperative signal and acquire the high-order similarity. Wherein the user embedding formula of the times I is

In the formula (I), the compound is shown in the specification,

representing the embedding of user u after propagation through l layers. Not only considering information propagation from neighbors

Also adds self-connection

Wherein

And

the weights in the propagation formula are the same,

and

the weight matrix is shared. The LeakyRelu (-) activation function in this equation is a function that allows information to encode both positive and negative signals. The embedded propagation formula of the item to the user is as follows:

user embedded vectors

And item embedding vectors

As an input to the propagation layer(s),

is a laplacian norm expressed as a signal strength coefficient, i.e., to control the association of the user u with each neighboring item v, where N is _u N _v Respectively representing single neighbors of the user u and the item v, namely adjacent nodes;

respectively, representing the weight matrix of the propagation process.

Similar to solving for user embedding, the embedding formula for the project is written as:

step 5.5 predicting the end result

Will pass through the different propagation results of the times l

And

performing connection operation to obtain final user p _u And item q _v The embedded vector of (a) is:

wherein, | | is a linking operation,

the embedded vectors of the user u and the project v in the operation for the first time are respectively represented, the range of l is controlled, the embedded vectors of the user and the project can be effectively obtained, and the problems that insufficient information extraction is caused due to too small l, parameters are too large due to too large l, space and time complexity is increased and the like can be avoided. The final prediction results are:

step 5.6: adopting BPR as a loss function and Adam as an optimization algorithm, and circularly reciprocating until the model converges;

where (u, v, z) ∈ O, (v, z) is a training sample pair, where v represents an item interacted with by user u, z represents an item not interacted with by user u, and σ () is a sigmoid activation function. Theta represents all trainable parameters in the model, and the parameters are usedλ to control L ₂ The regular strength prevents overfitting.

Step 5.6: comparing and adjusting parameters, and acquiring and storing optimal parameters; including learning rate, mini-batch size, L2 regularization coefficients, weight vectors, and bias vectors.

And 6: and (4) verifying the model effect through the evaluation indexes of the test set data to obtain the performance improvement of the recommendation algorithm.

Step 6.1: predicting the articles which are interested by the user on the test set by using the trained prediction model;

step 6.2: comparing all the prediction results with the real results, and obtaining the performance evaluation result of the model by using precision @ N, recall @ N and NDCG @ N as the evaluation indexes of the model;

step 6.3: the final prediction model is used for prediction of the recommendation algorithm.

Experiment:

1. data set

According to the invention, lastFM, douban and Ciao are respectively used for verifying the recommended performance of the experiment. LastFM is a data set of a user song listening sequence which is issued under the framework of information heterogeneity and fusion of a second international seminar of a fifth ACM recommendation system conference and has implicit feedback of context information; the double is a relatively famous theme social network in China, comprises themes including movies, books, music and the like, and uses a data set about movie scores; ciao was crawled from the category of DVDs of the well-known review site dvd. Ciao.co.uk site in 12 months 2013. The information analysis of the 3 data sets is shown in table 1.

TABLE 1 introduction Table of data set

2. Evaluation criteria

To verify the unbiasedness of the model, experiments were performed on three true datasets LastFM, douban, ciao. Before the formal experiment, a 5-fold cross validation method is adopted, a data set is randomly and averagely divided into 5 parts, 1 part of the data set is sequentially selected as a test set, the rest 4 parts of data are used as a training set, 5 models are trained to obtain 5 experimental results, and the average value of the 5 results is selected as a final experimental result. The invention selects 3 popular ranking indexes to evaluate the recommendation performance, namely Precision (Precision), recall (Recall) and Normalized broken cumulative Gain (NDCG), wherein the higher the values of the indexes are, the better the recommendation quality is. Similar to most recommendation algorithms, the top N items are recommended for the user, namely top-N recommendation, and the expressions of precision @ N, recall @ N and NDCG @ N are defined as follows:

where N is the number of items recommended to the user,

to test a set of pooled users, R _u Representing a list of items recommended to a user, I _u The method comprises the following steps that a user interacts with an item set, and precision @ N measures the ratio of correctly recommended items to total recommended items; recall @ N is the ratio of recommending correct items to items that the user has interacted with; DCG @ N is a consideration of the ranking factor, such that top ranked items gainHigher, when item i is adopted, rel _i =1, otherwise equals 0; IDCG @ N is DCG @ N and NDCG @ N which are ideal to measure and evaluate the search result algorithm. The larger the values of precision @ N, recall @ N, NDCG @ N, the better the recommendation effect.

3. Comparative method and experimental results

The experiment selects the existing recommended method (APR, CURE _ BPR, CFGAN, neuMF and NGCF) to carry out experimental comparison on the data sets LastFM, douban and Ciao with the NGCF-ITS. Other variables were set to the best state for this experiment, and other methods used the default parameters in their references, and the results of the experiment were compared as shown in Table 2

TABLE 2 Experimental comparison Table

Through the technical scheme, the method is based on an algorithm of graph collaborative filtering, not only considers the project-project time sequence relation to dynamically depict the interest change of the user, but also relieves the problems of data sparseness and cold start, can accurately represent the embedded vectors of the user and the project, provides more accurate recommendation, and improves the recommendation performance. Therefore, the method provided by the invention realizes a relatively accurate prediction effect.

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (7)

1. The graph nerve collaborative filtering recommendation method fusing the project time sequence relation is characterized by comprising the following steps of:

step 1: acquiring project data of each user interaction;

step 2: preprocessing project data interacted by a user;

and step 3: converting a data format, dividing a data set, and dividing the data set into a training set and a test set;

and 4, step 4: acquiring user-project interaction data and project-project time sequence data, and constructing a user-project interaction matrix and a project-project time sequence relation matrix and carrying out normalization processing on the project-project time sequence relation matrix;

step 4.1: initializing a user and item embedding matrix, and obtaining the user embedding matrix and the item embedding matrix by adopting a graph-based neural network embedding method;

step 4.2: constructing a project-project time sequence relation matrix, dividing a project time sequence into a plurality of groups of subsequences by adopting a sliding window strategy, constructing a project time sequence relation graph, and deeply depicting the dynamic change of user interest by aggregating project time information characteristics;

step 4.3: constructing a user-project interaction diagram through the preprocessed project data of the user interaction, and constructing a user-project interaction matrix according to the user-project interaction diagram, wherein the user-interacted projects indicate the interests of users;

step 4.4: normalizing the item-item time sequence relation matrix;

and 5: initializing various parameters of a hybrid recommendation model fusing user-project interaction information and project-project time sequence relation, inputting data of a training set into a network for training until the network is converged, and adjusting network hyper-parameters through a comparison test to finally obtain optimal parameters;

step 5.1: building an NGCF-ITS structure, which comprises an input layer, an item-item time sequence relation aggregation layer, an embedded propagation layer and a prediction layer; setting a feature vector of each layer;

step 5.2: initializing neuron states in each layer of a user embedded matrix, an item embedded matrix and NGCF-ITS;

step 5.3: calculating the time sequence dependency relationship of the project, performing feature fusion, embedding the learning project into a vector and transmitting the vector to the next layer;

step 5.4: calculating the high-order connectivity of user-project interaction, continuously learning a user embedded vector and a user embedded vector, performing multi-level embedded propagation, and connecting results;

step 6: and (4) verifying the model effect through the evaluation indexes of the test set data to obtain the performance improvement of the recommendation algorithm.

2. The graph neural collaborative filtering recommendation method fusing item temporal relationships according to claim 1, characterized in that: the step 1 specifically comprises the following steps: and acquiring project data of each user interaction, wherein the project data comprises three data characteristics of user id, project id and time stamp.

3. The graph nerve collaborative filtering recommendation method fusing item time series relations according to claim 1, characterized in that: the step 2 specifically comprises the following steps:

step 2.1: removing abnormal values, and directly deleting records with obvious abnormal values in the data;

step 2.2: users less than 5 in the project data amount of the user interaction are removed.

4. The graph nerve collaborative filtering recommendation method fusing the project time series relationship according to claim 1, characterized in that the step 3 specifically is:

step 3.1: converting a data format, converting an original data sequence into a format of a supervised learning sequence, and sequencing according to ascending sequence of timestamps of user interaction with projects;

step 3.2: dividing a data set, and dividing the historical interactive data set according to the ratio of 8: the scale of 2 is divided into a training set and a test set.

5. The graph neural collaborative filtering recommendation method fusing item temporal relationships according to claim 1, characterized in that: the network hyper-parameter adjustment in the step 5 specifically comprises the following steps: and adopting the BPR as a loss function and Adam as an optimization algorithm, and circulating until the model converges.

6. The graph nerve collaborative filtering recommendation method fusing item time series relations according to claim 1, characterized in that: in the step 5, in order to obtain the optimal parameters, the optimal parameters are obtained and stored by comparing and adjusting the parameters; including learning rate, mini-batch size, L2 regularization coefficients, weight vectors, and bias vectors.

7. The graph nerve collaborative filtering recommendation method fusing item time series relations according to claim 1, wherein the step 6 specifically is:

step 6.1: predicting the articles in which the user is interested by using the trained prediction model on the test set;

step 6.2: comparing all the prediction results with the real results, and obtaining the performance evaluation result of the model by using precision @ N, recall @ N and NDCG @ N as the evaluation indexes of the model;

step 6.3: the final prediction model is used for prediction of the recommendation algorithm.

Images