Search

CN-116452281-B - Sequence recommendation method, device and medium based on hypergraph and self-supervision learning

CN116452281BCN 116452281 BCN116452281 BCN 116452281BCN-116452281-B

Abstract

The invention discloses a sequence recommending method, device and medium based on hypergraph and self-supervision learning, wherein the method comprises the steps of obtaining a sequence recommending data set, carrying out data preprocessing, constructing user history interaction data into a sequence form, carrying out data set division on constructed sequence data, dividing the constructed sequence data into a training set, a verification set and a test set, setting a learnable embedded vector comprising a commodity embedded vector, an attribute embedded vector and their respective hyperedge embedded vectors, inputting the sequence data into a hypergraph neural network, carrying out hypergraph convolution, capturing context information of a sequence, finally learning user sequence representation, and carrying out recommendation prediction. According to the method, the hypergraph neural network is utilized to model global information of the sequence data, polygonal relations in the sequence are captured, real interest preference of a user is better modeled, and a better recommendation effect is obtained. The invention can be widely applied to the field of recommendation systems.

Inventors

  • XU YONG
  • XIE HAIQUAN

Assignees

  • 华南理工大学

Dates

Publication Date
20260512
Application Date
20230320

Claims (8)

  1. 1. The sequence recommending method based on hypergraph and self-supervision learning is characterized by comprising the following steps of: Acquiring a sequence data set of a user on an e-commerce platform, wherein the sequence data set comprises commodity information, user interaction time stamps and attribute information corresponding to commodities; Carrying out data cleaning on the sequence data, and constructing the goods subjected to data cleaning and the corresponding attributes thereof into a sequence structure according to a time sequence; Inputting the constructed commodity and attribute sequence data into a constructed hypergraph model to obtain embedded vectors of a commodity end sequence and an attribute end sequence of a user; performing unsupervised contrast learning on the embedded vector of the commodity sequence and the embedded vector of the attribute sequence; training and learning the sequence recommendation model by combining the supervised loss of the sequence recommendation task and the unsupervised loss of the comparison and learning; calculating the similarity between the embedded vectors of the final learned commodity sequence and the embedded vectors of all commodities in the sequence data set, and recommending the commodity with higher similarity to a user; Inputting the constructed commodity and attribute sequence data into a constructed hypergraph model to obtain embedded vectors of a user in a commodity end sequence and an attribute end sequence, wherein the method comprises the following steps: setting a learnable superside embedded vector at a commodity end and an attribute end; Constructing a hypergraph matrix by using the commodity sequence embedded vector and the hyperedge embedded vector to obtain a first correlation matrix of the learnable commodity and the hyperedge; Constructing a hypergraph matrix by using the attribute sequence embedded vector and the superside embedded vector to obtain a second association matrix of the learnable attribute and the superside; According to the obtained first incidence matrix and the second incidence matrix, stacking a plurality of layers of hypergraph networks, performing hypergraph convolution, and obtaining high-order information on a sequence to obtain a final embedded vector of a commodity sequence and an attribute sequence; The expression of the first incidence matrix is as follows: The expression of the second incidence matrix is as follows: Wherein, the A vector matrix is embedded for the superside of the commodity end, Is an embedded vector matrix of the commodity sequence, Is an embedded vector matrix of supersides of attribute ends, Is an embedded vector matrix of attribute sequences.
  2. 2. The sequence recommendation method based on hypergraph and self-supervised learning according to claim 1, wherein the performing data cleaning on the sequence data comprises: Sorting the user interactive commodities and the corresponding attributes thereof according to the time stamps to obtain commodity sequence data and attribute sequence data which are far and near in time; And deleting the commodity with low occurrence frequency and the attribute in the sequence data.
  3. 3. The sequence recommendation method based on hypergraph and self-supervised learning according to claim 1, further comprising the step of processing the sequence data into an equal length form after the step of constructing the washed commodity and its corresponding attribute into a sequence structure in time order: filling a shorter sequence into a predicted sequence length by adopting a padding mode; For longer sequences, the method of interception is adopted, and only commodities with relatively close time stamps and corresponding attributes thereof are adopted.
  4. 4. The sequence recommendation method based on hypergraph and self-supervised learning according to claim 1, wherein the performing unsupervised contrast learning on the embedded vector of the commodity sequence and the embedded vector of the attribute sequence comprises: Taking the commodity sequence and the attribute sequence as two views for comparison learning; InfoNCE loss, taking the embedded vector of the commodity sequence and the embedded vector of the corresponding attribute as a positive sample pair, and taking the embedded vector of the commodity sequence and the embedded vector of other attributes as a negative sample pair, and performing unsupervised comparison learning.
  5. 5. The sequence recommendation method based on hypergraph and self-supervised learning of claim 1, wherein the supervised loss of joint sequence recommendation task and the unsupervised loss of contrast learning train and learn the sequence recommendation model, comprising: summing the supervised loss of the sequence task and the unsupervised loss of the comparison learning to serve as a target loss function of the recommendation model; And updating parameters of the model by adopting a gradient descent method until the loss of the objective function reaches a set threshold value.
  6. 6. The sequence recommendation method based on hypergraph and self-supervised learning according to claim 1, wherein the expression with supervised loss is: Wherein, the In order to activate the function sigmoid, A positive sample is represented and a positive sample is represented, A negative sample is represented and is shown, An embedded vector representing the positive sample commodity, Representing the embedded vector of the negative-sample commodity, A final embedded vector representing the sequence of goods; the expression of the unsupervised loss is: Wherein, the As a function of the cosine similarity, Is a temperature coefficient of the silicon carbide material, A final embedded vector representing the sequence of attributes corresponding to the sequence of items, A final embedded vector representing a sequence of attributes to which the sequence of items does not correspond, Representing a negative-sample attribute sequence set, Representing all the commodity sequences.
  7. 7. A sequence recommendation device based on hypergraph and self-supervised learning, comprising: At least one processor; at least one memory for storing at least one program; The at least one program, when executed by the at least one processor, causes the at least one processor to implement the method of any one of claims 1-6.
  8. 8. A computer readable storage medium, in which a processor executable program is stored, characterized in that the processor executable program is for performing the method according to any of claims 1-6 when being executed by a processor.

Description

Sequence recommendation method, device and medium based on hypergraph and self-supervision learning Technical Field The invention relates to the technical field of artificial intelligence, deep learning and recommendation systems, in particular to a sequence recommendation method, device and medium based on hypergraph and self-supervision learning. Background Today's world Internet technology is rapidly evolving and huge amounts of data are being generated on the network from moment to moment. For an average user, such a huge amount of data makes it difficult for them to select content of interest to themselves. For the application service platform, the operation strategy is to recommend the content of interest to the user as much as possible, so that the user is reserved to the greatest extent, and the daily activity of the platform is improved. The recommendation system can model massive user data and commodity data, mine interests of users, push commodities or contents interested by the users to the users, and relieve information overload. The recommendation system is therefore an integral part of the current network application platform. Sequence recommendation is an important branch of the field of recommendation systems. On online shopping platforms like naughty, jingdong, etc., the behavior of users is dynamically changing over time. Modeling the continuous behavior of a user is necessary to better capture the user's interest changes and make accurate recommendations. Therefore, the sequence recommendation study has received extensive attention from researchers in recent years. Most of the existing sequence recommendation methods are based on deep learning methods, including the use of cyclic neural networks, convolutional neural networks, self-attention mechanisms, and the like. The self-attention mechanism method can well capture the weight relation of the sequence context, mine the correlation inside the sequence, and therefore the interest preference of the user on the sequence data is extracted. Therefore, most of the sequence recommendation methods of the current optimal methods take the self-attention mechanism model as a main body model, and then conduct expansion research. However, the self-attention mechanism can only model paired information in a sequence when modeling sequence data, namely can only calculate the influence of one item on another item, and cannot model the influence of a plurality of items on one item at the same time. In view of this problem, a hypergraph neural network is used to globally model the sequence data, aggregate a plurality of items into one block, and calculate the polygon relationship. In addition, sequence recommendations often suffer from sparsity issues such as sequence lengths that are too short. For such short sequences, it is difficult for the model to capture enough information from it that the final learned representation of the sequence is sub-optimal. Currently, self-supervised learning is in computer vision. The fields of natural language processing, graph representation learning and the like have good effects. The self-supervision learning uses a data enhancement method to directly generate supervision signals from original data, constructs auxiliary tasks and performs combined training with a main task to play a role in relieving data sparseness. Disclosure of Invention In order to solve at least one of the technical problems existing in the prior art to a certain extent, the invention aims to provide a sequence recommendation method, device and medium based on hypergraph and self-supervision learning. The technical scheme adopted by the invention is as follows: A sequence recommendation method based on hypergraph and self-supervision learning comprises the following steps: Acquiring a sequence data set of a user on an e-commerce platform, wherein the sequence data set comprises commodity information, user interaction time stamps and attribute information corresponding to commodities; Carrying out data cleaning on the sequence data, and constructing the goods subjected to data cleaning and the corresponding attributes thereof into a sequence structure according to a time sequence; Inputting the constructed commodity and attribute sequence data into a constructed hypergraph model to obtain embedded vectors of a commodity end sequence and an attribute end sequence of a user; performing unsupervised contrast learning on the embedded vector of the commodity sequence and the embedded vector of the attribute sequence; training and learning the sequence recommendation model by combining the supervised loss of the sequence recommendation task and the unsupervised loss of the comparison and learning; and carrying out similarity calculation on the embedded vectors of the final learned commodity sequence and the embedded vectors of all commodities in the sequence data set, and recommending the commodities with higher similarity to the user. Further, the dat