Research On Personalized Recommendation Technology For Sequence Data

With the rapid development of globalization and digitalization,human beings have entered an era of information explosion.In the context of big data,there is a wide range of commodities,news and information are changing every day,advertising information is everywhere,science and technology are advancing rapidly,people’s receptivity is seriously "overloaded",and the explosive growth of information brings great challenges to information consumers and producers.In order to help people solve these problems,recommendation systems have emerged as the times require.Among many recommendation fields,the research of recommendation system for sequential data is widely used in our daily life.The core of sequential recommendation is to predict the next possible behavior of users based on their historical behavior sequences,and recommend relevant items or contents to users.Compared with traditional recommendation methods,sequential recommendation can better explore the evolution of users’ interests and behavioral patterns,so as to predict users’ interests and needs more accurately.Moreover,it can update the recommendation results in real time according to the current behavior of users,so it can handle the real-time recommendation problem.In practical applications,sequence recommendation has been widely used in e-commerce,music and movies,social networks,news recommendation and other fields,providing users with more personalized and accurate recommendation services.Therefore,it is a challenging and meaningful research work to fully exploit the hidden features in sequence data and to extract the user’s interest features from the user’s historical interaction sequences.This thesis presents a comprehensive and in-depth analysis of the recommendation task based on user sequence data,and at the same time,analyzes and researches the existing sequence recommendation methods for the problems of not fully expressing the association relationships in sequence data,complex temporal information and data sparsity when modeling.The details of the research are as follows:(1)Aiming to the problems that existing methods fail to take into account the dynamic transformation of items in temporal patterns and multi-level interdependencies among users when modeling,as well as the inadequate expression capability of feature vectors,we design a sequence recommendation method that fusing collaborative transformation and temporal-aware target interaction networks.First,we construct a user-item bipartite graph and model the user sequences by graph convolutional neural network to obtain the synergistic signals with high-order connectivity to enhance the embedding representation of users and items.Second,we feed the enhanced embedding representation into the sequential model while fusing different forms of temporal information to extract behavioral patterns from different user perspectives,and finally,we enhance the corresponding interaction items by measuring the correlation between historical items and candidate items.Experiments demonstrate that our approach enables diverse user interest modeling and greatly improves the expressiveness of the model.(2)Aiming to the problem that the sparse data in the dataset affects the recommendation effect,we design a feature-based method to enhance the sequence recommendation data.First,we use the clustering algorithm to obtain the embedding distribution of all items,and obtain the loss of spatial distribution by calculating the distance between the items and the cluster centers after clustering,and back propagate this loss function,and we use the gradient value of each dimension to represent the size of the contribution made by that dimension in determining the spatial distribution of the items.Second,we use the gradient values embedded in each dimension to perform data expansion by adaptive means,so that the dimensions with high importance are always retained,while for the dimensions with low importance,we perform random discarding by dropout to obtain enhanced data that retains the important features.Third,a new sampling strategy based on spatial distribution is designed to avoid semantically similar users from being classified as negative samples.Finally,contrastive learning training is performed by positive and negative samples.Experiments prove that the method shows good results on sparse datasets.