Semantic-Guided Multi-Scene Recommendation System

The rapid development of the Internet has provided users with a diverse range of information content,but at the same time,it has also buried users in a vast amount of data,making it difficult for them to quickly and accurately find their areas of interest.In order to alleviate the problem of information overload,researchers have designed a series of intelligent recommendation systems to help users discover their interests within the massive amount of information.However,existing recommendation system technologies overlook the background factors that may influence user preferences,such as the time,location,weather,and the diversity of user behaviors when interacting with products,and they neglect the presence of noise and biases in user’s historical behavioral interaction data.As a result,the learning of user preferences is not comprehensive,there are deviations from the user’s true preferences,and this in turn affects the effectiveness of the recommendations.In recent years,major e-commerce platforms have introduced the concept of "scenes" to provide more granular guidance for inferring user preferences in different situations.A scene is a combination of semantic information factors,such as time,location,and different types of user interactions.Such combinations often represent specific intentions of users in certain circumstances.To further explore users’ recommendation needs in different scenes and extract their true intentions from implicit feedback data,providing accurate and efficient personalized recommendation services,this thesis focuses on the semantic information of scenes.It investigates the semantic information of scene tags and the semantic information of users’multiple behavioral interactions within scenes.This thesis proposes a multi-scene recommendation system model based on the semantic correlation of scene tags,namely SHNN(short of Semantic-guided Hypergraph Neural Network),and a multi-scene recommendation system framework based on semantic alignment of users’ multiple behaviors,namely MBA,(short of Multi-Behavior Alignment).SHNN utilizes a large-scale pre-trained language model to establish semantic connections between scene tags and transfers user preferences across scenes using hypergraph convolutional networks.It also designs a self-supervised task to alleviate the problem of insufficient representation caused by imbalanced data between scenes.MBA aligns the semantic information of users’ different behaviors using Kullback-Leibler divergence(KL-divergence)and infers users’ true preferences through multiple types of user behavioral data.It achieves effective knowledge transfer between different behavioral preferences and users’ true preferences while performing data denoising.In the task of multi-scene recommendation,due to the scarcity of relevant datasets,this thesis collected a large-scale multi-scene dataset from a Chinese e-commerce platform and conducted extensive experimental analyses.In the task of multi-behavior recommendation,two publicly available benchmark datasets and a large-scale multi-behavior dataset collected from the Chinese e-commerce platform were used for extensive experimental analyses.Compared to existing methods,SHNN and MBA demonstrated significant improvements in the tasks of multi-scene recommendation and multi-behavior recommendation,respectively.This thesis also designed relevant analytical experiments to explain the impact of scene tag semantic information and user’s multiple behavior semantic information on multi-scene recommendation systems.