Heterogeneous Graph Attention-based Multi-modal Personalized Recommendation

In the era of data explosion,the huge amount of mixed information on the Internet makes it increasingly difficult to filter the required information.Recommendation system has developed as an efficient way to filter information.By effectively mining massive amounts of data,it can recommend information that matches the user’s unique attributes and uncover potential interest,helping to improve the user’s web experience and achieve commercial realisation.The core of personalised recommendation system lies in accurately capturing user preferences.Although recommendation algorithms constantly improve,most of them suffer from data sparsity,neglecting the attributes of users or items,and ignoring the complex relationships generated during user-item interaction.This thesis focuses on the above challenges and develops a recommendation algorithm based on heterogeneous graph attention and multi-modal content,with four main parts of work.(1)A formal modelling approach based on heterogeneous graphs is proposed to maximise the retention and efficient use of information such as multiple interaction behaviours generated during user-content interaction and multiple attribute associations between contents.(2)A feature mining and fusion mechanism for complicated usercontent relationships is proposed to extract and convey the complex information generated during multi-behavioural interactions between users and content.This fusion mechanism adds timestamps and evaluates the importance of the interactions,and finally fuses various types of information to obtain feature embedding representations of user-contentside nodes and some content-side nodes.(3)A multi-modal content-side feature extraction algorithm based on heterogeneous graph attention mechanism is proposed to efficiently process the content-side heterogeneous graph and achieve the propagation and aggregation of messages over the whole heterogeneous graph to accomplish the personalised recommendation task.The model is trained by updating the user-side and content-side node representations to fuse the full domain information and perform computation,integrating the loss function,and finally outputting the Top-K personalised recommendation results ranked by the inner product.(4)Integrate the above work and finally transform it into heterogeneous graph-based personalised video recommendation prototype system,to clearly demonstrate the model recommendation results and present the personalised recommendation results obtained from the model training with front-end interface.