Micro-Video Recommendation on Social Media

Abstract

With the rapid development of social informatization, people are becoming accustomed to using social media platforms to record micro-videos to share life. However, due to the growth of Internet users, the user-centred information production model has caused an explosive growth of Internet information; people are suffering an increasingly severe problem of ‘’information overload’’. The ‘’Information overload’’ problem means that people cannot quickly and accurately find the required information from a large amount of information. Massive data reduces the efficiency of user information retrieval. Recommendation system, as an essential measure of information filtering, is an effective way to solve the problem of information overload.Although traditional recommendation algorithms have been successfully applied in commercial recommendation systems (e.g., Netflix, Amazon), they are incapable of micro-video recommendation scenarios. For example, in collaborative filtering, it only leverages the user's personal preferences and historical interaction information to mine user's preferences. Unlike long videos, micro-videos contain multi-modality information, such as social attributes, video descriptions, visual and audio, etc. How to effectively extract and represent the features from the heterogeneous data to boost the micro-video recommendation is a critical challenge in our research. Besides, the data sparseness and cold start problems also limit the performance of micro-video recommendations.This thesis focuses on the multi-modality data representation from micro-videos, intending to boost the performance of micro-video recommendation. In particular, we developed algorithms to address issues identified from micro-video data, such as feature extraction, fine-grained categories classification, social relations mining, and data sparseness problem, etc.The four main contributions of this research are as follows. First, we developed fine-grain genres of micro-video recommendation model called latent genre aware (LGA), which jointly analyze the context and visual contents of micro-video to facilitate the recommendation. Second, we considered the impact of social factors on micro-video recommendations. Third, we proposed a novel recommendation algorithm to tackle the data sparseness problem. Fourth, we proposed a novel model named RecKGC that generates a completed knowledge graph and recommends items for users simultaneously to address the problems of recommendation together with knowledge graph completion.As of now, there is little research on the micro-video recommendation. This research is a step forward in recommending micro-video based on social media. Mainly, it explores multi-modality data mining, modelling and addresses related challenges. In the era of big data, abundant and diverse multimedia data will lead to more powerful recommendation systems to facilitate people's lives.