Deep Learning And Multi-modal Learning Based Multimedia Recommender Ranking System

Benifiting from wide spread of the Internet and mobile devices,e-commerce,short-video or news feed has become a choice for the majority.To overcome the problem of"information overload",recommender systems have become a core technology.In this thesis,we study the multimedia recommender systems,focusing on deep learning and multi-modal learning for solving user and item cold start problem and data defficient rec-ommender systems.The contents are divided into three parts as follows.First,item cold start problem.When a new item is added into a recommender sys-tem,recommendation model cannot predict which user may be interested in the item due to lack of previous interaction information.Previous studies didn't make full use of items'multimedia information such as titles,descriptions or pictures.These information can be used to model the content of item and predict which user could be interested in.Compared with basedlines,this thesis proposed an adversarial learning based multi-modal learning method,weighting higher for content information of cold start items.The method im-proves the recommendation accuracy of cold start items.Evaluation on two benchmark datasets show the effectiveness of the method.The Hit Ratio and NDCG improve by over 14%.Second,User cold start problem.When a new user is added into a recommender system,recommendation model cannot offer customized recommendation due to lack of historical interactions.In this thesis,we proposed a Hybrid Interest Modeling(HIM)Net-work,using user-group similarity and group-item preference to denote the preference be-tween the user and the item.We also proposed an end-to-end training framework for HIM,which can learning user clustering seamlessly.As an user modeling framework,HIM can include any state-of-the-art recommender algorithms.Evaluation on two pub-lic benchmark datasets and one industrial real-world recommendation engine show the effectiveness of this method.Third,the multimedia recommendation algorithm in a small data volume scenario.Although deep learning can surpass traditional methods when the amount of data is suffi-cient and the labeling is accurate,decision trees may be more suitable when the amount of data is insufficient.In this thesis,we proposed a new embedding method called TSE for session-based recommendation.Given the user's watch list(session)and target items,TSE embeds each item in the conversation j ointly,taking the relative position and relation-ship between all items in the conversation into account.This embedding method allows the recommendation system to weight more on recent viewing items and tap the user's long and short interests.At the same time,multi-head attention is used to embed the sequence and then perform nonlinear transformation.After embedding,the recommendation comes down to a classification problem.Through the innovative use of gradient boosting deci-sion trees to process neural network embeddings,TSE has achieved better performance.The TSE method won the championship in the 2019 ACM Multimedia Conference and the content-based video relationship prediction challenge hosted by Hulu.