Research On Rating Prediction Models For Recommender System

With the advent of the era of mobile internet,a tremendous change of people's consum-ing habits and their information acquisition approaches is taking place.Recommender systems have been proved to be effective in helping people discover useful information,and are acquir-ing increasing attention from commercial and portal web sites.However,as the diversity of information grows on the internet and the quantity and scale of recommendation tasks increases,we are still facing many key problems and technical challenges that solicit effective solutions.This thesis mainly pays attention to the problems of insufficient utilization of diverse informa-tion by current recommender systems and the lack of knowledge accumulation and utilization approaches during the process of recommendation tasks.Based on the existing methods and previous work,this thesis mainly studies three types of important information mining problems in building recommender systems,which are jointly mining temporal dynamics in ratings and reviews,using review text to help mine visual dynamics in rating prediction tasks,and the meth-ods of knowledge accumulation and utilization in multi-task scenarios.The detailed research content and novelty are summarized as follows.In terms of jointly mining temporal dynamics in ratings and reviews,this thesis focuses on exploiting temporal dynamics in ratings and reviews that reflect the interim and intrinsic features of items,respectively,and proposes a temporal rating model with topics that jointly mines the temporal dynamics of both user-item ratings and reviews.Studying temporal drifts in reviews helps us understand item rating evolutions and user interest changes over time.Our model also automatically splits the review text in each time period into interim words and intrinsic words.By linking interim words and intrinsic words to short-term and long-term item features respectively,we jointly mine the temporal changes in user and item latent features together with the associated review text in a single learning stage.Through experiments on 28 real world datasets collected from Amazon,we show that the rating prediction accuracy of our model significantly outperforms the existing state-of-art recommender system models.And our model can automatically identify representative interim words in each time period as well as intrinsic words cross all time periods.This can be very useful in understanding the time evolution of users' preferences and items' characteristics.In terms of mining visual information and review text in rating prediction tasks,this the-sis studies discovering item features which match the fashion trends by jointly mining visual dynamics and review text,and presents visually-aware temporal rating model with topics using review text to help mine visual dynamics as well as non-visual features for rating prediction task.Understanding the reviews will help the recommender system know whether a user is attracted by the appearance of an item,and which aspect of an item's appearance contributes most to its ratings.To achieve this,we incorporate the visual information into the rating predicting func-tion and introduce a topic model that can automatically classify words in an item's reviews into non-visual words that explain the coherent feature,and visual words that are associated with its visual appearances in each time period,respectively.In terms of knowledge accumulation and utilization in rating prediction tasks,this thesis studies the approaches of building recommender systems in a lifelong machine learning manner.First,this thesis proposes a lifelong learning rating prediction model with outstanding appli-cability and scalability.The model is designed for multi-task scenarios in a lifelong machine learning manner.The only required data during the training of our system is users' ratings.It can accumulate knowledge from previous tasks and utilize it in future tasks to help improve the recommendation performance.We build a Knowledge Base for our model to store the knowl-edge gained from previous tasks.Then a specific rating model is developed to incorporate the knowledge into the training process.Second,this thesis presents a lifelong learning model that jointly mines ratings and topics.It can accumulate knowledge of words and topics from previous tasks and utilize it in future tasks to help improve the rating prediction accuracy.