Research Of Recommendation Model Based On Multi Fusion Methods

Nowadays,highly focus on the data mining topics,recommender systems is most popular and drawing pervasive attentions from both industry and academia.Among them recommendation is much experimental but suffers from the sparse data and cold start problem.The sparsity of rating data usually leads to the bad performance.Usually,users like to share their experience and opinion on social network instead of rating.Therefore,some works considered the user opinion or user review to design recommender system to alleviate the sparse data problem.Furthermore,with the rapid development of the social media,there are many different types of new or small social network introduced(e.g.foursquare and yelp),a lot of efforts have been made on different types of recommendation system to assist user to find the user interested and related information from the small or newly introduced social networks.Generally,users like to interact with big networks where they find more and interesting information easily(e.g.Twitter and Facebook).For this reasons,small or newly introduced social networks facing lack of information problem due to the less number of user attention to the networks.Some efforts in literature have made to endeavor recommender system for small and newly networks,but attracting more information for small and newly introduced networks from the explicit network in recommendation system has been ignored yet.Therefore,more information for small and newly introduced networks and data scarcity problem,got by link prediction and sentiment analysis respectively will be researched in this thesis.Moreover,recent years,deep learning has been widely used in various fields.It essentially extracts the abstract and adequate feature,learns the complicated map function from a big amount of data.However,as far as we know,there is not yet any work,incorporate the link prediction,sentiment analysis,and deep learning simultaneously,which is namely the recommendation model based on Multi-fusion methods.Therefore,a novel recommendation model proposed for small and newly introduced social networks and its function based on three different sub fusion methods in a series.This thesis mainly focuses on the following contents:(1)For lack of information problem of the small or newly introduced network,alleviate by proposed method AEUI(Attract explicit user information)which is based on the anchor link prediction approach in this thesis.AEUI attract the similar user's information from the source(explicit)network toward target(implicit)network to develop a stronger network for recommendations.The input dataset of the method based on multiples network graphs like(foursquare ‘user experience networks' and Twitter ‘microblogging social networks')and the outcome are predicting anchor users among networks.Further extract the spots and users reviews on the basis of user's and spot of small and newly introduced networks.The results show that AEUI is the cog in the wheel for propose recommendation model.(2)The data scarcity problem usually suffered by traditional recommender system which is based on the five start user rating system.Therefore user reviews sentiments are used in this thesis by proposed URSC(User review sentiment Classification)method.The task of USRC identifies the user's decision from reviews and covert into the user binary rating [1,0] positive and negative sentiments to find users interested and not interested information for recommendation model.The experimental results shows that the dataset is tune after AEUI and the accuracy of proposed method is well oiled.(3)For the Top-K recommendation,First,extract the features vectors from the AEUI method,afterward,identified the user decision from extracted user review feature vector in binary rating vector.Then,these latent features are regarded as the input data of the deep neural network model,which is the last part of the proposed model and is used to predict the rating scores,and called this proposed method Deep Learning based recommendation model(DLRM).The effectiveness of the proposed framework is evaluated against state-of-the-art alternatives and neural networks based solutions.Experimental results over publicly available datasets demonstrate that proposed recommendation model based on multi fusion methods significantly outperform against off baselines methods in the recommendation system.