Personalized Recommender Systems With Diversified Data

Recommender systems(RSs)are a subclass of information filtering systems which aim at predicting users' preference for items,and then recommending a few items that users would like.With the boosting of online services,the volume of available items on the Internet has been keeping growing exponentially,and users may get lost because of the information overload.Thus RSs play an important role in directing customers to their favourite items,and how to build more precise RSs has become research hotspot.Traditional RSs usually have one or more following problems:1)Cold start problem.Algorithms can not effectively learn behaviour patterns for new-join users or inactive users;2)multi-domain data fusion.When the dataset contains features from multiple sources,simply treating these features equally can not lead to the best result;3)gen-eralization problem.For example,some classical models such as Logistic Regression and Gradient Boosting Decision Tree can not generalize to unseen feature combina-tions in the training set.In this dissertation,we study how to address these challenges via effectively leveraging the big data on the Internet,which usually contains diversified information,including multi-source data,heterogeneous data,multimoal data and so on.We aim to improve traditional recommender systems via considering these diversified information.Specifically,we make the following contributions:(1)We observe that,in the real world,customers usually interact with multiple products.Researchers can leverage multiple-domains learning methods which can trans-fer knowledge across different products,thus the cold-start problem in recommender systems can be alleviated.In this thread,we propose two cross-domain models,namely the Content-boosted Collaborative Filtering Neural Network(CCCFNet)and the Mul-tifaceted Model for Cross Domain RSs(MCDRS).CCCFNet combines collaborative filtering and content-based filtering in a unified framework,and further embeds this framework into a multi-view neural network.Different from the CCCFNet,the MC-DRS contains a selective learning mechanism to overcome the potential inconsistency problem across different domains.(2)Recently,an increasing number of researchers are interested in employing deep learning techniques for RSs.Multi-layer fully connected neural networks are widely used in the current RSs literature.However,due to feature sparsity,heterogeneity and diversity,a plain MLP is hardly the best model.Thus we propose a novel deep fu-sion model(DFM)which comprises of two key components,i.e.the inception module which learns latent representations via leveraging various subnetworks in parallel,and the attention mechanism which contextually fuses diverse data across multiple channels.The inception componnet can learn better representation than a plain MLP because it can recover latent features from both shallower and deeper interactions of raw features.The attention mechanism can reduce the domain diversity and discrepancy problem.Through comprehensive experiments we have demonstrated that,on one hand,DFM learns good representations for users/items which are effective for candidate retrieval;on the other hand,when incorporated with a wide and deep part,the new model can provide superior performance for item re-ranking.(3)The most successful applications of deep learning techniques are in computer vision,speech recognition,and natural language understanding.The most popular neu-ral modules,i.e.the Convolutional Neural Network(CNN)and the Recurrent Neural Network(RNN)are designed based on the properties of images or sequence of words.However,for web-scale recommender systems,the input features are usually sparse,categorical-continuous-mixed,high-dimensional,and without local correlation.Thus,for many related works,the research focus is not to leverage deep learning techniques to learn elaborate representation from features,but to learn high-order feature interactions automatically based on the framework of factorization machine.In this research direc-tion,we propose a novel model named extreme deep factorization machine(xDeepFM).A multi-layer neural network only learns high-order feature interactions in an implicit fashion,and there is no theoretical conclusion on what the maximum degree of fea-ture interactions is.In contrast,xDeepFM has a component to explicitly learn certain bounded-degree high-order feature interactions.On the other hand,xDeepFM can learn feature interaction in a vector-wise fashion,which is more theoretically sound than a bit-wise fashion.We conduct comprehensive experiments on three real-world datasets,and the results demonstrate that xDeepFM outperforms state-of-the-art models.To summarize,in this dissertation we introduce our research on heterogeneous information based recommender systems,from the angle of cross-domain knowledge transferring,multi-source information fusion,and high-order diverse feature interac-tions.In addition,we also introduce some feature engineering techniques for extracting information from heterogeneous data.Since the recommender systems in industry usu-ally need to handle heterogeneous data,our research will be meaningful for modern recommender systems both in theory and practice.