Research On Recommendation Methods Over Diversified Data

In the era of information overload,recommender systems have been an inalienable apart of many online web applications,which have a profound impact on our daily lives.For instance,the product recommendation services in modern E-commerce websites assist customers in identifying desired products,point-of-interest recommendation services in social media platforms help users to find potentially interesting places,micro-video recommendation services in micro-video sharing applications help users to identify favourite micro-videos,music recommendation services in online music websites help users to generate the personalized playlists that match their tastes.In a way,recommender systems are everywhere.An effective recommender system can significantly improve user experiences and help business owners to earn more profits.In recent years,with the evolution of internet technology,the format and structure of the data in recommender systems have been much more diversified.The data,which has main characteristics of multi-source,heterogeneous,multimodal and temporally dynamic,is playing an important role.The immediate problem to be solved in recommender systems is how to deal with the complex diversified data effectively,which is also the research trends in the area of machine learning.Aiming to remedy the limitations of most existing recommendation methods over diversified data,this dissertation exploits the challenging research problems in aspects of long-tail distribution,data domain-awareness,data heterogeneity and temporal dynamics.As such,we put forward the corresponding effective and efficient methods for the tasks of rating prediction and item ranking.The main contributions of this thesis are summarized as follows:1)To settle the problem that it is insufficient to optimize the relative order among products only from user side,we propose a co-pairwise ranking recommendation method(CPR).Specifically,by integrating user-item interactions and item-item complementarity relationships,we exploit both user-and product-side pairwise rankings to optimize the ordered list of a given user.Considering the quality of the negative sample has a direct impact on the accuracy and the convergence rate of pairwise model,we devise a position-aware sampling strategy for the proposed method.To solve the data imbalance problem,we further design an alternative optimization algorithm to efficiently learn the model parameters.Extensive experiments on several real-world datasets demonstrate that CPR significantly outperforms other state-of-the-art pointwise and pairwise methods.Meanwhile,the co-pairwise ranking mechanism is capable of alleviating the long-tail problem.2)Motivated by the phenomenon that user's attentions would vary in different product domains,we propose a novel recommendation method,which integrates product's intrinsic and extrinsic characteristics in a unified manner.Specifically,we focus on fine-grained modeling of product characteristics to improve recommendation quality,and then contribute an intrinsic-extrinsic probabilistic matrix factorization(IEPMF).However,a downside is that such an expressive model poses efficiency challenges in learning model parameters with traditional optimization schemes.To make our method suitable for practical use,we develop a fast alternating least squares learning algorithm for the IEPMF model.Finally,we design an online updating strategy to adapt IEPMF to real-time streaming data.Extensive experiments show quantitatively that IEPMF significantly outperforms competing baselines,and qualitatively that it is capable of making meaningful recommendations.3)To cope with the problem that how to integrate multi-source heterogeneous data to achieve the maximum value of big data,we contribute a novel recommendation method with multi-source heterogeneous information.First,we provide the detailed definitions of various data among different information sources,and analyze the correlations between user ratings and functional relations.Second,we devised a computationally efficient learning algorithm,named MSRA(multi-source heterogeneous information based recommendation algorithm),to optimize the proposed model.Extensive experiments on four large-scale datasets demonstrate significant improvements of MSRA over a series of state-of-the-art recommendation methods.Meanwhile,our algorithm is capable of providing obtain more accurate results of various items.4)Translation-based recommendation method(Trans Rec)only uses the latest consumed item to model a user's short-term preference,which is insufficient for modeling fidelity.Towards this end,we propose a high-order translation method(HTM)for sequential recommendation.Specifically,we represent each user/item aggregate high-order Markov chains into one representation,we devise a position-aware attention mechanism to capture the varying importance of information at different orders.Extensive experiments on four real-world datasets show that HTM consistently and significantly outperforms other state-of-the-art sequentially non-aware and aware methods,verifying the interpretability of our position-aware attention mechanism by a case study.