Data Mining Techniques For Online Users’ Consumption Behavior

With the development of IT related industry, especially the blooming of the (mobile) Internet and the online social media, the worlds of commerce and Inter-net technology become more inextricably linked. All the online service providers have realized the great research and marketing values of the massive online users’ consumption footprints. Accurately mining and understanding online users’con-sumption behavior provides an unprecedented opportunity in many applications, such as targeted marketing and advertising for online service providers, person-alized services for customers, etc. However, existing online user consumption behavior algorithms suffer from challenges such as the data sparsity and hetero-geneity, the complex consumption decision process and the multi-discipline nature of users’decision behavior. To that end, in this dissertation, by leveraging social theories from various disciplines, we make a focused study on designing models for mining and understanding users’ consumption behavior. To be specific, the contributions of this dissertation can be summarized as follows.Firstly, we propose consumption behavior understanding and product rec-ommendation algorithms based on user interest modeling. The underlying as-sumption for recommender systems is that a user’s future consumption behavior is motivated by her historical consumption interests. Thus user interest modeling is the key point of recommender systems. To deal with the data sparsity and cold-start problem in user interest modeling process, we propose a two-stage col-laborative filtering framework NHPMF. NHPMF utilizes the user neighborhood and item neighborhood information from user-generated tags, and then incorpo-rates the local neighborhood information into matrix factorization process. Thus it has the complementary advantages of two main kinds of algorithms in collabo-rative filtering (local neighborhood-based methods and global matrix factorization models). Experimental results on two real world datasets validate the proposed model performs much better than baselines with regard to recommendation accu-racy. On the other hand, as the traditional product recommendation algorithms are successful at providing accurate recommendations that match some of a user’s dominant interests, but fail to provide diversified recommendations that cover all of a user’s interests. We also provide a REC framework to promote the diversity measure in user interest modeling and product recommendation. Specifically, the REC based framework first introduces a coverage notion to measure the usefulness of a whole set, and then designs an efficient algorithm to simultaneously optimize the relevance based measures in traditional collaborative filtering and the new coverage measure. Experimental results on three real world datasets demonstrate that REC based models can provide more diversified recommendations without losing accuracy.Secondly, we provide online users’ consumption behavior understanding and product prediction from social networks. A key characteristic is that due to the information diffusion in social networks, users’ behaviors are not isolated but cor-related with social environments. We collect a large scale social networked users’ mobile phone adoption data from a leading microblogging service platform. By leveraging the social theories in user decision making, we identify three key fac-tors for product adoption prediction: i.e., users’ personal interests, the homophily effect and the social influence effect in social networks. We propose a SHIP mod-el to automatically learn the relative contribution of each factor that underlines people’s smartphone adoption. Experimental results show the superiority of the SHIP model for smartphone adoption. Also, compared to the traditional binary representation of users’ product adoption status, we introduce a product adop-tion rate measure to accurately describe users’ adoption status and commitment to a product over time. We introduce a decision function to capture the various factors that may influence users’decisions, and further propose the generalized and personalized assumptions to model users’ unique preferences in balancing various factors in the decision function. Experimental results on two real-world datasets show the improvement of our proposed models in product adoption rate prediction.Lastly, we provide to analyze the evolution of users’ consumption behav-ior (user-item interaction) and social link behavior (user-user interaction) in so-cial networking services (SNS). In fact, researchers have long converged that the evolution of SNSs is driven by the interplay between users’ two kinds of behaviors, with both behaviors changing over time. By leveraging the underlying social the-ories (i.e., social influence and the homophily effect), we propose a probabilistic approach to fuse these social theories for jointly modeling users’temporal behav-iors in SNSs. Experimental results on two real-world datasets demonstrate the effectiveness of our proposed model in both user consumption behavior prediction and social link behavior prediction.