User Capability Mining With Behavior Analysis

For modern users, there is a clear trend to augment the physical devices/objects with sensing, computing and communication capabilities in society. Besides, many companies have established plenty of advanced information management and opera-tion systems, which make users' life and study more convenient and efficient. Along this line, the digital traces left by users while interacting with cyber-physical spaces are accumulating at an unprecedented breadth, depth and scale, and we call all those traces the "digital footprints". The explosive growth in the amount of available data has created an opportunity for us to automatically analyze users' behavior using novel information technologies. Furthermore, users behavior analytics becomes important for both researchers and developers, because it can reveal the patterns of users behaviors in individual, group and societal scales, and thus enables a variety of applications.Unlike classic recommender systems, some platforms should take users' capability into account, because of unique characteristics. For example, if one want to buy a prod-uct, he/she probably not just favors it but also has such purchasing power. Indeed, in the decision process of recommendation, we should not only take the personalized pref-erence into account, but also consider the degree of matching their psychologic/physical factor to the recommended item.In this work, we firstly mathematically model users in the real world. Multiple aspects of users' behaviors have been analyzed. We collect users' consumption data,trajectory data, etc. Based on the user understanding, we also consider users' ablility,e.g. purchasing power, in our methods, because it usually plays an important role of behavior analysis in many senerios. Therefore, we take both users' personalized pref-erence and competence into account, such that we could improve users' experience and recommendation performance. For a given user, we aim to quantify the matching de-gree between user and the available products, such that we could prove the effectiveness of the proposed methods.In conclusion, the main contribution and innovation of this dissertation can be sum-merized as follows:· Based oil consumption data,Inet usage data and trajectory data, from the pur-chasing power perspective, we propose techniques for identifying users who are qualified to obtain the funding support. Specifically, we investigate users' com-plex behavior within an area from multiple perspectives, and develop a learning framework by jointly incorporating the heterogeneous features to predict the port-folio of stipends the given users should be rewarded. Our framework formalizes the above problem as a multi-label learning problem. Along this line, we first extract discriminative features from three perspectives: (i) smartcard usage be-havior, (ii) internet usage behavior and (iii) trajectory within an area.· From the psychology factor perspective, we try to discover potential users of new released products, which plays an important role in marketing campaigns, espe-cially for companies have many established products and users. Intuitively, one of the straightforward solutions is to directly recommend all the users of histor-ical products to the new product. However, according to our real-world obser-vations, users often have different forms of acceptance to the different products.Therefore, it is appealing to estimate users' personal preferences for making rec-ommendation. We adopt transer learning to leverage the user data from historical products (i.e., auxiliary domains) for recommendation of new product (i.e., target domain).· Based on audio records, from the singing perspective, we propose a method to rec-ommend songs which could highly match users' vocal competence. Unlike clas-sic music recommendation, online karaoke has unique characteristics. In karaoke systems, users may receive high ratings, if his vocal competence could meet the requirment of songs. When choosing karaoke songs, users often care about whether their vocal competence meets the vocal requirements of songs. With the development of computational acoustic analysis, we first extract multi-aspect vo-cal ratings. After preprocessing the karaoke records, we obtain the audio records which encode users' vocal performance. At last, we exploit a non-negative ten-sor factorization method to model the generative process of vocal ratings Karaoke recommendation can help these users identify appropriate karaoke songs, receive high ratings, and moreover, improve karaoke experience.