User Behaviors in Social Services: Models and Applications
Peifeng Yin · 2014
The Web 2.0 prompts the appearance of many online social services. They serve as a platform where users can interact with each other by sharing content as well as providing feedbacks, e.g., ``like'' and ``dislike'' to others' shared contents. The large quantity of user behavior log offers a potential good source to study and model user behaviors. And the research results can help the development of multiple applications such as popular item highlighting, personalized recommendation. These applications not only improve user experience but also increase engagement. However, there are several issues in converting raw logs into useful applications. We summarize them as i) timeliness requirement, ii) silent user, iii) item conflict and iv) non-textual content. The first one is related to some applications characterized by timeliness, i.e., the earlier the better. Services providers usually highlight in the homepage the popular shared contents, i.e., those receiving a large quantity of ``like''. We propose the predict the popularity of the shared content at an early time based on its early vote situation. Specifically, we model two personalities of voters, defined as \emph{conformer} and \emph{maverick}. The former one represents the probability that the voter's opinion conforms with the majority while the latter one stands for the inconsistence with the majority. By learning the two personalities of each person via her voting history, we may quickly predict the potential popularity of each item based on the early votes. The second issue concerns the user's behavior of silent viewing. Many applications, especially personalized recommendation, needs the explicit opinion of people on items to model user preference. However, if a user only views an item without clicking ``like'' or ``dislike'', the current technique can not exploit it. We argue that the reason of silent viewing is that the quality of the item is not good enough to motivate the person to vote. However, the time she spends on that item suggests her opinion. We propose a viewing-voting model to capture the user's dwell time and voting behavior. We demonstrate an improved performance for conventional recommendation techniques when combined with VV model. The third issue represents the blind spots of current personalized recommendation techniques. Conventional recommended items such as books, music, movie have no conflict among them, i.e., current consumption of one time has no impact on future ones. Other items may have. For example, if a user buys a TV, it is unlikely that she may buy another one in near future. We thus propose a actual-tempting model to capture such conflict relationship between items. Experiments running on a dataset of mobile app installation demonstrate that the AT model outperforms conventional recommendation techniques. And optimal performance is achieved for hybrid methods.