A heterogeneous hidden Markov model for mobile app recommendation

Recent proliferation of ubiquitous smart phones has led to the emergence of a wide variety of apps. Selecting apps through keyword search or recommendations from friends or social networks (e.g., Facebook) may not match the real preferences of users, especially when the need is just-in-time and context specific. Although there are many collaborative filtering approaches that are capable of generating time-aware recommendations, most of them work on modeling of the time stamps (the time that events happen) rather than modeling of the sequential patterns (in cases that time stamps are not available) as well as investigating the factors behind those patterns. In this paper, we propose a mechanism for modeling three important factors governing the app installation of smart phone users: (1)short-term context, (2)co-installation pattern, and (3)random choice. Specifically, we use a hidden Markov model equipped with heterogeneous emission distributions to incorporate these factors. Apps being installed are probabilistically categorized into one of these factors, and app recommendations for users are carried out accordingly. This coherent model can be inferred effectively by using Gibbs sampling. The formulation has a significant advantage that the performance is less sensitive to data sparsity and incompleteness. Empirical results show that it has higher performance in recommending mobile apps to smart phone users, measured in terms of precision and area under the ROC curve (AUC). Besides, the proposed model allows the nature of the apps, with respect to the three factors, to be revealed as well as the extent to which each user is affected by the three factors to be inferred, providing additional insights on the users' behavior.