I3

The scrolling interaction is a pervasive human-computer interaction on smartphones, which can reflect intrinsic characteristics during dynamic browsings. Different from extrinsic statistical measures like frequency of visits and dwell time, intrinsic features underlying scrolling interactions reveal fine-grained implicit feedbacks about user interests. Toward this end, we explore user interest inference by extracting efficient browsing features from scrolling human-computer interactions on smartphones. In this paper, we first analyze browsing traces of 40 volunteers, and find two intrinsic browsing features underlying scrolling interactions, i.e., browsing velocity stability and browsing velocity sequence, which are tightly related to user interests. Inspired by the observation, we propose an Intelligent Interest Inference system, I3, which infers user interests through sensing scrolling interactions during browsings. Specifically, I3 first extracts the two intrinsic browsing features from users' scrolling interactions. Then, I3 applies a Naive Bayesian-based approach to construct an interest discriminator for coarse-grained user interest (i.e., like-preferred or dislike-preferred) inference. Furthermore, we develop a deep learning-based approach for I3 to train rating classifiers for fine-grained rating inference in like-preferred and dislike-preferred browsings respectively. Finally, I3 utilizes the interest discriminator and rating classifiers to infer exact user ratings about browsing contents on smartphones. Experimental results under browsing traces of 46 volunteers present that I3 achieves 92.4% overall accuracy in interest inference.