| Article ID | Journal | Published Year | Pages | File Type |
|---|---|---|---|---|
| 407197 | Neurocomputing | 2016 | 7 Pages |
As the amount of online multimedia keeps growing rapidly, study of online user behavior is highly demanded in many applications to provide better user experience and online services. Due to the surging online user population in recent years, more and more online contents are now associated with advertisements. For marketers, online advertising can reach broader audience with reduced cost. Meanwhile, the revenue for many IT companies such as Google mainly comes from advertisement hosting. Therefore, for both marketers and advertisement hosts, effective online advertising is the primary goal. Thus, it is of great interest to predict and prevent the users from zapping, i.e., skipping the advertisement. Reduced zapping would increase the possibility for the user to become potential consumer. Despite its importance, zapping prediction for online advertising has received very limited attention. Since the zapping behavior is related to the user׳s emotional states, in this paper, we propose to predict zapping from user׳s facial expressions. Our prediction is non-intrusive, meaning that the user׳s experience during advertisement watching is not disturbed. Specifically, a robust encoding of the smile responses termed as Cumulative Smile Sparse Representation (CSSR) is extracted from the user׳s facial expressions. Then, this representation is incorporated in a semi-supervised hypergraph learning framework to predict the moment-to-moment zapping behavior. The hypergraph is able to discover both pairwise and higher order data relationships, leading to more accurate prediction as compared to conventional classifiers. Experiments are conducted on a recently published zapping behavior analysis dataset and state-of-the-art performance is achieved as compared to other competing methods.
