Deep Interest Network (DIN) is a state-of-the-art model which uses attention mechanism to capture user interests from historical behaviors. User interests intuitively follow a hierarchical pattern such that users generally show interests from a higher-level then to a lower-level abstraction. Modelling such interest hierarchy in an attention network can fundamentally improve the representation of user behaviors. We therefore propose an improvement over DIN to model arbitrary interest hierarchy: Deep Interest with Hierarchical Attention Network (DHAN). In this model, a multi-dimensional hierarchical structure is introduced on the first attention layer which attends to individual item, and the subsequent attention layers in the same dimension attend to higher-level hierarchy built on top of the lower corresponding layers. To enable modelling of multiple dimensional hierarchy, an expanding mechanism is introduced to capture one to many hierarchies. This design enables DHAN to attend different importance to different hierarchical abstractions thus can fully capture a user's interests at different dimensions (e.g. category, price or brand). To validate our model, a simplified DHAN is applied to Click-Through Rate (CTR) prediction and our experimental results on three public datasets with two levels of one-dimensional hierarchy only by category. It shows DHAN's superiority with significant AUC uplift from 12% to 21% over DIN. DHAN is also compared with another state-of-the-art model Deep Interest Evolution Network (DIEN), which models temporal interest. The simplified DHAN also gets slight AUC uplift from 1.0% to 1.7% over DIEN. A potential future work can be combination of DHAN and DIEN to model both temporal and hierarchical interests.

This paper presents a framework for a customizable MPEG-4 face player using a FAPs (Facial Animation Parameters) sequence recovered from a real-time image sequence. First, the 3D nonrigid motion and structure of the facial features is recovered from a 2D image sequence and a "person-specific" model of the face. The model consists of the intensity and range image of the face. Then, the FAPs are computed from the recovered 3D structure. The customizable MPEG4 face animation is generated using a FAP sequence and a model of a specific person. The dataset of four face image sequences from three different face orientations equipped with range images are used. The GT (ground truth) results are generated using the 3D structure provided by range images. The results are evaluated quantitatively by comparing recovered FAP values, and qualitatively by comparing the generated MPEG-4 animations. FAPs are recovered up to 8% (relative) and 16 FAP units (absolute) accuracy and the animation is nearly identical to the MPEG-4 animation using GT FAPs.