Representation Learning for Activity with Multi-level Attention Kinematic Similarity Computation PROJECT TITLE : Representation Learning with Multi-level Attention for Activity Trajectory Similarity Computation ABSTRACT: The widespread adoption of GPS-enabled hardware and wireless Communication technology has resulted in the accumulation of massive amounts of trajectory data. Specifically, activity trajectory from Location-based Social Networks (LBSN) enriches traditional trajectory data with additional user semantic activities, such as visiting work/home/entertainment places. This adds a new layer of meaning to the data. Comparing the proximity of two activity trajectories along multiple dimensions, such as time, location, and semantics, is one way to evaluate the degree to which they are similar to one another. We are able to mine implicit user preference and apply it to route planning, point of interest recommendation, or any other online tasks in this way. There are two aspects that make comparing activity trajectories, also known as computing their similarity, a particularly difficult task. One problem is that the sampling rate is not consistent in either time or space. The other issue is the disparity between the activities of each individual. Previous work finds a solution to the problem of uneven sampling rate by employing trajectory complements, which are restricted to information that is spatial-temporal in nature. In this paper, we propose learning a representation for one activity trajectory by jointly considering the spatio-temporal characteristics and the activity semantics. This will allow us to learn a representation for one activity trajectory. In order to compute the degree to which two trajectories are similar to one another, individual trajectory points and contextual features are given different amounts of weight using multi-level attention mechanisms. To be more specific, we propose a point-level and feature-level attention mechanism as a means of adaptively selecting critical elements and contextual factors for the purpose of learning trajectory representation. In a comprehensive experimental evaluation on real trajectory databases, our proposed method, which we have given the name At2vec, demonstrates superior performance when compared to existing baselines. Did you like this research project? To get this research project Guidelines, Training and Code... Click Here facebook twitter google+ linkedin stumble pinterest Enhancing Quick and Accurate Static 3D Cloth Draping by Curvature Loss with GarNet++ For robust image classification, regularization on augmented data to diversify the sparse representation is necessary.