Computing trajectory similarity is a fundamental task in trajectory analysis. Traditional heuristic methods suffer from quadratic computational complexity, which limits their scalability to large datasets. Recently, Trajectory Representation Learning (TRL) has been extensively studied to address this limitation. However, most existing TRL algorithms face two key challenges. First, they prioritize spatial similarity while neglecting the intricate spatio-temporal dynamics of trajectories, particularly temporal regularities. Second, these methods are often constrained by predefined single spatial or temporal scales, which can significantly impact performance, since the measurement of trajectory similarity depends on spatial and temporal resolution. To address these issues, we propose MSST, a Multi-Scale Self-supervised Trajectory Representation Learning framework. MSST simultaneously processes spatial and temporal information by generating 3D spatial-temporal tokens, thereby capturing spatio-temporal characteristics of trajectories more effectively. Further, MSST explore the multi-scale characteristics of trajectories. Finally, self-supervised contrastive learning is employed to enhance the consistency between the trajectory representations from different views. Experimental results on three real-world datasets for similarity trajectory computation provide insight into the design properties of our approach and demonstrate the superiority of our approach over existing TRL methods. MSST significantly surpasses all state-of-the-art competitors in terms of effectiveness, efficiency, and robustness. We explore the multi-scale characteristics of trajectories. To the best of our knowledge, this is the first effort in the TRL literature. Compared to previous TRL research, the proposed method can balance the noise and the details of trajectories, enabling a more comprehensive analysis by accounting for the variability inherent in trajectory data across different scales.

Social relationship understanding from videos bears an enormous potential for social media analysis. However, most existing researches only explored spatio-temporal features from videos, ignoring rich contextual information hidden in semantic objects. In this paper, we propose a Two-Stage Model (TSM), first introducing object information for social relationship understanding from videos, to our best knowledge. In the first stage, rich and robust representation for social relationship is obtained by the extraction of both spatio-temporal features and semantic objects information. In the second stage, we utilize a propagated knowledge graph to capture the interaction between semantic objects and video scenes. Specially, an attention mechanism is employed to measure the effectiveness of each semantic object on different scenes. Extensive experiments demonstrate our TSM achieves the state-of-the-art performance on the dataset SRIV.

Entity Disambiguation is commonly faced in semantic search and knowledge base population. However, it is a challenging task because of the diversity of mentions. Previous methods can be classified into two main groups. One focuses on disambiguating mentions in a document independently and mainly relies on the local context similarity. The other collectively disambiguates mentions only taking into account link information. These are not appropriate when the context and the link information are poor or misleading. In this paper, we propose a new method to collectively disambiguate mentions in documents. Our proposed framework considers three features, including text similarity, entity popularity, and entity relationship. First we adopt LeaderRank algorithm on the graph model to rank entities according to the link information among entities. Then we combine with global text similarity between entity and document to disambiguate mentions. Our detailed experimental evaluation on two benchmark datasets demonstrates our methods is effective.

This paper presents the method that underlies our submission to the popularity prediction task of Social Media Prediction Challenge 2017. The task is designed to predict the impact of sharing different posts for a publisher on social media. There are many factors that influence image popularity; these include not only the visual features of the image, but also the social features, such as user characteristics of its poster and even the upload time. In this project, we propose a fast and effective framework for popularity prediction. First, we investigate and extract visual and social features of images. For the visual feature, we introduce 1) global feature descriptors, such as Local Binary Pattern and Color Names, 2) local feature descriptors, such as Local Maximal Occurrence, and 3) deep features. For the social feature, we adopt users features (average views, group count, and member count), post features (title length, description length, and tag count), and time features (month, weekday, day, and hour). Furthermore, we fed a fusion of multi-feature to Linear Regression, Matrix Factorization based on Time and feature Cluster, and Support Vector Regression models respectively, and present comparative analysis of the prediction results. Finally, we choose the best model to predict the popularity scores of the test images. Experimental results demonstrate that our method can achieve 0.8581, 1.4062 and 0.8625 in terms of Spearman Ranking Correlation, Mean Absolute Error, and Mean Squared Error, respectively.

With the huge amount of web video data and its exponential growth in recent years, there are new challenges in Near-Duplicate Video Detection (NDVD) which have attracted much attention owing to its wide applications. One of the problems is how to extract discriminative features to achieve higher precision, and the other problem is how to improve the efficiency of large scale video analysis. Existing methods have predominantly focused on improving the precision and achieve some success on NDVD. However, it is an extremely challenging task to guarantee the higher precision and efficiency simultaneously. To balance the efficiency and precision, we propose a Multi-Feature based Parallel System (MFPS) for large scale NDVD to overcome these challenges. For feature extraction, we combine local and global features to accurately represent the critical information of video. In our work, not only Scale-Invariant Feature Transform (SIFT) but also Local Maximal Occurrence (LOMO) is introduced as local features and Color Name (CN) is adopted as global feature. Meanwhile, for SIFT and CN features, Bag-of-Visual-Words (BoVW) method is used to quantize the features due to its accuracy and efficiency in NDVD. Furthermore, we design a similarity measure for video pairs which is inspired by hough transform and sliding window idea. Finally, we implement the system on parallel cloud computing platform based on spark. A comprehensive evaluation is conducted on CC_WEB_VIDEO dataset which includes 12790 videos and 27% near-duplicates. Experimental results show that the proposed method is able to achieve some improvements in term of mAP against other methods. Moreover, our approach is efficient and achieves 5.8X speedup.