Xingzhong Du

Incremental Multimodal Query Construction for Video Search

Recent improvements in content-based video search have led to systems with promising accuracy, thus opening up the possibility for interactive content-based video search to the general public. We present an interactive system based on a state-of-the-art content-based video search pipeline which enables users to do multimodal text-to-video and video-to-video search in large video collections, and to incrementally refine queries through relevance feedback and model visualization. Also, the comprehensive functionalities enhance a flexible formulation of multimodal queries with different characteristics. Quantitative and qualitative analysis shows that our system is capable of assisting users to incrementally build effective queries over complex event topics.

Multiple graph unsupervised feature selection

Feature selection improves the quality of the model by filtering out the noisy or redundant part. In the unsupervised scenarios, the selection is challenging due to the unavailability of the labels. To overcome that, the graphs which can unfold the geometry structure on the manifold are usually used to regularize the selection process. These graphs can be constructed either in the local view or the global view. As the local graph is more discriminative, previous methods tended to use the local graph rather than the global graph. But the global graph also has useful information. In light of this, in this paper, we propose a multiple graph unsupervised feature selection method to leverage the information from both local and global graphs. Besides that, we enforce the l 2 , p norm to achieve more flexible sparse learning. The experiments which inspect the effects of multiple graph and l 2 , p norm are conducted respectively on various datasets, and the comparisons to other mainstream methods are also presented in this paper. The results support that the multiple graph could be better than the single graph in the unsupervised feature selection, and the overall performance of the proposed method is higher than the other comparisons. HighlightsA novel unsupervised feature selection algorithm is proposed which combines multiple graphs to uncover the manifold.The l 2 , p norm has more flexibility in controlling the sparse learning, thereby resulting in better performance.Combining multiple graph and l 2 , p norm results in better performance.

Minimal on-road time route scheduling on time-dependent graphs

On time-dependent graphs, fastest path query is an important problem and has been well studied. It focuses on minimizing the total travel time (waiting time + on-road time) but does not allow waiting on any intermediate vertex if the FIFO property is applied. However, in practice, waiting on a vertex can reduce the time spent on the road (for example, resuming traveling after a traffic jam). In this paper, we study how to find a path with the minimal on-road time on time-dependent graphs by allowing waiting on some predefined parking vertices. The existing works are based on the following fact: the arrival time of a vertex v is determined by the arrival time of its in-neighbor u, which does not hold in our scenario since we also consider the waiting time on u if u allows waiting. Thus, determining the waiting time on each parking vertex to achieve the minimal on-road time becomes a big challenge, which further breaks FIFO property. To cope with this challenging problem, we propose two efficient algorithms using minimum on-road travel cost function to answer the query. The evaluations on multiple real-world time-dependent graphs show that the proposed algorithms are more accurate and efficient than the extensions of existing algorithms. In addition, the results further indicate, if the parking facilities are enabled in the route scheduling algorithms, the on-road time will reduce significantly compared to the fastest path algorithms.

Discrete Deep Learning for Fast Content-Aware Recommendation

Cold-start problem and recommendation efficiency have been regarded as two crucial challenges in the recommender system. In this paper, we propose a hashing based deep learning framework called Discrete Deep Learning (DDL), to map users and items to Hamming space, where a user»s preference for an item can be efficiently calculated by Hamming distance, and this computation scheme significantly improves the efficiency of online recommendation. Besides, DDL unifies the user-item interaction information and the item content information to overcome the issues of data sparsity and cold-start. To be more specific, to integrate content information into our DDL framework, a deep learning model, Deep Belief Network (DBN), is applied to extract effective item representation from the item content information. Besides, the framework imposes balance and irrelevant constraints on binary codes to derive compact but informative binary codes. Due to the discrete constraints in DDL, we propose an efficient alternating optimization method consisting of iteratively solving a series of mixed-integer programming subproblems. Extensive experiments have been conducted to evaluate the performance of our DDL framework on two different Amazon datasets, and the experimental results demonstrate the superiority of DDL over the state-of-the-art methods regarding online recommendation efficiency and cold-start recommendation accuracy.

Streaming Ranking Based Recommender Systems

Studying recommender systems under streaming scenarios has become increasingly important because real-world applications produce data continuously and rapidly. However, most existing recommender systems today are designed in the context of an offline setting. Compared with the traditional recommender systems, large-volume and high-velocity are posing severe challenges for streaming recommender systems. In this paper, we investigate the problem of streaming recommendations being subject to higher input rates than they can immediately process with their available system resources (i.e., CPU and memory). In particular, we provide a principled framework called as SPMF (Stream-centered Probabilistic Matrix Factorization model), based on BPR (Bayesian Personalized Ranking) optimization framework, for performing efficient ranking based recommendations in stream settings. Experiments on three real-world datasets illustrate the superiority of SPMF in online recommendations.

Exploiting detected visual objects for frame-level video filtering

Videos are generated at an unprecedented speed on the web. To improve the efficiency of access, developing new ways to filter the videos becomes a popular research topic. One on-going direction is using visual objects to perform frame-level video filtering. Under this direction, existing works create the unique object table and the occurrence table to maintain the connections between videos and objects. However, the creation process is not scalable and dynamic because it heavily depends on human labeling. To improve this, we propose to use detected visual objects to create these two tables for frame-level video filtering. Our study begins with investigating the existing object detection techniques. After that, we find object detection lacks the identification and connection abilities to accomplish the creation process alone. To supply these abilities, we further investigate three candidates, namely, recognizing-based, matching-based and tracking-based methods, to work with the object detection. Through analyzing the mechanism and evaluating the accuracy, we find that they are imperfect for identifying or connecting the visual objects. Accordingly, we propose a novel hybrid method that combines the matching-based and tracking-based methods to overcome the limitations. Our experiments show that the proposed method achieves higher accuracy and efficiency than the candidate methods. The subsequent analysis shows that the proposed method can efficiently support the frame-level video filtering using visual objects.

An empirical study on user-topic rating based collaborative filtering methods

User based collaborative filtering (CF) has been successfully applied into recommender system for years. The main idea of user based CF is to discover communities of users sharing similar interests, thus, in which, the measurement of user similarity is the foundation of CF. However, existing user based CF methods suffer from data sparsity, which means the user-item matrix is often too sparse to get ideal outcome in recommender systems. One possible way to alleviate this problem is to bring new data sources into user based CF. Thanks to the rapid development of social annotation systems, we turn to using tags as new sources. In these approaches, user-topic rating based CF is proposed to extract topics from tags using different topic model methods, based on which we compute the similarities between users by measuring their preferences on topics. In this paper, we conduct comparisons between three user-topic rating based CF methods, using PLSA, Hierarchical Clustering and LDA. All these three methods calculate user-topic preferences according to their ratings of items and topic weights. We conduct the experiments using the MovieLens dataset. The experimental results show that LDA based user-topic rating CF and Hierarchical Clustering outperforms the traditional user based CF in recommending accuracy, while the PLSA based user-topic rating CF performs worse than the traditional user based CF.

Using Detected Visual Objects to Index Video Database

In this paper, we focus on how to use visual objects to index the videos. Two tables are constructed for this purpose, namely the unique object table and the occurrence table. The former table stores the unique objects which appear in the videos, while the latter table stores the occurrence information of these unique objects in the videos. In previous works, these two tables are generated manually by a top-down process. That is, the unique object table is given by the experts at first, then the occurrence table is generated by the annotators according to the unique object table. Obviously, such process which heavily depends on human labors limits the scalability especially when the data are dynamic or large-scale. To improve this, we propose to perform a bottom-up process to generate these two tables. The novelties are: we use object detector instead of human annotation to create the occurrence table; we propose a hybrid method which consists of local merge, global merge and propagation to generate the unique object table and fix the occurrence table. In fact, there are another three candidate methods for implementing the bottom-up process, namely, recognizing-based, matching-based and tracking-based methods. Through analyzing their mechanism and evaluating their accuracy, we find that they are not suitable for the bottom-up process. The proposed hybrid method leverages the advantages of the matching-based and tracking-based methods. Our experiments show that the hybrid method is more accurate and efficient than the candidate methods, which indicates that it is more suitable for the proposed bottom-up process.

A unified framework with a benchmark dataset for surveillance event detection

Abstract As an important branch of multimedia content analysis, Surveillance Event Detection (SED) is still a quite challenging task due to high abstraction and complexity such as occlusions, cluttered backgrounds and viewpoint changes etc. To address the problem, we propose a unified SED detection framework which divides events into two categories, i.e., short-term events and long-duration events. The former can be represented as a kind of snapshots of static key-poses and embodies an inner-dependencies, while the latter contains complex interactions between pedestrians, and shows obvious inter-dependencies and temporal context. For short-term event, a novel cascade Convolutional Neural Network (CNN)–HsNet is first constructed to detect the pedestrian, and then the corresponding events are classified. For long-duration event, Dense Trajectory (DT) and Improved Dense Trajectory (IDT) are first applied to explore the temporal features of the events respectively, and subsequently, Fisher Vector (FV) coding is adopted to encode raw features and linear SVM classifiers are learned to predict. Finally, a heuristic fusion scheme is used to obtain the results. In addition, a new large-scale pedestrian dataset, named SED-PD, is built for evaluation. Comprehensive experiments on TRECVID SEDtest datasets demonstrate the effectiveness of proposed framework.

Mobi-SAGE-RS: A sparse additive generative model-based mobile application recommender system

Abstract With the rapid prevalence of smart mobile devices and the dramatic proliferation of mobile applications (Apps), App recommendation becomes an emergent task that will benefit different stockholders of mobile App ecosystems. However, the extreme sparsity of user-App matrix and many newly emerging Apps create severe challenges, causing CF-based methods to degrade significantly in their recommendation performance. Besides, unlike traditional items, Apps have rights to access users’ personal resources (e.g., location, message and contact) which may lead to security risk or privacy leak. Thus, users’ choosing of Apps are influenced by not only their personal interests but also their privacy preferences. Moreover, user privacy preferences vary with App categories. In light of the above challenges, we propose a mobile sparse additive generative model (Mobi-SAGE) to recommend Apps by considering both user interests and category-aware user privacy preferences in this paper. To overcome the challenges from data sparsity and cold start, Mobi-SAGE exploits both textual and visual content associated with Apps to learn multi-view topics for user interest modeling. We collected a large-scale and real-world dataset from 360 App store - the biggest Android App platform in China, and conducted extensive experiments on it. The experimental results demonstrate that our Mobi-SAGE consistently and significantly outperforms the other existing state-of-the-art methods, which implies the importance of exploiting category-aware user privacy preferences and the multi-modal App content data on personalized App recommendation.