Xiangmin Zhou

Online Video Recommendation in Sharing Community

The creation of sharing communities has resulted in the astonishing increasing of digital videos, and their wide applications in the domains such as entertainment, online news broadcasting etc. The improvement of these applications relies on effective solutions for social user access to video data. This fact has driven the recent research interest in social recommendation in shared communities. Although certain effort has been put into video recommendation in shared communities, the contextual information on social users has not been well exploited for effective recommendation. In this paper, we propose an approach based on the content and social information of videos for the recommendation in sharing communities. Specifically, we first exploit a robust video cuboid signature together with the Earth Movers Distance to capture the content relevance of videos. Then, we propose to identify the social relevance of clips using the set of users belonging to a video. We fuse the content relevance and social relevance to identify the relevant videos for recommendation. Following that, we propose a novel scheme called sub-community-based approximation together with a hash-based optimization for improving the efficiency of our solution. Finally, we propose an algorithm for efficiently maintaining the social updates in dynamic shared communities. The extensive experiments are conducted to prove the high effectiveness and efficiency of our proposed video recommendation approach.

Enhancing online video recommendation using social user interactions

The creation of media sharing communities has resulted in the astonishing increase of digital videos, and their wide applications in the domains like online news broadcasting, entertainment and advertisement. The improvement of these applications relies on effective solutions for social user access to videos. This fact has driven the research interest in the recommendation in shared communities. Though effort has been put into social video recommendation, the contextual information on social users has not been well exploited for effective recommendation. Motivated by this, in this paper, we propose a novel approach based on the video content and user information for the recommendation in shared communities. A new solution is developed by allowing batch video recommendation to multiple new users and optimizing the subcommunity extraction. We first propose an effective technique that reduces the subgraph partition cost based on graph decomposition and reconstruction for efficient subcommunity extraction. Then, we design a summarization-based algorithm which groups the clicked videos of multiple unregistered users and simultaneously provide recommendation to each of them. Finally, we present a nontrivial social updates maintenance approach for social data based on user connection summarization. We evaluate the performance of our solution over a large dataset considering different strategies for group video recommendation in sharing communities.

An Approach for Clothing Recommendation Based on Multiple Image Attributes

Currently, many online shopping websites recommend commodities to users according to their purchase history and the behaviors of others who have similar history with the target users. Most recommendations are conducted by commodity tags based similarity search. However, clothing purchase has some specialized characteristics, i.e. users usually don’t like to go with the crowd blindly and will not buy the same clothing twice. Moreover, the text tags cannot express clothing features accurately enough. In this paper we propose a novel approach that extracts multi-features from images to analyze its content in different attributes for clothing recommendation. Specifically, a color matrix model is proposed to distinguish split joint clothing. ULBP feature is extracted to represent fabric pattern attribute. PHOG, Fourier, and GIST features are extracted to describe collar and sleeve attributes. Then, some classifiers are trained to classify clothing fabric patterns and split joint types. Experiments based on every attribute and their combinations have been done respectively, and have achieved satisfied results.

Discovery of stop regions for understanding repeat travel behaviors of moving objects

Different from simplifying trajectories using turning points to keep their coarse shapes, this paper summarizes trajectories using stop points.We define a new concept of stay stability (i.e., reciprocal of speed) between any two GPS points to detect stop points on individual trajectories.We discover repeat travel behaviors by finding common sequences of stop regions where a certain number of objects visit with similar stop duration.The experiments on 20 labeled trajectories in Geolife demonstrated the semantic effect, accuracy and near linear efficiency of our method. GPS trajectory dataset with high sampling-rates is usually in large volume that challenges the processing efficiency. Most of the data points on trajectories are useless. This paper summarizes trajectories using stop points. We define a new concept of stay stability (i.e., time dividing distance or reciprocal of speed) between any two GPS points to detect stop points on individual trajectories. We propose a novel Mining Repeat Travel Behaviors Using Stop Regions (MRTBUSR) method. In MRTBUSR, a stop region is a popular region containing a certain number of close stop points that can be grouped into a cluster. We then retrieve common sequences of stop regions to denote repeat route patterns and further analyze the stop durations on a stop region to find repeat travel behaviors. The experiments on 20 labeled trajectories selected from GeoLife demonstrated the semantic effect, accuracy and near linear efficiency of our proposed method.

Change Detection from Media Sharing Community

This paper investigates how social images and image change detection techniques can be applied to identify the damages caused by natural disasters for disaster assessment. We propose a framework that takes advantages of near duplicate image detection and robust boundary matching for the change detection in disasters. First we perform the near duplicate detection by local interest point-based matching over image pairs. Then, we propose a novel boundary representation model called relative position annulus (RPA), which is robust to boundary rotation, location shift and editing operations. A new RPA matching method is proposed by extending dynamic time wrapping (DTW) from time to position annulus. We have done extensive experiments to evaluate the high effectiveness and efficiency of our approach.

Modeling User Preferences on Spatiotemporal Topics for Point-of-Interest Recommendation

With the development of the location-based social networks (LBSNs) and the popular of mobile devices, a plenty of users check-in data accumulated enough to enable personalized Point-of-Interest recommendations services. In this paper, we propose a scheme of modeling users preferences on spatiotemporal topics (UPOST scheme) for accurate individualized POI recommendation. In the UPOST scheme, we discover temporal topics from semantic locations (i.e., peoples description words for a location) to learn users preferences. UPOST infers users preference for different types of places during different periods by learning the spatiotemporal topics from the historical semantic locations of users. We have developed two algorithms under the UPOST scheme: the time division LDA algorithm (TDLDA) and the time adaptive topic discovery algorithm (TATD). In TDLDA, we divide the check-in dataset into different time segments and use one LDA for one segment. Then we improve TDLDA further by developing a new TATD algorithm to discover spatiotemporal topics. The experimental results demonstrate the effectiveness of our UPOST scheme, both TDLDA and TATD outperform the counterpart method that do not consider semantic locations.

Clustering of Multiple Density Peaks

Density-based clustering, such as Density Peak Clustering (DPC) and DBSCAN, can find clusters with arbitrary shapes and have wide applications such as image processing, spatial data mining and text mining. In DBSCAN, a core point has density greater than a threshold, and can spread its cluster ID to its neighbours. However, the core points selected by one cut/threshold are too coarse to segment fine clusters that are sensitive to densities. DPC resolves this problem by finding a data point with the peak density as centre to develop a fine cluster. Unfortunately, a DPC cluster that comprises only one centre may be too fine to form a natural cluster. In this paper, we provide a novel clustering of multiple density peaks (MDPC) to find clusters with arbitrary number of regional centres with local peak densities through extending DPC. In MDPC, we generate fine seed clusters containing single density peaks, and form clusters with multiple density peaks by merging those clusters that are close to each other and have similar density distributions. Comprehensive experiments have been conducted on both synthetic and real-world datasets to demonstrate the accuracy and effectiveness of MDPC compared with DPC, DBSCAN and other base-line clustering algorithms.

Real-time context-aware social media recommendation

Social media recommendation has attracted great attention due to its wide applications in online advertisement and entertainment, etc. Since contexts highly affect social user preferences, great effort has been put into context-aware recommendation in recent years. However, existing techniques cannot capture the optimal context information that is most discriminative and compact from a large number of available features flexibly for effective and efficient context-aware social recommendation. To address this issue, we propose a generic framework for context-aware recommendation in shared communities, which exploits the characteristics of media content and contexts. Specifically, we first propose a novel approach based on the correlation between a feature and a group of other ones for selecting the optimal features used in recommendation, which fully removes the redundancy. Then, we propose a graph-based model called content–context interaction graph, by analysing the metadata content and social contexts, and the interaction between attributes. Finally, we design hash-based index over Apache Storm for organizing and searching the media database in real time. Extensive experiments have been conducted over large real media collections to prove the high effectiveness and efficiency of our proposed framework.

Cross the data desert: generating textual-visual summary on the evolutionary microblog stream

Effectively and efficiently summarizing social media is crucial and non-trivial to analyze social media. On social streams, events which are the main concept of semantic similar social messages, often bring us a firsthand story of daily news. However, to identify the valuable news, it is almost impossible to plough through millions of multi-modal messages one by one with traditional methods. Thus, it is urgent to summarize events with a few representative data samples on the streams. In this paper, we provide a vivid textual-visual media summarization approach for microblog streams, which exploits the incremental latent semantic analysis (LSA) of detected events. Firstly, with a novel weighting scheme for keyword relationship, we can detect and track daily sub-events on a keyword relation graph (WordGraph) of microblog streams effectively. Then, to summarize the stream with representative texts and images, we use cross-modal fusion to analyze the semantics of microblog texts and images incrementally and separately, with a novel incremental cross-modal LSA algorithm. The experimental results on a real microblog dataset show that our method is at least 1.31% better and 23.67% faster than existing state-of-the-art methods, and cross-modal fusion can improve the summarization performance by 4.16% on average.

Social event detection with retweeting behavior correlation

Event detection over microblogs has attracted great research interest due to its wide application in crisis management and decision making etc. In natural disasters, complex events are reported in real time on social media sites, but these reports are invisible to crisis coordinators. Detecting these crisis events helps watchers to make right decisions rapidly, reducing injuries, deaths and economic loss. In sporting activities, detecting events helps audiences make better and more timely game viewing plans. However, existing event detection techniques are not effective at handling complex social events that evolve over time. In this paper, we propose an event detection method that takes advantage of retweeting behavior for handling the events evolution. Specifically, we first propose a topic model called RL-LDA to capture the social media information over hashtag, location, textual and retweeting behavior. Using RL-LDA, a complex event can be well handled by exploring the correlation between retweeting behavior and the event. Then to maintain the RL-LDA in a dynamic environment, we propose a dynamic update algorithm, which incrementally updates events over real time streams. Experiments over real-world datasets show that RL-LDA detects the temporal evolution of complex events effectively and efficiently.