Huan Huo

Energy Efficient Clustering Algorithm for Data Gathering in Wireless Sensor Networks

Wireless sensor networks are characterized by centralized data gathering, multi-hop communication and many-to-one traffic pattern. These three characteristics may give rise to funneling effects that can lead to severe packet collision, network congestion, packet loss and even congestion collapse. This can also result in hotspots of energy consumption that may cause premature death of sensor nodes and even premature death of entire network. In this paper, exploiting spatial correlation of nodes to form clusters of nodes sensing similar values, and only cluster head sensor reading is transmit to sink, such can efficiently alleviates the funneling effects. A novelty clustering algorithm is proposed which can greatly reduce the number of cluster heads. Experimental results validate the effectiveness of this approach.

Collaborative Filtering Fusing Label Features Based on SDAE

Collaborative filtering (CF) is successfully applied to recommendation system by digging the latent features of users and items. However, conventional CF-based models usually suffer from the sparsity of rating matrices which would degrade model’s recommendation performance. To address this sparsity problem, auxiliary information such as labels are utilized. Another approach of recommendation system is content-based model which can’t be directly integrated with CF-based model due to its inherent characteristics. Considering that deep learning algorithms are capable of extracting deep latent features, this paper applies Stack Denoising Auto Encoder (SDAE) to content-based model and proposes DLCF(Deep Learning for Collaborative Filtering) algorithm by combing CF-based model which fuses label features. Experiments on real-world data sets show that DLCF can largely overcome the sparsity problem and significantly improves the state of art approaches.

LTMF: Local-Based Tag Integration Model for Recommendation

There are two primary approaches to collaborative filtering: memory- based and model-based. The traditional techniques fail to integrate with these two approaches and also can’t fully utilize the tag features which data contains. Based on mining local information, this paper combines neighborhood method and matrix factorization technique. By taking fuller consideration of the tag features, we propose an algorithm named LTMF (Local-Tag MF). After the real data validation, this model performs better than other state-of-art algorithms.

Recognizing Textual Entailment with Attentive Reading and Writing Operations

Inferencing the entailment relations between natural language sentence pairs is fundamental to artificial intelligence. Recently, there is a rising interest in modeling the task with neural attentive models. However, those existing models have a major limitation to keep track of the attention history because usually only one single vector is utilized to memorize the past attention information. We argue its importance based on our observation that the potential alignment clues are not always centralized. Instead, they may diverge substantially, which could cause the problem of long-range dependency. In this paper, we propose to facilitate the conventional attentive reading operations with two sophisticated writing operations - forget and update. Instead of utilizing a single vector that accommodates the attention history, we write the past attention information directly into the sentence representations. Therefore, higher memory capacity of attention history could be achieved. Experiments on Stanford Natural Language Inference corpus (SNLI) demonstrate the superior efficacy of our proposed architecture.

A Weighted K-AP Query Method for RSSI Based Indoor Positioning

The paper studies the establishment of offline fingerprint library based on RSSI (Received Signal Strength Indication), and proposes WF-SKL algorithm by introducing the correlation between RSSIs. The correlations can be transformed as AP fingerprint sequence to build the offline fingerprint library. To eliminate the positioning error caused by instable RSSI value, WF-SKL can filter the noise AP via online AP selection, meanwhile it also reduces the computation load. WF-SKL utilizes LCS algorithm to find out the measurement between the nearest neighbors, and it proposes K-AP (P,Q) nearest neighbor queries between two sets based on Map-Reduce framework. The algorithm can find out K nearest positions and weighted them for re-positioning to accelerate the matching speed between online data and offline data, and also improve the efficiency of positioning. According to a large scale positioning experiments, WF-SKL algorithm proves its high accuracy and positioning speed comparing with KNN indoor positioning.

Anomalous Region Detection on the Mobility Data

Mobility data records the change of location and time about the crowd activities, reflecting a large amount of semantic knowledge about human mobility and hot regions. From the perspective of regional semantic knowledge, mining anomalous regions of overcrowded area is essential for disaster-aware resilience system scheme. This paper studies how to discover anomalous regions of moving crowds over the mobility data. From the perspective of spatial information analysis about the location sequence of moving crowds, the paper introduces grid structure to index activity space and proposes a density calculation method of grid cells based on kernel function. By adopting Top-k sorting method, the algorithm determines the density thresholds to detect the anomalous regions. Finally, experimental results validate the feasibility and effectiveness of the above method on practical data sets.

A Trajectory Prediction Method for Location-Based Services

Most existing location prediction techniques for moving objects on road network are mainly short-term prediction methods. In order to accurately predict the long-term trajectory, this paper first proposes a hierarchical road network model, to reduce the intersection vertexes of road network, which not only avoids unnecessary data storage and reduces complexity, but also improves the efficiency of the trajectory prediction algorithm. Based on this model, this paper proposes a detection backtracking algorithm, which deliberately selects the highest probability road fragment to improve the accuracy and efficiency of the prediction. Experiments show that this method is more efficient than other existing prediction methods.

The cost model for fragmented XML streams

The cost-model for XML stream is the theory foundation in the research of XML stream dissemination, fragmentation and evaluation. According to the matching operations in fragmented XML stream system, this paper analyzes the features of the system overhead on client, network and server, and brings in the cost model for fragmented XML stream based on Hole-Filler model. Based on the cost model, this paper illustrates the influence of the fragmentation policy on XML dissemination system by an application and proposes a series of selection policies for XML stream fragmentation to improve the system efficiency, adaptivity and scalability.

Keyword Search on Streaming XML Fragments

With the growing popularity of XML and emergence of streaming data model, processing streaming XML has become an important topic. This paper proposes keyword search solution over XML fragment streams based on hole-filler model. Two efficient indexes, dual list and sketch are developed to further improve the performance: dual list indexes the candidate XML fragments to keep track of the relationship among fragments that include keywords and sketch summarizes the candidate XML elements to compute SLCA. SLCA computing algorithm, which is triggered by certain keywords, avoids redundant operations on computing the SLCA of elements that not contribute to the final result. The algorithm produces partial answers continuously without having to wait for the end of the stream. We illustrate the effectiveness of the algorithms developed with experiments.