Donghua Yang's research while affiliated with Harbin Institute of Technology and other places

As the fundamental phrase of collecting and analyzing data, data integration is used in many applications, such as data cleaning, bioinformatics and pattern recognition. In big data era, one of the major problems of data integration is to obtain the global schema of data sources since the global schema could be hardly derived from massive data sources directly. In this paper, we attempt to solve such schema integration problem. For different scenarios, we develop batch and incremental schema integration algorithms. We consider the representation difference of attribute names in various data sources and propose ED Join and Semantic Join algorithms to integrate attributes with different representations. Extensive experimental results demonstrate that the proposed algorithms could integrate schemas efficiently and effectively.

In the context of big data, the interactive analysis database system needs to answer aggregate queries within a reasonable response time. The proposed AQP++ framework can integrate data preprocessing and AQP. It connects existing AQP engine with data preprocessing method to complete the connection between them in the process of interaction analysis. After the research on the application of materialized views in AQP++ framework, it is found that the materialized views used in the two parts of the framework both come from the accurate results of precomputation, so there’s still a time bottleneck under large scale data. Based on such limitations, we proposed to use approximate materialized views for subsequent results reuse. We take the method of identifying approximate interval as an example, compared the improvement of AQP++ by using approximate materialized view, and trying different sampling methods to find better time and accurate performance results. By constructed larger samples, we compared the differences of time, space and accuracy between approximate and general materialized views in AQP++, and analyzed the reasons for the poor performance in some cases of our methods. Based on the experimental results, it proved that the use of approximate materialized view can improve the AQP++ framework, it effectively save time and storage space in the preprocessing stage, and obtain the accuracy similar to or better than the general AQP results as well.

This paper explores the routing problem in ambient backscatter wireless sensor networks (AB-WSNs) using reinforcement learning approaches. Ambient RF signals serve as the only power source for battery-less sensor nodes and are also leveraged to enable backscatter communication among these nodes. This results in intermittent connection and dynamic topology within AB-WSNs, thereby making it difficult to route data to the sink, e.g., data may not reach the sink in a timely manner. We first introduce a multi-agent network model with two mechanisms to address this issue. We then model the routing problem with the Markov decision process, allowing each node to make informed route decisions based on the current state of its neighbors. With the aim of enabling each node to learn the optimal routing policy and do adaptive stateless routing, we propose two learning algorithms. The first, a value-based learning algorithm, is designed for sparse AB-WSNs. And the second, a policy-based learning algorithm, is intended to tackle the curse of dimensionality in dense AB-WSNs. We analyze the convergence of both learning algorithms and evaluate their performance through extensive experiments. The experiment results validate the convergence and efficiency of the proposed learning algorithms under various conditions.

Analysis of higher-order organizations, usually small connected subgraphs called motifs, is a fundamental task on complex networks. This paper studies a new problem of testing higher-order clusterability: given query access to an undirected graph, can we judge whether this graph can be partitioned into a few clusters of highly-connected motifs? This problem is an extension of the former work proposed by Czumaj et al. (STOC’ 15), who recognized cluster structure on graphs using the framework of property testing. In this paper, a good graph cluster on high dimensions is first defined for higher-order clustering. Then, query lower bound is given for testing whether this kind of good cluster exists. Finally, an optimal sublinear-time algorithm is developed for testing clusterability based on triangles.