Complex graph stream mining
- Publication Type:
- Thesis
- Issue Date:
- 2015
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Recent years have witnessed a dramatic increase of information due to the ever development of modern technologies. The large scale of information makes data analysis, particularly data mining and knowledge discovery tasks, unprecedentedly challenging. First, data is becoming more and more interconnected. In a variety of domains such as social networks, chemical compounds, and XML documents, data is no longer represented by a flat table with instance-feature format, but exhibits complex structures indicating dependency relationships. Second, data is evolving more and more dynamically. Emerging applications such as social networks continuously generate information over time. Third, the learning tasks in many real-life applications become more and more complicated in that there are various constraints on the number of labelled data, class distributions, misclassification costs, or the number of learning tasks etc.
Considering the above challenges, this research aims to investigate theoretical foundations, study new algorithm designs and system frameworks to enable the mining of complex graph streams from three aspects, including (1) Correlated Graph Stream Mining, (2) Graph Stream Classifications, and (3) Complex Task Graph Classification.
In particular, correlated graph stream mining intends to carry out structured pattern search and support the query of similar graphs from a graph stream. Due to the dynamic changing nature of the streaming data and the inherent complexity of the graph query process, treating graph streams as static datasets is computationally infeasible or ineffective. Therefore, we proposed a novel algorithm, CGStream, to identify correlated graphs from a data stream, by using a sliding window, which covers a number of consecutive batches of stream data records. Experimental results demonstrate that the proposed algorithm is several times, or even an order of magnitude, more efficient than the straightforward algorithms.
Graph stream classification aims to build effective and efficient classification models for graph streams with continuous growing volumes and dynamic changes. We proposed two methods for complex graph stream classification. Due to the inherent complexity of graph structure, labelling graph data is very expensive. To solve this problem, we proposed a gLSU algorithm, which aims to select discriminative subgraph features with minimum redundancy by using both labelled and unlabelled graphs for graph streams. The second approach handles graph streams with imbalanced class distributions and noise. Both frameworks use an instance weighting scheme to capture the underlying concept drifts of graph streams and achieve significant performance gain on benchmark graph streams.
Complex task graph classification aims to address the graph classification problems with complex constraints. We studied two complex task graph classification problems, cost-sensitive graph classification of large-scale graphs and multi-task graph classification. As in medical diagnosis the misclassification cost/risk for different classes is inherently different and large scale graph classification is highly demanded in real-life applications, we proposed a CogBoost algorithm for cost-sensitive classification of large scale graphs. To overcome the limitation of insufficient labelled graphs for a specific learning task, we further proposed effective algorithms to leverage multiple graph learning tasks to select subgraph features and regularize multiple tasks to achieve better generalization performance for all learning tasks.
Please use this identifier to cite or link to this item: