Classification and outlier detection based on topic based pattern synthesis

• Authors: Samrat Kokkula;Narasimha Murty Musti
• Venue: MLDM'13 Proceedings of the 9th international conference on Machine Learning and Data Mining in Pattern Recognition
• Year: 2013

In several pattern classification problems, we encounter training datasets with an imbalanced class distribution and the presence of outliers, which can hinder the performance of classifiers. In this paper, we propose classification schemes based on the pre-processing of data using Novel Pattern Synthesis (NPS), with the aim to improve performance on such datasets. We provide a formal framework for characterizing the class imbalance and outlier elimination. Specifically, we look into the role of NPS in: Outlier elimination and handling class imbalance problem. In NPS, for every pattern its k-nearest neighbours are found and a weighted average of the neighbours is taken to form a synthesized pattern. It is found that the classification accuracy of minority class increases in the presence of synthesized patterns. However, finding nearest neighbours in high-dimensional datasets is challenging. Hence, we make use of Latent Dirichlet Allocation to reduce the dimensionality of the dataset. An extensive experimental evaluation carried out on 25 real-world imbalanced datasets shows that pre-processing of data using NPS is effective and has a greater impact on the classification accuracy over minority class for imbalanced learning. We also observed that NPS outperforms the state-of-the-art methods for imbalanced classification. Experiments on 9 real-world datasets with outliers, demonstrate that NPS approach not only substantially increases the detection performance, but is also relatively scalable in large datasets in comparison to the state-of-the-art outlier detection methods.