Grant: Discovery Grants Program - Individual

My research proposes the use of unsupervised clustering to enhance datasets prior to applying other data mining and business intelligence techniques including data summarization, second order unsupervised clustering, classification, prediction, and association mining. Inspired by unsupervised pre-training used in the convergence of deep feed-forward neural networks, the proposed work will be a significant step forward in this area. In the first phase, we will use state-of-the-art clustering techniques. In subsequent phases, we will explore data mining techniques beyond neural networks. To illustrate an application of the proposed work, consider a financial transaction dataset that can be used for customer relationship management, inventory management and fraud detection. A customer's representation can include temporal, geographical, or spending profiles. In phase one, these profiles can be used as attributes for representing customers in addition to other raw attributes such as total spending, frequency and recency of visits. The augmented dataset can be analyzed in the second phase using a diverse set of data mining techniques including business intelligence, association mining, supervised learning (classification and prediction), and second order unsupervised learning (clustering). The association mining will provide rules such as: “Those who spend more in summer tend to spend more on weekends”. The supervised learning will use known incidences of default, fraud and other anomalous behavior in financial transaction data to create models that predict the chances of fraud and anomalies based on customer profiles. For example, "holiday spenders" tend to default more on their loans. The subsequent unsupervised learning will also help us understand correlations between different profiles. For instance, "summer spenders" tend to be "non-local spenders".We will demonstrate the usefulness of our proposal using data from one wholesaler, one retailer, and open datasets for financial defaults to test and refine the proposed sequential data mining techniques. The application will not be restricted to commercial transactions. In addition, we will also use weather and energy consumption patterns to suggest optimal energy control strategies for buildings.While this proposal derives its inspiration from the unsupervised pre-training in deep learning, it will benefit from utilizing broader clustering research from the last fifty years. An example of one improvement is using cluster validity indices to determine the appropriate number of clusters. Furthermore, first-phase supervised learning will also be applied to other classification and prediction techniques such as decision trees, random forests and support vector machines. It can also enhance other data mining techniques such as optimization, association mining and business intelligence.