Employing Clustering Techniques and Association Rules for Client Segmentation and Attribute Dependency Mining in the Domain of Car Insurance

Segmenting clients according to common characteristics and determining the needs for each group is an important objective in the domain of car insurance. We perform segmentation through specific methods, comparing multiple clustering techniques, considering both classical and deep learning algorithms. Regarding the classical methods, clustering techniques appropriate for both numerical and categorical data were adopted, such as k-means clustering, X-means clustering, respectively k-prototype clustering. Regarding the deep-learning techniques, a stacked denoising autoencoder, followed by conventional clustering techniques, was experimented, the performance being compared with that achieved after the individual application of the classical techniques. After employing the clustering methods, the relevant attributes that separate among clusters were determined, the dependencies between the policy insurance type and other attributes being also analyzed through graphical representations and association rules. The experiments were performed considering the data extracted from a relational database specific for car insurance, containing 1000 instances for the main tables.