Identifying user habits through data mining on call data records

TL;DR

This paper introduces a framework using unsupervised clustering methods to identify user habits from anonymized call data records, enabling profile discovery without prior domain knowledge.

Contribution

The paper presents two novel clustering approaches, LD-ABCD and PROCLUS, for extracting user behavior patterns from CDR data without using pre-existing labels.

Findings

LD-ABCD effectively discovers meaningful clusters with local dissimilarity measures.

Additional features can be generated to reveal implicit information in sparse data.

Graphical visualization aids in interpreting clustering results for practical use.

Abstract

In this paper we propose a framework for identifying patterns and regularities in the pseudo-anonymized Call Data Records (CDR) pertaining a generic subscriber of a mobile operator. We face the challenging task of automatically deriving meaningful information from the available data, by using an unsupervised procedure of cluster analysis and without including in the model any \textit{a-priori} knowledge on the applicative context. Clusters mining results are employed for understanding users' habits and to draw their characterizing profiles. We propose two implementations of the data mining procedure; the first is based on a novel system for clusters and knowledge discovery called LD-ABCD, capable of retrieving clusters and, at the same time, to automatically discover for each returned cluster the most appropriate dissimilarity measure (local metric). The second approach instead is based on PROCLUS, the well-know subclustering algorithm. The dataset under analysis contains records characterized only by few features and, consequently, we show how to generate additional fields which describe implicit information hidden in data. Finally, we propose an effective graphical representation of the results of the data-mining procedure, which can be easily understood and employed by analysts for practical applications.