Data Mining Applied to CRM for a Long Distance Telephone Company HUGO L.C. AZEVEDO, MARLEY M.B.R. VELLASCO, EMMANUEL P.L. PASSOS Laboratório de Inteligência Computacional Aplicada Departamento de Engenharia Elétrica Pontifícia Universidade Católica do Rio de Janeiro Rua Marquês de S. Vicente, 225, Gávea, Rio de Janeiro, RJ, Brasil, 22453-900 BRASIL Abstract: - Nowadays, it is very important for retail companies to understand their customers and establish a good relationship with them. It is then crucial to be able to segment the customers according to their buying patterns and needs. Unfortunately, this is not an easy task, requiring the consideration of several different mathematical techniques. In this work, a database from a long distance telephone company containing client- calling patterns was clustered and characterized. First, the Kohonen algorithm was used to cluster a subset of the database. Then, the clusters extracted were characterized using rules and some statistical visualization techniques. Finally, a classification model, based on neural networks, was built to classify future customers into the clusters extracted. Key-Words: - data mining, KDD, neural networks, Kohonen Algorithm, clustering, pattern classification, CRM 1 Introduction Nowadays, it is very important for a company to understand its customers and to create and manage relationships with them (CRM- Customer Relationship Management) [7]. In markets where the competition is high, this becomes a survival issue. Therefore, it is very important for a company to be able to segment its clients into clusters with similar characteristics. Segmentation allows the company to provide specific services and products to each group, according to its needs. A good performance in this task, usually increases the level of satisfaction and fidelity of the clients, consequently increasing company’s revenues and profit. Such task is usually easy for companies with few clients. For large companies, however, the situation is very different, requiring sophisticated techniques to accomplish an efficient segmentation. In such companies, client databases are generally large and complex, with usually spread all over the company departments. Clustering, cluster characterization and classification, are some of the tasks that are usually performed in a process called data mining, which is the most important step of a larger process called KDD (Knowledge Discovery in Databases) [3]. The mining step usually uses statistics, computational intelligence, OLAP, and other techniques, with the ultimate goal of extracting some useful and non trivial knowledge from a complex database. In the present work, a KDD process was performed on a complex database containing calling patterns from the clients of a long distance telephone company. 2 Description of the Work 2.1 Database The database analyzed in this work contained summarized data about the last 3 monthly bills of a sample of 4,000 company’s clients. The study was limited to non-international long-distance calls made from/to fix (not mobile) phones. The data retrieved from the company’s data warehouse was condensed into a single table containing 49 attributes. The attributes were related to: - Distance between the caller and the phone called; - Day of the week in which the call was made (weekday, Saturday or Sunday); - Time period of the call (peak or off-peak); - Type of the call (inter-regional, intra-regional or intra-sector). The measures used to quantify the attributes above were: - Number of calls made in a month; - Number of minutes spent in a month; - Revenues generated by the client in a month. For each measure, the average of the last 3 months was calculated.