The Analysis of the Telecommunications Sector by The Means of Data Mining Techniques

Introduction

The international telecommunications play an important role in today’s market dynamic. In the last decade a dramatic change in the ownership structure of telecommunications companies has taken place, from public (state-owned) monopolies to private companies. The rapid development of mobile telephone networks and video and Internet technologies has created enormous competitive pressure on the companies. As new competitors arise, companies need intelligent tools to gain a competitive advantage. Also, stock market expectations are enormous, and investors and financial analysts need tested tools to gain information about how companies perform financially compared to their competitors, what they are good at, who the major competitors are, etc. (Karlsson et al., 2001). In other words, the telecom companies need to benchmark their performances against competitors in order to remain important players in this market.

There is a huge amount of information about these companies’ financial performance that is now publicly available. This amount greatly exceeds our capacity to analyze it, the problem being that we often lack tools to quickly and accurately process these data. The aim of this paper is to analyze comparatively the financial performance of international telecom companies by the means of Data Mining techniques.

Data-mining techniques

Data-Mining techniques have different features that make them suitable for analyzing great amount of data. The goal of the DM is to discover new patterns in data, while most analytical tools (query and reporting tools, statistical analysis packages, online analytical processing - OLAP - and visualisation tools) are based on verification where "the system is limited to verifying user’s hypotheses" (Fayyad et al., 1996). The problem with the verification-based approach is that it "relies on the intuition of the analyst to pose the original question and refine the analysis based on the results of potentially complex queries against a database" (Moxon, 1996). Among the factors that limit the reliability of verification-based analytical tools are the ability of the analyst to pose appropriate questions and to manage the complexity of the attribute space. DM supports the discovery-based approach since "one defining data-mining characteristic is that research hypotheses and relationships between data variables are obtained as a result of (instead of as a condition for) the analyses activities" (Romeu, 2001). The discovery goal of the DM process can be further divided into prediction, where the system finds patterns or models for the purpose of future predictions and description, where the discovered patterns are presented in a human-understandable way to the user. In this paper we combine the two different goals of the DM process: we are interested in finding both patterns (models) that describe the financial situation of companies as well as models for financial (class) predictions.

In order to fulfil its role DM could perform a number of tasks such as clustering, classification, regression, dependency modelling, summarisation, and change and deviation detection. The link between these tasks and the real-world applications or business problems (the final goal of DM is to address these problems) is not straightforward, because real-world applications rarely have a simple single solution. Many different tasks may match a particular application, depending on how one approaches the problem (Smyth, 2002). For example, in this paper, the real-world application would be to assess telecom companies’ financial performance. Treating our problem as a supervised learning task implies that we already have financial performance classes for all the observations used to train the classifier. Actually there are no labelled data available, thus, the class variable has to be created at the beginning, by treating our problem as an unsupervised task. Only after the class variable has been constructed, can a classifier be trained. Smyth (2002) pinpoints various advices worth consideration when linking real-world applications with the data-mining task. The author states that it is advisable to start with only one task to address a real-world application and, only if necessary, add more complex ones. He also suggests removing the irrelevant details of the original formulation of the problem so that it resembles more closely a standard textbook task description. In order to select the proper task for a given problem, the data miner should have a complete understanding of both the business problem addressed and the task linked to it. Finally, Smyth (2002) states that it is better to approximate the solution to the right problem than it is to solve the wrong problem exactly.

In this paper, we have used Self-Organizing Map (SOM) algorithm to build the class variable and, then, we applied the induction of decision trees (famous Quinlan’s See5.0 tree-induction algorithm) in order to extract some rules of how to predict the performance of different telecom companies as data become available. For a detailed technical explanation of the two algorithms (SOM and See5.0) the reader is referred to Kohonen (1997) for SOM and Quinlan (1993) for See5.0.

The telecommunication dataset

We characterise companies’ financial performance by calculating the following ratios (see Table 1). This ratio classification is the most common financial ratio classification (Lev 1974, p. 12; Lehtinen 1996, p. 44) and has proposed the following categories for financial ratios: profitability, short-term solvency (liquidity), long-term solvency, and efficiency ratios.

Experiment

In this experiment we assess comparatively the financial performance of international telecommunication companies. Our results are presented in the following publications: Costea et al. (2002a, b), extended in Costea and Eklund (2004), applies SOM and the three classification techniques to assess telecom companies’ financial performance, Alcaraz and Costea (2004) introduces Weighting FCM to benchmark the telecom companies, Costea and Nastac (2005) studies three factors that can affect the classification performance of ANNs in the telecom sector data. In this paper we will present how a clustering technique (SOM) and a tree-induction technique (See 5.0) can be combined to assess comparatively financial performance of telecom companies.

Firstly, we applied SOM to the telecom data set. We used data from 1995 to 1999 (462 rows) to build the SOM map (see Figure 1) and train our See 5.0 classification model and the data for 2000 for the Scandinavian companies (11 rows) to test the prediction power of the tree-induction model.

Figure 1. Final 9x6 SOM for the telecom data set with identified "real"

clusters and feature planes

Source: Costea, 2005

The borders of the "real" clusters are identified by the dotted lines. Feature planes for each financial ratio are shown at the top of the figure ("warm" colours indicate high values, whereas "cold" colours indicate small values). The financial ratios are presented in Table 1.  Trajectories for the Scandinavian companies between 1995 and 1999.

After we attached the classes to each data row, we applied the tree-induction classification technique See5.0 (Quinlan, 1993). The training and testing accuracy rates for See 5.0 are presented in Table 2.

Conclusion

In this paper we showed how Data Mining techniques can be used to build classification models with which companies from different sectors, e.g. the telecommunications sector, can be compared based on their financial performance. In this attempt we have build a financial performance map (using the SOM clustering algorithm) and identified several financial performance clusters. Then, we have used a tree-induction algorithm, namely Quinlan’s See5.0 to model the relationship between the newly-built class variable (which shows the financial position/cluster for each observation) and the financial ratios which depict four performance dimensions: profitability, solvency, liquidity and efficiency.

Our results are good with regard to the validation criteria as well as to their interpretability and practical implications.

References

1.   Alcaraz-Garcia AF, Costea A. A Weighting FCM Algorithm for Clusterization of Companies as to their Financial Performances, Proceedings of the "IEEE 4th International Conference on Intelligent Systems Design and Applications" (ISDA 2004), Rudas I. (ed.), CD-ROM Edition, Budapest, Hungary , August 26-28, 2004, Track: Intelligent Business, p. 589-594.

2.   Costea A, Eklund T, Karlsson J. Making Financial Class Predictions Using SOM and Two Different Classification Techniques: Multinomial Logistic Regression and Decision Tree Induction. Proceedings of the "Central & Eastern European Workshop on Efficiency and Productivity Analysis". AES Press, Bucharest, Romania , June 28-29, 2002a

3.   Costea A, Eklund T, Karlsson J. A framework for predictive data mining in the telecommunication sector. Proceedings of the "IADIS International Conference" - WWW/Internet, Isaías P. (ed.). IADIS Press, Lisbon, Portugal . November 13-15, 2002b

4.   Costea A, Eklund T. Combining Clustering and Classification Techniques for Financial Performance Analysis. Proceedings of "8th World Multi-Conference on Systemics, Cybernetics and Informatics" (SCI 2004), Callaos et al. (eds.), Organized by IIIS, Orlando, Florida, USA , July 18-21, 2004, Volume I: Information Systems, Technologies and Applications, Track: Management Information Systems, p. 389-394, 2004

5.   Costea A. Computational Intelligence Methods for Quantitative Data Mining. Turku Centre for Computer Science, Ph. D. thesis, No. 67, Turku, Finland , November 2005

6.   Costea A, Nastac I. Assessing the predictive performance of ANN-based classifiers based on different data preprocessing methods, distributions and training mechanisms. In International Journal of Intelligent Systems in Accounting, Finance and Management (IJISAFM) 13(4): 217 - 250, December 2005

7.   Fayyad U, Piatetsky-Shapiro G, Smyth P. Knowledge Discovery and Data Mining: Towards a Unifying Framework. Proceedings of the "Second International Conference on Knowledge Discovery and Data Mining" (KDD’96), E. Simoudis, J. Han, and U. Fayyad (eds.), AAAI Press, Portland, Oregon, pp. 82-88, August 2-4, 1996

8.   Karlsson J, Back B, Vanharanta H, Visa A. Financial Benchmarking of Telecommunications Companies. TUCS Technical Report 395, February 2001.

9.   Karlsson J. Data-Mining, Benchmarking and Analysing Telecommunications Companies. Licentiate dissertation, Abo Akademi University , Turku, Finland , 2002.

10.  Kohonen T. Self-Organizing Maps. 2nd edition, Springer-Verlag, Heidelberg , 1997

11.  Lehtinen. Financial Ratios in an International Comparison. Validity and Reliability. Acta Wasaensia 49, Vaasa, Finland , 1996

12.  Lev B. Financial Statement Analysis. Englewood Cliffs, New Jersey, Prentice-Hall Inc, 1974

13.  Moxon B. Defining Data Mining - The Hows and Whys of Data Mining, and How It Differs From Other Analytical Techniques, 1996 [Available at:

http://www.dbmsmag.com/9608d53.html] (Accessed on: 10.10.2006).

14.  Quinlan JR. C4.5 Programs for Machine Learning. Morgan Kaufmann Series in Machine Learning. Morgan Kaufmann Publishers, San Mateo , 1993

15.  Romeu JL. Operations Research/Statistics Techniques: A Key to Quantitative Data Mining. Proceedings of FCSM "Federal Committee on Statistical Methodology" Conference, Key Bridge Marriott, Arlington,Virginia , November 14-16, 2001

16.  Smyth P. Selection of Tasks. Handbook of Data Mining and Knowledge Discovery - Task and Method Selection (Section 17.1 - pp. 443-444). Willi Klosgen and Jan M. Zytkow (eds.), Oxford University Press, New York, NY , 2002