Pattern Recognition in Business Cases Analysis and Scoring

Usually business activity can be represented by several sets of cases: one set (or class of cases) contains successful, another – only tolerable, next – just so-so or even bad samples of cases. The case itself very often is fuzzily described and presented by people and members of business community. It means that some parameters of the certain case are presented quantitatively (in numbers) when others – only qualitatively (using verbal expressions, such as, for example, “very good”, “good”, “acceptable”, “bad” or “very bad”).

In 2006 BSS has introduced the Pattern recognition method (PRc) for case evaluation in risk assessment. The key ideas of the introduced method are:

1. to use the multiple balanced scorecards (one per each class) instead of using one general balanced scorecard;
2. to classify the case by comparison of the calculated values of scorecards for each class;
3. to use in the scorecards mathematically calculated values of coefficients (as a result of linear programming problem solution) instead of statistically calculated values.

Background

The primary idea was born in the Centre of Cybernetics at the Genoa University (Italy) in late sixties when the research in character recognition field on the probabilistic analyzer (PAPA) was performed. Practical implementation started at the Kaunas University of Technology (Lithuania) during the academic activity and research according to the PhD program studies. The picture below presents the evolution of BSS PRc:

Fig. 1 - Evolution of BSS Pattern Recognition method.

Operational model

On the background presented above an operational model for business activity analysis has emerged. The model consists of two parts (Fig.2):

1. The first one - training part, where parameters of the multiple scorecards are calculated. In real business processes this step is done periodically (once a year);
2. The second - recognition part, where the analyzed case is assigned to one of defined classes according to the maximum value of the multiple score. This step is performed for every new case.

Fig. 2 - Patter Recognition diagram.

The kernel of the BSS PRc software automatically performs: 1) certain preparation of cases (defuzzification, normalization and so on), 2) (the training part) solution of a Linear Programming problem, automatically formulated on the basis of business activity experience (“history of cases”), and determination of coefficients for case pattern multiple scoring, and 3) the case pattern recognition and assignment of the case under consideration to one of the several predetermined sets (or classes).

BSS PRc enables us to analyze even fuzzy history of each case set, extract its decisively important typical features and use them later for scoring and evaluating of a newly coming case pattern.

Main features of BSS PRc:

• It does not require huge amount of statistical data in any class set. It can work even with very modest business history;
• If history is rich BSS PRc is able to generalize it and extract the most important data;
• If business is new and no case history available at all BSS PRc enables you to start from the scratch creating some hypothetical cases for the purpose of simulation at the beginning. On top of those cases your own history will be enriched according to business experience gradually obtained;
• It permits to determine the most informative (the most influential) features and parameters of your business cases;
• BSS PRc encapsulates self-educating algorithm. While in operation it continuously improves the quality of recognition by analyzing the most recent historical data;
• All possible estimations and recommendations are soundly mathematically based;
• Linear Programming is used instead of rules of thumb;
• Ability to avoid average guesstimates guarantees individual approach to every case under consideration;
• BSS PRc can be easily integrated in other application fields such as: financial services market operations, different business evaluations, medical diagnosis, military operations planning, earthquake predictions,  signal and process recognition and many others.

Pattern recognition investigation tool

It is obvious that PRc method is more powerful than the ordinary scorecard or decision tree methods are. The recognition quality depends on the proper settings of algorithm parameters and on the correct definition of the samples of introduced cases (classes). To facilitate solution of business analysis problems BSS has developed the PRc Investigation tool. This tool has user friendly interface for the managing of investigation scenarios and presentation of BSS PRc behaviour.

Investigation tool allows importing the piece of operation history data and running certain training and recognition scenarios. The rich reporting functionality provides the comprehensive information on recognition quality, most sensitive pattern parameters, etc. At the end of investigation session the business experts obtain the best algorithm parameter setting and expected recognition resolution. During this self-learning session business experts also learn how to select samples of business cases for each class to get correct PRc training process results.

Main Investigation tool features:

• User-friendly package interacting with the business people by means of several windows simple to use;
• Ready to use for training of  business experts;
• Allows to perform the evaluation of correctness of your business organization ideas;
• The tool can be used by business experts in their daily work for the investigation of trends in business behaviour changes.

Baltic Software Solutions successfully absorbed scientific background of the Pattern recognition idea as well as its own experience in the field of leasing process management, financial operations analysis, business risk management, technical diagnosis and process evaluation.



For more information please download a leaflet Pattern recognition based scoring.