Contextual Improvement Planning by Fuzzy-Rough Machine Learning: A Novel Bipolar Approach for Business Analytics

Abstract

Nearly all companies need to retrieve valuable information from business data to increase its efficiency or value, and the rising interests of research in this domain could be named as business analytics. Because most of the problems (obstacles) faced by business have to consider a group of complex and interrelated factors, conventional statistics models (e.g., regression) have constraints in resolving these interrelated and complex problems. Therefore, this study proposes a novel multiple attribute decision-making model to resolve—from ranking/selection to improvement planning—the problems of business analytics in finance, based on the similarity with positive contexts (rules) and the dissimilarity with negative ones. The proposed model not only enhances the previous method (i.e., dominance-based rough set approach, DRSA) on ranking within the same decision class, but also provides a contextual approach to guide businesses for systematic improvements. Infused with the modified VIKOR method, the proposed model could support a company to transform analytics into priority contexts, which may guide improvement planning. To show the proposed model, a group of semiconductor companies in Taiwan is analyzed as an empirical case, and three companies are taken as examples to illustrate the ranking and improvement planning processes. The obtained findings thus contribute to bridge the applications of data-driven business analytics to the field of decision science in practice.

Appendices

Appendix 1

#Positive Rules (associated with Good DC)

• 1. (AR_turnover ≥ H) & (Inventory ≥ H) & (ROA ≥ M) & (EPS ≥ M) => (DC ≥ Good)

• 2. (Speed ≥ M) & (AR_turnover ≥ H) & (EPS ≥ M) & (CashFlow ≥ H) & (CashFlow_adq ≥ M) => (DC ≥ Good)

• 3. (Debt ≥ M) & (InterestCoverage ≥ M) & (AR_turnover ≥ H) & (Inventory ≥ M) & (EPS ≥ H) & (CashFlow ≥ M) => (DC ≥ Good)

• 4. (NetPafterTax ≥ H) & (EPS ≥ M) & (CashFlow ≥ M) & (CashFlow_adq ≥ H) & (CashFlow_reinv ≥ M) => (DC ≥ Good)

• 5. (LongCapital ≥ M) & (AR_days ≥ H) & (ROA ≥ M) & (CashFlow ≥ H) => (DC ≥ Good)

• 6. (LongCapital ≥ M) & (Speed ≥ M) & (AR_turnover ≥ M) & (AR_days ≥ H) & (ROE ≥ M) & (CashFlow_adq ≥ M) => (DC ≥ Good)

• 7. (LongCapital ≥ H) & (ROE ≥ H) & (CashFlow ≥ H) => (DC ≥ Good)

• 8. (FixAssetTurnover ≥ M) & (EPS ≥ M) & (CashFlow ≥ H) & (CashFlow_adq ≥ H) => (DC ≥ Good)

• 9. (LongCapital ≥ H) & (Liquidity ≥ H) & (Speed ≥ M) & (EPS ≥ H) => (DC ≥ Good)

• 10. (LongCapital ≥ H) & (Liquidity ≥ H) & (AR_days ≥ M) & (NetPbeforeTax ≥ H) => (DC ≥ Good)

• 11. (LongCapital ≥ M) & (InterestCoverage ≥ M) & (ROA ≥ H) & (NetPafterTax ≥ M) & (CashFlow ≥ M) => (DC ≥ Good)

• 12. (LongCapital ≥ M) & (InterestCoverage ≥ M) & (ROA ≥ H) & (CashFlow ≥ M) & (CashFlow_reinv ≥ H) => (DC ≥ Good)

• 13. (LongCapital ≥ M) & (InterestCoverage ≥ H) & (Inventory ≥ M) & (ROE ≥ M) & (EPS ≥ H) => (DC ≥ Good)

#Negative Rules (associated with Bad DC)

• 14. (EPS ≤ L) => (DC ≤ Bad)

• 15. (LongCapital ≤ L) & (Liquidity ≤ M) => (DC ≤ Bad)

• 16. (Speed ≤ L) & (ROA ≤ M) => (DC ≤ Bad)

• 17. (AR_turnover ≤ M) & (AR_days ≤ L) => (DC ≤ Bad)

• 18. (AR_days ≤ M) & (Inventory ≤ L) => (DC ≤ Bad)

• 19. (Debt ≤ M) & (CashFlow ≤ L) => (DC ≤ Bad)

• 20. (CashFlow ≤ L) & (CashFlow_adq ≤ L) => (DC ≤ Bad)

• 21. (NetPbeforeTax ≤ L) & (CashFlow_adq ≤ L) => (DC ≤ Bad)

• 22. (Debt ≤ L) & (CashFlow_adq ≤ L) => (DC ≤ Bad)

• 23. (Liquidity ≤ M) & (InterestCoverage ≤ L) => (DC ≤ Bad)

• 24. (Debt ≤ M) & (AR_turnover ≤ M) & (FixAssetTurnover ≤ M) & (CashFlow_reinv ≤ M) => (DC ≤ Bad)

Appendix 2

The calculations of fuzzy evaluations on the three alternatives were based on collecting the five domain experts’ perceptions regarding three scales: Unsatisfied (U), Neutral (N), and Satisfied (S), ranging from 0.00 to 1.00 by the triangular fuzzy membership function. The collected fuzzy numbers on the three scales from the five experts are in Table 9. The figures from a company, industrial average, max, min, and standard deviation on each attribute were provided with the requirements of attributes in the questionnaire. Take “A ₁ ≥ M” in C ₃ (company A) for example, the evaluation of the requirement (A ₁ ≥ M) for A was based on the five experts’ opinions; their opinions on this requirement were “N,” “N,” “N,” “U,” and “U”.

Table 9 Fuzzy numbers for the three scales by the five experts

Based on Table 9 and (17), the fuzzy evaluations (by the five experts) of company A on this requirement (i.e., A ₁≥M) was (0.40, 0.50, 0.60) ⊕ (0.35, 0.50, 0.65) ⊕ (0.40, 0.50, 0.60) ⊕ (0.00, 0.00, 0.25) ⊕ (0.00, 0.00, 0.30)/5 = (0.23, 0.30, 0.48). The fuzzy-averaged triangular membership function can be illustrated in Fig. 3, and the defuzzified performance score was (0.23 + (0.48−0.23) + (0.30−0.23))/3 = 0.34. (Table 9 and Fig. 3.)

Fig. 3

Averaged fuzzy evaluation as triangular membership function

Appendix 3

Take the case of A ₃ attribute (i.e., Liquidity) of company A for example, the industrial average $\overline{A}_{3} = 333.42$ and SD₃ = 327.35, and the raw figure of company A on Liquidity = 212.98. Therefore, its Z-score-based transformation equals to −0.3679 ((212.98–333.42)/327.35 = 0.3679) by (19); thus, it was categorized as “M” because −0.5244 < −0.3679 < 0.5244. In this case, it would be regarded as fully satisfied for the premise “A ₃ ≤ M” in Table 5; in other words, its performance evaluation on the requirement “A ₃ ≤ M” would be 1.00 for calculating the synthesized bipolar decision model. Otherwise, if Z-score-based transformation was higher than 0.5244 (e.g., 0.8888), its performance evaluation on the requirement “A ₃ ≤ M” would be 0.00.