Rough-DRSA (Dominance-Based Rough Set Approach)

Rough-DRSA is useful when you hold a set of examples labelled with past decisions. This example set is a decision table in which it is known who was assigned to which class. The method sorts new alternatives into classes using rules learned from these examples; this is sorting, not ranking. Its output is not a score but a class label: "certainly good," "certainly bad" or "boundary region" (uncertain). Rough-DRSA is the adaptation of Pawlak's 1982 rough set theory to ordinal criteria, made by Greco, Matarazzo and Słowiński in 2001.

Classical rough set theory groups together objects that cannot be distinguished from one another, that is, objects sharing the same attributes. But in multiple-criteria decisions, criteria are usually ordinal: a "good" grade is always superior to an "average" grade. Starting from this observation, Greco, Matarazzo and Słowiński replaced indiscernibility with a dominance relation: an object dominates another if it is at least as good as that object on every criterion. From this dominance relation, DRSA builds, for every class, a "certain" lower set (objects that undoubtedly belong to that class) and a "possible" upper set (objects that could belong to that class).

The philosophical consequence of this approach is that it does not hide inconsistency. In real data, two objects with identical criterion values may have been assigned to different classes; this shows that some unmeasured factor influenced the decision. DRSA calls this situation the "boundary region" and marks it openly, rather than averaging it away. For this reason DRSA is not a compensatory method; it is a rule-based, transparent classification approach.

The method proceeds through five steps.

First, building the decision table. You need past examples that have already been classified, together with the ordinal criteria that describe them. A class assignment (for example, "good" or "bad") is known for every object. The dominance relation between criteria is defined: an object dominates another if it is at least as good on every criterion.

Second, building the lower and upper approximations. For every union of classes (for example, the "at least good" class), a lower approximation and an upper approximation are built. The lower approximation gathers the objects that certainly belong to that class, those the dominance relation shows to be unaffected by any contradiction. The upper approximation covers every object that could belong to that class.

Third, the quality and accuracy of the approximation. The ratio of the number of objects in the lower approximation to the total number of objects gives the overall quality of the classification. This ratio shows how much inconsistency exists in the dataset; the closer it is to 1, the more consistent the data.

Fourth, extracting decision rules. Certain rules of the form "if ... then at least this class" are derived from the lower approximations, and possible rules of the form "probably this class" are derived from the boundary region. These rules show directly which criterion thresholds drive a change of class.

Fifth, applying the rules to new objects. A new alternative is assigned to whichever class or boundary region matches the rule it satisfies.

The formulas behind each step and the intermediate tables are given on the DecisionMind method page; this card carries no formulas.

DRSA's real output is a class label: certainly good, certainly bad, or the boundary region (uncertain). This is not the single number that ranking methods produce. DecisionMind's comparison screen may show a computed numerical "dominance score" to display a rank; this score is an auxiliary indicator of how many objects it dominates, and it is not an output produced by DRSA itself. Reading this score as a measure of confidence or certainty is wrong; the real information is the class label.

An object in the boundary region does not mean "somewhat good, somewhat bad"; it means "which class it belongs to cannot be determined with the criteria at hand." A low quality of approximation does not mean "the method failed"; it is a true reflection of inconsistency in the data (examples with the same criteria assigned different decisions), and it shows that the criterion set may be incomplete.

Thus instead of writing:

"Rough-DRSA ranked the alternatives, the first alternative is this one"

the report should read:

"Rough-DRSA assigned this alternative to the certainly good class; these alternatives lie in the boundary region and cannot be placed in a definite class with the criteria at hand"

Rough-DRSA belongs to the rough data-type card; this means the computation is performed not on crisp numbers but on the dominance classes the numbers form. DecisionMind currently holds no other data-type extension of this method.

You need criteria scored on an ordinal scale (for example, poor/average/good), past examples containing a class assignment for every object, and at least two classes. Rough-DRSA neither asks for nor produces weights; it measures a criterion's importance not by weight but by how much the quality falls when that criterion is removed from the classification (indispensability). The recommended problem size is between two and twenty-five objects and three to ten criteria; the number of rules can grow rapidly in large datasets.

Rough-DRSA is suitable if you hold examples classified in the past and your aim is to sort new alternatives into classes using explicit rules learned from these examples. Its typical territory is credit-risk classification, quality control and classification based on expert assessment; it works especially well when the unit of measurement between criteria is not exact but the order information is reliable.

It should not be used where the aim is a ranking (who comes first, who comes second); DRSA produces classes, not ranks. If no past examples exist, or if there are too few to extract rules from, DRSA cannot produce reliable rules. If even the order information between criteria is unreliable (only unordered, categorical data is available), the classical rough set approach is more suitable.

Aim is classification, past examples exist → Rough-DRSA

Aim is ranking (who comes first) → TOPSIS, RAPS, REGIME

Criteria are unordered, purely categorical → Classical rough sets (Pawlak, 1982)

Not a ranking but weights are needed → AHP, BWM, SWARA (subjective); Entropy, CRITIC (objective)

Rough-DRSA's most important advantage is transparency: the result is not a black-box score but readable rules of the form "if this criterion is above this threshold, it belongs to at least this class." Instead of hiding real inconsistencies in the dataset, it shows them openly as a boundary region; this tells the decision-maker where more information is needed. It requires no criterion weights; it derives which criterion is indispensable (the core) directly from the data. It needs no precision in the unit of measurement for ordinal data.

Its limitations stem from its rule-based structure. First, when working with few examples, the extracted rules can become overly specialised (overfitting) and may not generalise well to new objects. Second, DRSA gives no definite answer for objects in the boundary region; the decision-maker must resolve this uncertainty with additional information. Third, as the number of criteria grows, the number of possible rules grows rapidly and interpretability becomes harder. Fourth, the numerical auxiliary scores fitted onto ranking screens for DRSA can overshadow the method's real philosophy (classification, not ranking); this distinction must be made explicit in every report (Greco, Matarazzo and Słowiński, 2004).

The most common mistake is treating DRSA's boundary-region output as an error or a shortcoming; the boundary region correctly reflects a real inconsistency in the data and should not be hidden. A second mistake is presenting the auxiliary numerical score as a ranking and never showing the class labels that are DRSA's real output in the report. A third mistake is extracting rules from too few examples and generalising them to a large population. A fourth mistake is overlooking the core (indispensable) criterion and assuming all criteria carry equal importance.

The governing principle is this:

Rough-DRSA's output is a class label; giving a definite "good or bad" answer about an alternative in the boundary region goes beyond the information the method provides.