Fuzzy PCA (Fuzzy Principal Component Analysis, TFN-component profile)

Fuzzy PCA is not a ranking method; it does not order alternatives, it produces criterion weights. Instead of a single figure in each cell, it takes a triangular fuzzy number (TFN) as input: a lower bound, a most likely value and an upper bound. This represents situations in which data is not exact but given as "approximate" or as expert judgement.

The idea the method rests on is principal component analysis (PCA), a statistical technique defined by Karl Pearson in 1901, which reduces the shared variation (variance) among criteria to a small number of "components." DecisionMind's profile applies the classical PCA calculation separately, and independently, to each of the TFN's three components, lower bound, most likely value, upper bound, and averages the three results. This approach is not a standard method known in the literature under the name "fuzzy PCA weighting" and resting on a single founding paper; it is DecisionMind's own engineering profile, and the manifest states this explicitly. It is a practical solution for a user who wants weights derived from the relationships between criteria but whose data is fuzzy; beyond carrying the name "PCA," however, it is not a formulation endorsed by a single academic source.

The idea behind PCA is this: if a group of criteria is strongly linked, one rising as the others rise, that group is in effect measuring a single underlying trend, and the criteria that reflect this trend most strongly deserve more weight. The method mathematically decomposes the shared variation among criteria and measures how much each criterion contributes to it. The fuzzy adaptation adds a further question: when data is not exact but given as an interval (l, m, u), how much do the weights calculated under "the most pessimistic reading," "the most likely reading" and "the most optimistic reading" differ from one another? If the three readings sit close together, uncertainty is not affecting the weight much; if the three readings are far apart, the width of the TFN genuinely changes the result.

The consequence of this philosophy is that the method shows both the structure among the criteria and how uncertain the data is, at the same time. But running the three components independently does not account for the TFN's own internal consistency, that the lower bound, the most likely value and the upper bound together represent a single fuzzy quantity; three separate crisp problems are solved and then averaged.

The method proceeds through three steps.

First, input validation. Every cell's triangular fuzzy number is checked to satisfy lower bound ≤ most likely value ≤ upper bound, and a direction (more is better / less is better) is confirmed for every criterion. A minimum of two alternatives and two criteria is required; if no criterion distinguishes between the alternatives (all values identical at all three components), the calculation halts.

Second, three independent PCA runs. The classical PCA calculation consists, in order, of min-max scaling, standardisation, decomposing the inter-criteria correlation matrix into eigenvalues and eigenvectors, and retaining the components whose eigenvalue exceeds 1 (the Kaiser criterion). This calculation is applied separately to the TFN's lower-bound table, most-likely-value table and upper-bound table. Each table on its own produces a classical PCA result, that is, its own weight vector.

Third, averaging and an uncertainty measure. The three weight vectors are averaged and rescaled so that they sum to 1; this is the final weight vector. DecisionMind also computes and reports how far this average sits from the weight derived from the most-likely value alone, and the largest weight difference among the three components; these two figures indicate how much the width of the TFN affects the result.

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

As in classical PCA, the weight vector reflects the criteria's contribution to the shared variance; here, however, that contribution is the average of three separate readings (pessimistic, most likely, optimistic). The weight alone is incompletely interpreted without reading the two additional figures reported alongside it: if the largest difference among the components is small (a few thousandths, say), the width of the TFN has changed the weight almost not at all, and the result can be approached with the same confidence as with crisp data. If this difference is large, the choice of lower and upper bound genuinely changes the result, and the report should state this sensitivity.

The weight measures a criterion's contribution to shared variation in the data, not its importance to the decision-maker; this is the same interpretation given on the classical PCA-WEIGHT card, only here it is the average of three fuzzy readings.

Thus instead of writing:

"The fuzzy PCA analysis showed that this criterion is the most important criterion"

the report should read:

"On the average of these three fuzzy readings, this criterion contributes most to shared variance; the difference between the lower and upper bound changes the weight by this much, which shows the effect of uncertainty on the result"

This profile works with fuzzy data (triangular fuzzy numbers, TFNs): every cell holds a triple, lower bound, most likely value and upper bound, instead of a single figure. If your data is crisp (a single figure in every cell), this profile is unnecessary; DecisionMind's crisp counterpart of the same idea is the separate card, PCA Weighting. This fuzzy profile has no separate extension of its own within DecisionMind.

You need the following: alternatives in rows, criteria in columns, a triple satisfying l ≤ m ≤ u in every cell, and "more is better or less is better" information for every criterion. Weights are not entered; the method produces them. A minimum of two alternatives and two criteria is required. If the lower and upper bounds are very close to one another (nearly crisp data), the results of the three components converge as well, and fuzziness adds little information.

If your criteria are numerical but not exact, given as an expert-judgement interval ("at least this much, likely this much, at most this much"), and you want weights derived from the shared structure among criteria, this profile is an option. Where you want to see how much uncertainty affects the weight, comparing these three readings is more useful than giving only a single crisp figure.

The cases where it should not be used follow both from its philosophy and from the nature of its source. If a strong shared structure among criteria is not expected, PCA's basic assumption, that a small number of components explain most of the variance, does not hold, and the weights can come out unbalanced. If the number of alternatives is small (three or four alternatives), the correlation matrix cannot be built reliably. If the result must be an academically established method resting on a single literature source, for example to be presented in a thesis or peer-reviewed publication as "the fuzzy PCA method defined in the literature," a method with a clear source should be chosen instead of this profile. This profile's originator is not an academic paper but DecisionMind's own engine.

Data is fuzzy, shared structure among criteria matters → Fuzzy PCA

Data is crisp → PCA Weighting

Only distinguishing power is needed, shared structure does not matter → Fuzzy Entropy

An academic publication needs a single established method source → Choose a fuzzy weighting method with a clear source, not this profile

Number of alternatives is very small → The correlation matrix is unreliable, read the result with caution

This profile's most practical feature is that it carries classical PCA's power to capture shared structure among criteria into fuzzy data. Calculating and comparing the three components separately directly shows how much uncertainty affects the weight; this is information that does not appear in a weight derived from a single crisp figure. The calculation can be followed step by step, and each component can be reported separately.

The most important limitation is the nature of its source: this profile has no single academic paper behind it; it is an engineering profile derived from DecisionMind's own source code (a software repository), and the manifest flags it as "evidence level C," that is, an evidence level without independent literature verification. Second, running the three components independently disregards the TFN's internal consistency, that the lower, middle and upper bound together represent a single fuzzy quantity; in effect, three separate crisp problems are solved and averaged, and no genuine fuzzy arithmetic is carried out. Third, classical PCA's own limitations, an unreliable correlation with few alternatives, an inability to see non-linear relationships among criteria, apply here as well. Fourth, if the gap between the lower and upper bound is not the same size for every criterion, that is, if one criterion carries a more "symmetric" uncertainty and another a more "skewed" one, this asymmetry can affect only the upper-bound component's weight. This effect is not visible from the average; the separate components must be examined.

The most common mistake is presenting this profile as "the named fuzzy PCA method defined in the literature"; this profile is DecisionMind's own engineering decision and must be stated as such in the report.

A second mistake is looking only at the final weight and never examining the difference among the three components; how much uncertainty affects the result is understood from that difference. A third is entering the lower and upper bound very close to one another, nearly as crisp data, and then claiming "a fuzzy analysis was carried out"; if the TFN's width is close to zero, all three components give the same result and fuzziness adds nothing. A fourth is reporting weights obtained from few alternatives (three or four) as though they were an exact proportion; the correlation matrix is not reliable at that size.

The governing principle is this:

The fuzzy PCA weight is the average of three separate crisp calculations; the report must show both this average and the difference among the components, and must state clearly that this profile is not a literature standard but DecisionMind's own engineering profile.