ELECTRE Tree AutoElicit: Auto-Calibration Methods

• ELECTRE Tree AutoElicit is a family of structured elicitation schemes that leverage tree, ensemble, or hierarchical decompositions to infer parameters in ELECTRE models.
• It reduces elicitation complexity by replacing the need for m(m-1)/2 pairwise comparisons with efficient tournament tree and Deck of Cards methods.
• The approach supports various applications from weight calibration in TTM and ELECTRE Tri-B to simulation-driven parameter tuning for network ranking.

ELECTRE Tree AutoElicit denotes, in the available literature, a set of structured elicitation procedures for ELECTRE-type multi-criteria decision models in which a tree, ensemble, or hierarchical decomposition is used to infer parameters that would otherwise require heavier expert input. The most direct formulation couples ELECTRE with the Tournament Tree Method (TTM): a tournament tree is built over the criteria set, exactly $m-1$ pairwise judgments are collected through the Deck of Cards method, a complete reciprocal and consistent comparison system is reconstructed, and the resulting additive scale or multiplicative weights are normalized for use in ELECTRE concordance calculations (García-Zamora et al., 9 Oct 2025). Closely related uses of the label include the ELECTRE Tree ensemble for ELECTRE Tri-B parameter inference and a simulation-driven ELECTRE-like weight-tuning scheme for network ranking, which suggests that the expression refers to a family of AutoElicit mechanisms rather than a single standardized algorithm (Barros et al., 2020, Cao et al., 2024).

1. Terminological scope and main lineages

The phrase combines three ideas that recur in the ELECTRE literature: structured decomposition, automatic elicitation, and downstream outranking. What changes across papers is the meaning of “tree.” In one lineage, the tree is a spanning tree induced by tournament matches over criteria. In another, it is a Random-Forest–inspired ensemble of ELECTRE Tri-B base models. In a third, it is a prospective hierarchy of local and global criteria whose weights are tuned against diffusion outcomes.

Lineage	Meaning of “tree”	Auto-elicited parameters
Tournament Tree Method	Spanning tree over criteria induced by $m-1$ pairwise matches	Additive scale $u_j$ or multiplicative weights $w_j$; optionally $q_j,p_j,v_j$ through separate elicitation
ELECTRE Tree algorithm	Ensemble of bootstrap ELECTRE Tri-B models	Any subset of $w_j,q_j,p_j,v_j,r_h,\lambda$
SK-E extension	Possible hierarchy such as Local vs Global criteria	Optimal local/global weight; hierarchical extension proposed

In the TTM formulation, AutoElicit is primarily a low-burden preference elicitation device for weights and, optionally, pseudo-criteria thresholds. In the ensemble-based ELECTRE Tree algorithm, AutoElicit means inferring all or any combination of ELECTRE Tri-B parameters from assignment examples or ordered clusters. In SK-E, AutoElicit is a simulation-based search for the weight that maximizes spreading performance. This terminological plurality is important because the same label can denote distinct mathematical objects and optimization targets (García-Zamora et al., 9 Oct 2025, Barros et al., 2020, Cao et al., 2024).

2. Tournament-tree elicitation over criteria

In the TTM-based formulation, criteria weights are elicited by organizing the criteria set $J=\{1,\ldots,m\}$ as a sports-like tournament. In each round, items are paired; if the number of items is odd, one item is virtually doubled to pair with itself, so no comparison is needed for that case. For each actual pair, the expert identifies the winner and specifies the difference of attractiveness via the Deck of Cards method, with $t$ blank cards encoded as $p=t+1$ units. Winners advance to the next round until one champion remains.

This structure yields exactly $m-1$ pairwise inputs. Those $m-1$0 winner-loser judgments form a spanning tree over the $m-1$1 criteria. The method is explicitly organized in three stages. Stage I is the tournament elicitation itself. Stage II reconstructs a complete pairwise comparison matrix from the tree edges. Stage III derives a global value scale and applies a normalization. The paper states the guarantees in exact terms: total matches $m-1$2, number of rounds $m-1$3, elicitation complexity $m-1$4, and $m-1$5 rounds. It also states that filling the complete matrix is $m-1$6, whereas computing the additive scale $m-1$7 or multiplicative weights $m-1$8 from the tree is $m-1$9.

The central claim of the method is dimensionality reduction without post hoc consistency repair. Classical Fuzzy Preference Relations and multiplicative/AHP matrices require $u_j$0 distinct pairwise judgments and are exposed to reciprocity and transitivity violations. TTM reduces the required judgments to $u_j$1, collapses the preference representation to $u_j$2 parameters, and guarantees consistency by construction rather than by ex post correction. The same framework is stated to be compatible with the classical Deck of Cards method, so interval and ratio scales can both be handled. A web application guides experts through the targeted comparisons across $u_j$3 rounds, constructs the consistent matrix, derives and normalizes the value scale, displays the ranking, and allows interactive adjustment of the ranking or the number of cards between consecutive objects (García-Zamora et al., 9 Oct 2025).

3. Mathematical reconstruction and insertion into ELECTRE

The additive TTM construction defines a matrix $u_j$4 in which $u_j$5 is the signed difference of attractiveness of criterion $u_j$6 over criterion $u_j$7. The sign convention is $u_j$8 if $u_j$9, $w_j$0 if $w_j$1, and $w_j$2 for no difference, with reciprocity $w_j$3. For an elicited pair $w_j$4, the Deck of Cards intensity is encoded as $w_j$5. The tournament outputs are stored in a match matrix $w_j$6, whose last row records the champion. Reconstruction then proceeds by initializing $w_j$7, placing the observed edge values, propagating comparisons from each loser to the champion, and completing all remaining entries through the champion $w_j$8 via

$w_j$9

The key lemma states that if there exists $q_j,p_j,v_j$0 such that $q_j,p_j,v_j$1 for all $q_j,p_j,v_j$2, then $q_j,p_j,v_j$3 is additively consistent. The TTM theorem states that the algorithm ensures exactly this property with $q_j,p_j,v_j$4 equal to the champion, hence additive reciprocity and transitivity are guaranteed. The representation theorem further states that additive consistency is equivalent to the existence of $q_j,p_j,v_j$5 such that $q_j,p_j,v_j$6; under TTM, one sets $q_j,p_j,v_j$7.

The multiplicative extension defines a comparison matrix $q_j,p_j,v_j$8 with reciprocity $q_j,p_j,v_j$9 and multiplicative transitivity $w_j,q_j,p_j,v_j,r_h,\lambda$0. Each elicited tree edge is encoded as a ratio $w_j,q_j,p_j,v_j,r_h,\lambda$1 if $w_j,q_j,p_j,v_j,r_h,\lambda$2 is judged more attractive than $w_j,q_j,p_j,v_j,r_h,\lambda$3, otherwise inverted. For any two criteria $w_j,q_j,p_j,v_j,r_h,\lambda$4, the complete matrix entry is obtained by multiplying edge ratios along the unique path $w_j,q_j,p_j,v_j,r_h,\lambda$5. Equivalently, fixing a root $w_j,q_j,p_j,v_j,r_h,\lambda$6 and setting $w_j,q_j,p_j,v_j,r_h,\lambda$7, the unique path from $w_j,q_j,p_j,v_j,r_h,\lambda$8 to $w_j,q_j,p_j,v_j,r_h,\lambda$9 yields $J=\{1,\ldots,m\}$0, and then $J=\{1,\ldots,m\}$1. The paper also gives the log-additive version $J=\{1,\ldots,m\}$2, $J=\{1,\ldots,m\}$3, with propagation on the tree and recovery by exponentiation.

Inserted into ELECTRE, the workflow is explicit. One builds a TTM over criteria, collects $J=\{1,\ldots,m\}$4 pairwise intensity judgments, computes either an additive scale $J=\{1,\ldots,m\}$5 or multiplicative weights $J=\{1,\ldots,m\}$6, and normalizes them, for example by $J=\{1,\ldots,m\}$7. If pseudo-criteria thresholds are needed, one may build a separate TTM over predefined performance levels for each criterion or use a simple ordered Deck of Cards scaling to elicit coherent sets for $J=\{1,\ldots,m\}$8, $J=\{1,\ldots,m\}$9, and $t$0. For two alternatives $t$1 and $t$2, and criterion $t$3, the generic partial concordance given in the paper is

$t$4

with global concordance

$t$5

The resulting weights feed directly into ELECTRE I, II, III, or IV; discordance and veto are then applied if required by the selected variant, followed by distillation or kernel extraction. In the additive case, the paper uses a max-based normalization $t$6, where $t$7 is the least attractive item, so that $t$8 (García-Zamora et al., 9 Oct 2025).

4. Ensemble-based ELECTRE Tree for ELECTRE Tri-B

A distinct meaning of ELECTRE Tree appears in the algorithm for inferring ELECTRE Tri-B parameters through a Random-Forest–inspired ensemble. Its goal is to elicit all or any combination of Tri-B parameters, including criteria weights, indifference, preference, and veto thresholds, reference profiles, and the lambda cutting level. The decision-maker may keep trusted parameters fixed and let the algorithm infer the remainder. Two sources of supervision are admitted: assignment examples supplied by the decision-maker and ordered clusters derived from the data itself.

The ensemble construction samples, with replacement, both alternatives and criteria. Each base model uses at least two criteria and between $t$9 and $p=t+1$0 of the alternatives. For each sampled subproblem, a genetic algorithm optimizes the selected Tri-B parameters against the reference labels. The paper specifies chromosome components for weights, thresholds, profiles, and $p=t+1$1, together with feasibility conditions $p=t+1$2, $p=t+1$3, monotone profiles for gain criteria, and $p=t+1$4. Fitness is assignment accuracy on the bootstrap sample. Selection can be tournament or roulette-wheel; crossover is Simulated Binary Crossover; mutation is Real-coded Jumping Gene GA.

Aggregation is performed in two different ways. In the merged model, parameters from all base models are averaged and projected back onto the feasible region. For weights, the average is taken over models containing criterion $p=t+1$5, then normalized so that $p=t+1$6. Thresholds are averaged and clipped to preserve $p=t+1$7. Profiles are averaged and then made monotone. Lambda is averaged and clipped to $p=t+1$8. The paper associates this merged regime with effectively linear decision boundaries. In the voting regime, each base model classifies each alternative and the final class is chosen by majority or plurality vote; this regime is associated with non-linear boundaries because the decision regions are the majority superposition of multiple parameterized models.

The empirical section reports both synthetic and applied results. On a synthetic dataset with $p=t+1$9 alternatives, $m-1$0 gain criteria, and $m-1$1 categories, training used $m-1$2 base models and $m-1$3 per model; base-model accuracies averaged approximately $m-1$4, voting produced smooth non-linear boundaries, and the merged model achieved approximately $m-1$5 accuracy. On the ESL Employment dataset with $m-1$6 alternatives and $m-1$7 normalized gain criteria, the reported training accuracies were approximately $m-1$8 for voting and $m-1$9 for the merged model, while test accuracies were approximately $m-1$00 and $m-1$01, respectively. In this lineage, AutoElicit therefore means automatic calibration of a full ELECTRE Tri-B model, not merely low-burden weight elicitation (Barros et al., 2020).

5. Simulation-driven AutoElicit and the criteria-tree extension

A third usage appears in SK-E, an improved ELECTRE multi-attribute fusion method for important node identification in complex networks. Here the published model is not an explicit criteria tree: it uses two flat criteria, a local metric and a global metric, and then tunes their trade-off automatically. The local metric is

$m-1$02

where $m-1$03 is the degree of node $m-1$04. The global metric is $m-1$05, built from the K-shell index $m-1$06 and the first-order neighbor K-shell sum $m-1$07: $m-1$08

After min-max normalization, the weighted decision matrix is formed as $m-1$09, with $m-1$10, $m-1$11, $m-1$12, and $m-1$13. The method then computes a harmony index

$m-1$14

where $m-1$15, and a disharmony index

$m-1$16

where $m-1$17. Comprehensive dominance is $m-1$18, and ranking is based on the net dominance coefficient

$m-1$19

AutoElicit is implemented as a grid search over $m-1$20, with $m-1$21, and validation through SIR diffusion. For each candidate weight, the top-$m-1$22 ranked nodes are used as seeds; $m-1$23 Monte Carlo SIR simulations are run for $m-1$24 days; and the chosen weight maximizes

$m-1$25

The reported optimal $m-1$26 values are $m-1$27 for Social, $m-1$28 for Friend, $m-1$29 for Power, $m-1$30 for Road, $m-1$31 for Dolphin, and $m-1$32 for Usair. The authors state explicitly that SK-E uses two flat criteria rather than an explicit criteria tree, but they also describe how the same AutoElicit logic can be extended into an ELECTRE Tree in which a root criterion splits into Local and Global branches and weights are optimized at each level against diffusion outcomes or controllability measures. In that sense, SK-E provides a direct foundation for a tree-structured AutoElicit scheme, although the published implementation remains flat (Cao et al., 2024).

6. Related outranking elicitation, assumptions, and recurring misconceptions

A related but distinct development is Electre-Score, the first outranking-based method for assigning a score to each alternative. Instead of producing a point estimate, it assigns a score interval $m-1$33 by comparing each alternative with ordered reference sets $m-1$34, each set being assigned a reference score $m-1$35 through a deck-of-cards procedure. The method uses pseudo-criteria, concordance, discordance, credibility, and a cutting level $m-1$36, but it does not support hierarchical criteria directly in its presented form. Its relevance to ELECTRE Tree AutoElicit lies in the common use of Deck of Cards for interval scaling, the explicit treatment of imperfect knowledge, and the possibility of interactive elicitation of reference structures and spacing constraints (Figueira et al., 2019).

Several assumptions and misconceptions recur across the three AutoElicit lineages. A common misconception is that “tree” always means a hierarchical criteria tree. In fact, the TTM tree is a spanning tree over elicited pairwise comparisons, the ELECTRE Tree algorithm is an ensemble metaphor inspired by Random Forests, and SK-E is published as a flat two-criterion ELECTRE-like fusion with a proposed hierarchical extension. The TTM formulation further assumes strict positivity for ratio judgments, requires a winner per match, interprets “0 cards” as the minimal difference equal to $m-1$37 unit rather than indifference, and handles odd $m-1$38 by virtual doubling. It states that any spanning tree yields a consistent matrix, although different trees may induce different cognitive paths. The ensemble-based Tri-B approach is computationally heavier than a single optimization, and the paper notes an identifiability issue: multiple parameter sets can induce the same assignments, so partial elicitation improves identifiability. SK-E, by contrast, ties weight elicitation to the chosen diffusion model and its parameters, and its pairwise outranking stage is $m-1$39, which can be burdensome for very large networks. Taken together, these points indicate that ELECTRE Tree AutoElicit is best understood as a class of principled, non-compensatory elicitation schemes whose exact semantics depend on whether the tree encodes a comparison skeleton, an ensemble architecture, or a criteria hierarchy (García-Zamora et al., 9 Oct 2025, Barros et al., 2020, Cao et al., 2024).