{"id":"b9d790b4-5e98-4a6d-9742-8d32a276a33c","arxiv_id":"2507.12192","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"This paper proposes IEMM, a decision-tree explainer for evidential clustering that minimizes a utility-weighted 'evidential mistakeness' cost.","lead":"Evidential clustering assigns each data point a belief spread over several possible clusters, so a single 'which cluster?' answer is often wrong. This paper introduces a utility-based error measure and a decision-tree algorithm, IEMM, that builds cautious explanations tuned to how much a decision-maker tolerates mistakes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IEMM's greedy split search is not shown to minimize the total evidential mistakeness that Theorem 2 links to representativeness, so the algorithm's practical claim remains unsupported.","rationale":"I checked the Theorem 2 proof. Up to the Appendix C sign/notation errors, the equivalence is correct: with one centroid per leaf, each leaf's label A determines Delta(x), and total Eq(6) = κ−|X|R while total Eq(7) = |X|−|X|R, both affine decreasing in R. So the stated theoretical core is not false. The vulnerability is the step from the objective to the algorithm. Algorithm 1 is a greedy recursive partitioning method; local minimization of split cost does not imply global minimization of the leaf-summed objective. The paper provides no approximation guarantee, and its experiments do not include an optimal or even a stronger baseline. The 93% headline is selected across λ and evaluated with the same utility being optimized, as the reader noted. I do not think these issues make the theorem wrong; they make the paper's central practical claim conditional. The reader's weakest assumption was about expressiveness of the one-leaf-per-metacluster class; that is a related but distinct concern. I would keep the CONDITIONAL verdict: the theory can stand, but the algorithm needs either an optimality/approximation analysis or a clearly heuristic status, and the evaluation needs an external or at least non-circular check.","tokens_in":17768,"tokens_out":21687,"duration_ms":242086,"concrete_test":"Run a brute-force enumeration of all IMM-like trees on small synthetic instances (2 features, 4 focal sets, ~100 points, fixed utility such as U^0) by dynamic programming over sorted coordinate thresholds. Compute total mistakeness and R for every tree. Run IEMM on the same instances and compare its output's R and total mistakeness with the enumerated optimum. Repeat over at least 100 random evidential partitions. If IEMM's output is ever strictly worse than the optimum, the greedy search does not implement the theoretical minimization, and the paper must either add an optimality/approximation analysis or reframe IEMM as a heuristic without claiming the Theorem 2 equivalence applies to its output.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2 is internally sound once the Appendix C notation is fixed: for any IMM-like tree with one centroid per leaf, the two total mistakeness forms are affine functions of R (T6 = κ−|X|R, T7 = |X|−|X|R), so a global minimizer of either mistakeness among IMM-like trees also maximizes R. The load-bearing gap is that Algorithm 1 is not proven to find such a global minimizer. Line 11 greedily chooses the split minimizing the local 'mistakeness of cut' and then recurses; no exchange argument, approximation guarantee, or optimality proof is given. Greedy splits can preclude better subsequent splits and yield a final tree whose total mistakeness is higher (and R lower) than another IMM-like tree. The experimental section reports only the R of the greedy output and never compares it with the R of the optimal IMM-like tree. Therefore the inference 'minimizing evidential mistakeness is equivalent to maximizing representativeness, and IEMM provides a practical greedy algorithm to do so' conflates a property of the global objective with a property of the search heuristic. This is load-bearing because the 93% satisfaction figure and the claimed connection to Theorem 2 depend on the algorithm actually minimizing (or approximately minimizing) the stated objective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a framework and an algorithm for explaining evidential clustering results with decision trees. It introduces a utility-based representativeness score R_{M,U} for cautious explainers and two forms of evidential mistakeness (Eqs. 6 and 7). Theorem 2 claims that, for IMM-like decision trees with one centroid per leaf, maximizing evidential representativeness is equivalent to minimizing either mistakeness measure. The IEMM algorithm greedily grows a decision tree by splitting on coordinate thresholds to minimize a local mistakeness. Experiments on synthetic and real datasets report representativeness values up to 93%.","tokens_in":17974,"tokens_out":7755,"duration_ms":85544,"significance":"The paper addresses a genuine gap—explainability for evidential clustering—and the parameterized utility family for 'tolerable mistakes' is a useful idea. If the theoretical and algorithmic claims are made precise, the work would be a solid extension of the IMM approach to cautious, imprecise settings. The release of the code is a strength. However, the practical claim currently rests on a greedy algorithm with no optimality analysis, and the experiments are largely a self-consistency check rather than a comparative validation. With those gaps addressed, the paper could be a meaningful contribution to explainable clustering.","major_comments":[{"comment":"The split-selection function M(x, m, v, F, i, θ) is never defined in the text. This function is the core of IEMM: it is called in every non-terminal node to choose the best threshold, and the values labeled 'mistakeness of cut' in Figure 5 must be its output. Without a precise definition, the algorithm is not reproducible and the claimed connection to Eqs. (6) and (7) cannot be checked.","section":"3.2, Algorithm 1 (line 11)"},{"comment":"The paper states that IEMM 'fits a decision tree based on an evidential clustering by minimizing the evidential mistakeness function', but no result links the greedy split selection to global minimization of the total mistakeness. Theorem 2 is a statement about global optimizers over the class of IMM-like trees; it does not imply that the recursively chosen locally optimal cuts produce such a global optimizer. An exchange argument, an approximation guarantee, or an explicit acknowledgment that IEMM is a heuristic (with a gap evaluation against exact search on small instances) is needed. Without this, the 93% representativeness result cannot be attributed to the theoretical equivalence.","section":"3.2 and Theorem 2"},{"comment":"The proof of Theorem 2 contains several notational and algebraic errors. In the first displayed equation, the inner sum is written as Σ_{Δ(x)≠C} U(C,B), whereas Eq. (6) sums over metaclusters whose centroids lie outside S. In the second displayed equation, the denominator from Eq. (7) is omitted; it equals 1 for a singleton leaf, but this should be stated. The sums in the theorem range over A⊂Ω, although leaves are indexed by focal sets A∈F_M. Finally, κ_{M,U} is used as a scalar, but the preceding expression defines a sum over x with varying masses m_x; the definition of κ_{M,U} should be made explicit.","section":"Appendix C, proof of Theorem 2"},{"comment":"The experiments report only the representativeness R_{M,U} of the trained trees, which is the same objective that the algorithm optimizes during training. There are no confidence intervals, no repeated runs, and no baselines such as random thresholds, CART trained on the hard partition, or hard-label IMM. The '93% satisfactory' statement is therefore a self-consistency check: it shows that the greedy tree can achieve high values of the optimized score, not that the explanations are satisfactory relative to alternative explainers. Baselines and variability estimates should be added.","section":"3.2, 'The Tests' and Tables 3–4"},{"comment":"Algorithm 1 takes as input the focal sets F and their centroids v, assuming every focal set has a centroid. This is not guaranteed for arbitrary evidential partitions produced by common algorithms, which typically return prototypes for individual clusters rather than for every metacluster. The paper should state this assumption explicitly, discuss how such centroids are obtained, or restrict the algorithm's scope accordingly.","section":"3.2, Algorithm 1 input"}],"minor_comments":[{"comment":"The term 'IMM-like decision tree' is used in Theorem 2 but never formally defined; please define it (e.g., a decision tree with axis-aligned splits and exactly one centroid per leaf) before the theorem.","section":"3.1, Theorem 2"},{"comment":"The two mistakeness functions are both denoted with similar symbols 'M_{M,U}' and '\\mathcal{M}_{M,U}' in the text; please use clearly distinct notation to avoid confusion.","section":"3.1, Eqs. (6)–(7)"},{"comment":"There are small typos and grammatical issues, such as 'attributtes' in Appendix A and 'this have drawn criticism' in Section 1; these should be corrected.","section":"Appendix A and Section 1"},{"comment":"Leaf labels such as 'label: 3' are ambiguous; they should indicate the metacluster, e.g., '{ω3}', and the 'mistakeness of cut' values should explicitly reference Eq. (6) or Eq. (7).","section":"Figure 5"},{"comment":"The expression 'FM = S x∈X Fmx' should be written as the union ⋃_{x∈X} F_{m_x}.","section":"Section 2.2"},{"comment":"For the real-world datasets, please state the number of samples, the number of attributes, and how the evidential partitions were generated, since this information is needed to interpret the reported representativeness values.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core is promising, but the authors should be asked to either prove an optimality or approximation result for the greedy procedure, or explicitly reposition IEMM as a heuristic. The experimental section needs baselines and variance estimates; the current metric is the same as the training objective, so the 93% claim is a self-consistency result rather than evidence of superiority. I do not see these as unfixable, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper has a genuinely new cost function for explaining evidential clustering and a correct-looking core theorem, but the algorithm that carries the practical weight is underspecified and its optimality is unproven. The experiments are a self-consistency check rather than independent validation.\n\nWhat's new and good: the utility-based evidential mistakeness (Equations 6 and 7) is a sensible generalization of IMM's mistake count, and Theorem 2 correctly shows that, for IMM-like trees with one centroid per leaf, minimizing either mistakeness form is equivalent to maximizing U-evidential representativeness. The representativity-abductivity equivalence for decision trees (Theorem 1) is a modest complement to Amgoud and Ben-Naim. Code and data are provided. The paper is also honest about its constraints: shallow trees, informed audience, possibly rigid explanations.\n\nSoft spots, in order of seriousness:\n\n1. Algorithm 1 is not proven to minimize the global objective. The split-selection function M(x,m,v,F,i,theta) is never defined in the text, and there is no exchange argument or approximation bound showing the greedy recursive split search achieves the global minimum that Theorem 2 characterizes. This would be fine if the paper said 'greedy heuristic' and validated against external criteria, but the abstract and Section 3.2 imply the algorithm minimizes evidential mistakeness. The 93% figure is the representativeness of the greedy output at a selected lambda, not an optimality bound.\n\n2. The Appendix C proof contains algebraic and notational errors (e.g., indexing leaves by metaclusters but summing over A subset of Omega, and messy double-sum manipulation). The result is probably true; the errors are fixable. A referee should ask for a clean rewrite.\n\n3. The evaluation metric is the same representativeness the algorithm optimizes. Tables 3 and 4 show self-consistency, not explanation quality. No baselines, no error bars. The 'satisfactory up to 93%' claim is a maximum over lambda and datasets, so it should not be read as a typical-case result.\n\nThe theoretical contribution is worth preserving. I'd send this to peer review with a request for major revision: define the split objective, either prove an optimality/approximation guarantee or explicitly characterize the greedy heuristic, fix Appendix C, and add external validation (e.g., fidelity to the original credal partition under a different metric, or a small user study).\n\nWho's it for: researchers in evidential clustering and explainability who want a cautious decision-tree explainer with a principled cost. I'd bring it to reading group only if that's a working area for you; otherwise it's a browse-and-file paper.\n\nRecommendation: accept for review, but tell the authors the load-bearing algorithm claims need to be reined in or proven.","headline":"Sound theoretical core for a new evidential-clustering explanation cost, but the greedy algorithm that bears the practical claims is unproven and the experiments validate it against its own objective.","tokens_in":18546,"tokens_out":2151,"would_cite":false,"duration_ms":25668,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Evidential clustering can be explained by shallow decision trees, and for one-centroid-per-leaf trees minimizing evidential mistakeness is provably equivalent to maximizing explanation quality.","keywords":["explainability","cautiousness","unsupervised classification","evidential clustering","decision trees","belief functions","utility functions","abductive explanation"],"falsifier":"Construct a two-feature evidential clustering in which two metaclusters have centroids that lie on the same side of every allowed axis-aligned threshold, yet their member points are interleaved so that a deeper or oblique tree separates them cleanly. If on such data IEMM's best shallow tree has evidential representativeness far below 1 while a non-IMM tree is near-perfect, then minimizing evidential mistakeness within the IMM-like class fails to deliver faithful explanations.","tokens_in":17545,"feed_emoji":"🌳","tokens_out":12763,"duration_ms":128015,"temperature":0.7,"pith_summary":"This paper asks a question that existing explainability work skips: when a clustering algorithm returns not crisp labels but a credal partition — a mass function over subsets of clusters, expressing uncertainty and imprecision — what should an explanation even be? The authors show that, for decision trees, being representative (matching the clustering on every point) is both necessary and sufficient to give abductive explanations. They then carry this criterion into the evidential setting by defining utility functions $U(A,B)$ that score how tolerable it is to say $A$ when the truth is $B$, which turns explanation cost into an evidential mistakeness. The paper's main theoretical result is that, for the class of one-centroid-per-leaf trees used in the IMM algorithm, minimizing either form of evidential mistakeness orders trees exactly like maximizing evidential representativeness. On that foundation, the Iterative Evidential Mistake Minimization (IEMM) algorithm builds shallow, cautious decision trees, and the authors report explanations satisfactory up to 93% of the time when the decision-maker's preferences are taken into account.","feed_headline":"Shallow trees provably explain evidential clusters","feed_subtitle":"Utility-aware decision trees reach up to 93% explanation satisfaction by encoding tolerable mistakes.","key_machinery":"The load-bearing object is the pair of evidential mistakeness measures from Equations (6) and (7), together with the utility function $U$ that defines what counts as a tolerable mistake. Theorem 2 shows that, for IMM-like trees (one centroid per leaf), the total of either mistakeness and the evidential representativeness $R_{M,U}(\\Delta)$ are affinely related by a tree-independent constant, which is all that is needed for the equivalence of orderings. This affine identity is what carries the argument: it licenses replacing maximize explanation quality with minimize evidential mistakeness in the algorithm, and it is what makes the greedy per-node split choice in IEMM (Algorithm 1) a sound heuristic inside the restricted family.","core_discovery":"The central claim is that evidential clustering results can be faithfully explained by shallow decision trees once the notion of an explanation error is made utility-aware. Concretely, the paper proves two things. First, Theorem 1: for decision-tree explainers, representativity (the explainer's leaves match the clustering on every point) is equivalent to abductivity (every leaf gives a sufficient reason for its label), so a tree explains a clustering exactly when it reproduces it. Second, Theorem 2: for any two IMM-like trees — trees in which every leaf contains exactly one metacluster centroid — the ordering by evidential representativeness $R_{M,U}$, the ordering by the cost-of-not-assigning mistakeness, and the ordering by the expected-assignment-cost mistakeness are all identical, because the three quantities differ only by a constant that depends on the clustering and utility but not on the tree. Hence minimizing evidential mistakeness is provably equivalent to maximizing utility-weighted explanation quality inside the IMM-like family, and the choice of $U$ — equivalently the parameter $\\lambda$ — encodes how cautious the explanation should be. This is what the IEMM algorithm exploits greedily, and it is why the resulting trees are interpretably shallow while still being optimal within their class.","pith_inferences":["The theorem's equivalence is proven only within the IMM-like family; a natural extension would be to test whether representativeness and mistakeness also order trees identically when leaves may hold several centroids or splits may be oblique, which would justify applying the same greedy principle in a richer hypothesis space.","Because the proposed utilities depend only on set cardinalities, they treat every cluster symmetrically; plugging in an asymmetric utility that weights confusion between particular clusters more heavily would change the explanation without altering the algorithm.","The reported 93% satisfactory-explanation rate suggests that alignment between the decision-maker's tolerance and the tree's cautiousness drives perceived quality; a testable follow-up is whether this alignment matters more than tree size for real users.","The algorithm presupposes a centroid for every focal set; evidential clustering methods producing focal elements without a natural centroid would need an extended input representation before IEMM applies."],"forward_implications":["Because representativity and abductivity coincide for decision-tree explainers, any tree that reproduces a hard clustering gives provably sufficient reasons for its assignments, and this property extends pointwise to the evidential case through utility-weighted representativeness.","The parameter $\\lambda$ is a single dial that turns explanation policy: negative $\\lambda$ penalizes assigning points to metaclusters that cover too much (over-cautious errors), positive $\\lambda$ penalizes assigning metaclusters that cover too little (under-cautious errors), and the extremes $\\lambda \\to \\pm\\infty$ recover pure inclusion rules based on subset relations.","Minimizing evidential mistakeness is guaranteed to order IMM-like trees identically to maximizing evidential representativeness, so the greedy split choice in IEMM is aligned with its stated objective rather than being a loose approximation.","When the clustering is hard and $\\lambda = 0$, IEMM's mistakeness reduces to the IMM mistake count, so the new algorithm contains the classical explanatory decision-tree method as a special case.","IEMM trees are shallow by construction — at most $|F|-1$ internal nodes, one leaf per focal set — so interpretability is built into the algorithm, at the price of possible rigidity for complex patterns."],"supporting_citations":[{"why":"Supplies the definition of a representative explainer and the result that abductive explanations are representative, which Theorem 1 turns into an equivalence for decision trees.","marker":"[2]"},{"why":"Provides the linear-time conversion of decision trees to DNF expressions, the mechanism by which a tree induces an explainer.","marker":"[4]"},{"why":"The IMM algorithm and the definition of a clustering mistake that IEMM generalizes, plus the complexity baseline.","marker":"[29]"},{"why":"Foundational belief-function theory: mass functions, focal sets, belief and plausibility, on which evidential clustering rests.","marker":"[40]"},{"why":"Introduces evidential clustering and credal partitions, the objects the paper explains.","marker":"[28]"},{"why":"A library used to generate the evidential partitions in the synthetic experiments.","marker":"[42]"},{"why":"Supplies the real-world richly labeled datasets used for validation.","marker":"[18]"},{"why":"Basis of the decision-tree implementation accepting evidential labels, on which IEMM is built.","marker":"[17]"}],"fun_headline_variants":["Shallow trees provably explain evidential clusters with utility","Utility-aware trees hit 93% satisfaction for evidential clustering","Representativity equals abductivity: key to explaining evidential clusters","IEMM: interpretable optimal shallow trees for evidential clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that a decision tree with at most $|F|-1$ axis-aligned splits, one leaf per metacluster centroid, can faithfully represent the evidential clustering; if the true credal boundaries need deeper, oblique, or centroid-independent splits, the explanation can be systematically wrong while still being optimal within that restricted class.","fun_headline_variants_meta":{"raw":{"variants":["Shallow trees provably explain evidential clusters with utility","Utility-aware trees hit 93% satisfaction for evidential clustering","Representativity equals abductivity: key to explaining evidential clusters","IEMM: interpretable optimal shallow trees for evidential clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1459,"prompt_tokens":998,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":390}},"tokens_in":614,"tokens_out":461,"duration_ms":5344,"temperature":1.0,"reasoning_tokens":390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:52:16.367049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a two-feature evidential clustering in which two metaclusters have centroids that lie on the same side of every allowed axis-aligned threshold, yet their member points are interleaved so that a deeper or oblique tree separates them cleanly. If on such data IEMM's best shallow tree has evidential representativeness far below 1 while a non-IMM tree is near-perfect, then minimizing evidential mistakeness within the IMM-like class fails to deliver faithful explanations.","supporting_citations":[{"cited_title":"In: Proceedings of the 37th International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"The IMM algorithm and the definition of a clustering mistake that IEMM generalizes, plus the complexity baseline."},{"cited_title":"Princeton University Press (Apr 1976)","cited_arxiv_id":null,"evidence_quote":"Foundational belief-function theory: mass functions, focal sets, belief and plausibility, on which evidential clustering rests."},{"cited_title":"Pattern Recognition41(4), 1384–1397 (Apr 2008).https://doi.org/ 10.1016/j.patcog.2007.08.014","cited_arxiv_id":null,"evidence_quote":"Introduces evidential clustering and credal partitions, the objects the paper explains."},{"cited_title":"evclust: Python library for evidential clustering","cited_arxiv_id":"2502.06587","evidence_quote":"A library used to generate the evidential partitions in the synthetic experiments."},{"cited_title":"In: FUZZ (Jan 2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world richly labeled datasets used for validation."},{"cited_title":"Expert Systems with Applications230, 120652 (Nov 2023).https://doi.org/10","cited_arxiv_id":null,"evidence_quote":"Basis of the decision-tree implementation accepting evidential labels, on which IEMM is built."}],"review_version":1}