{"id":"de71d312-647a-48e7-9526-24c489c8580c","arxiv_id":"2505.04019","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new explainability method converts Isolation Forest models into directed graphs and assigns each split an Inlier-Outlier Propagation Score measuring how strongly a feature directs data toward outlier classification.","lead":"This paper turns a trained Isolation Forest, an algorithm for detecting outliers, into a graph that shows which features and split directions push samples toward outlier or inlier labels. The authors add a score to each graph node so users can see whether a feature is strongly tied to outlier detection, and they test the idea on synthetic and thyroid data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deleting split values in Section 3.1 erases the threshold logic needed to support the paper's claimed 'feature boundary' explanations, leaving the central 'comprehensive' claim unsupported.","rationale":"The paper is clear, reproducible, and the IOP-Score is mathematically simple. The synthetic experiments are sanity checks on data where the true outliers are known, but they do not establish that the pair-based DPG faithfully represents iForest's decision logic. The reader's weakest-assumption analysis identifies deletion of split values and deletion of max-depth outlier paths; my stress-test lands on the same threshold-deletion issue, which is the more fundamental one. The method's own stated goal is to explain decision boundaries, yet Section 3.1 removes the only information that specifies a boundary. The middle-gap counterexample is not exotic: it is a natural one-dimensional analogue of the multi-modal inlier structure common in anomaly detection, and it shows that direction-only predicates can become ambiguous or contradictory. The max-depth removal is a separate distortion because iForest's score averages paths over all trees, so discarding long outlier paths for 'clarity' changes what is being explained. Neither issue is a mathematical inconsistency in the IOP-Score formula itself; both are failures of the construction to match the paper's stronger interpretive claims. Since the authors already acknowledge in the Limitations section that a deeper comparison with SHAP and DIFFI is needed, and since the empirical evidence is qualitative, the conditional verdict remains appropriate. My read does not change the reader's verdict, hence UNCHANGED.","tokens_in":13771,"tokens_out":4300,"duration_ms":49997,"concrete_test":"Run the released code on a one-dimensional synthetic dataset with inliers drawn uniformly from [0, 0.1] U [0.9, 1.0] and one outlier at 0.5, using 200 trees as in the paper. If the resulting DPG contains both X> and X<= predicates with comparable IOP-Scores and no output encodes that the outlier lies between the two inlier modes, the removal of v in Section 3.1 is responsible. Then repeat with a variant that keeps the full triple (f, sigma, v) and check whether the graph reveals the central interval. If the pair-based graph cannot while the triple-based graph can, the central 'boundary explanation' claim requires either a threshold-preserving representation or a softened claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the DPG representation plus IOP-Score gives a comprehensive global explanation of iForest, including 'feature boundaries' and the direction in which features drive outlierness. Section 3.1 explicitly removes the split value v from each predicate, keeping only pairs (f, sigma). This is not a harmless aggregation: an Isolation Tree separates a point because of a specific threshold, and the same feature-direction pair can describe very different regions depending on where v falls. Concretely, with inliers clustered near 0 and near 1 and one outlier at 0.5, splits isolating the outlier will sometimes be 'X <= v' (when v is chosen above 0.5) and sometimes 'X > v' (when v is chosen below 0.5). The DPG will then contain both X<= and X> predicates with comparable frequencies leading to the Outlier class, IOP-Scores near zero, and no way to tell the user that the outlier lies in the middle gap. The actual boundary information has been discarded. Since the abstract and contribution list promise 'insights into decision boundaries' and 'details on feature boundaries for both inliers and outliers,' the method overclaims what its representation can deliver. A second, smaller distortion is the max-depth exclusion in Section 3.1: outlier traversals that reach full tree depth are deleted, even though iForest's anomaly score averages path lengths over all trees, so the graph over-represents short paths and may misstate how the model actually reaches its decisions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a post-hoc, global explanation method for Isolation Forest (iForest) based on Decision Predicate Graphs (DPG). The authors extract predicate triples (feature, operator, split value) from the internal nodes of iTrees, then delete the split values, keeping only (feature, direction) pairs. They construct a weighted directed graph over these predicates, introduce a class-based frequency weighting scheme, and define the Inlier-Outlier Propagation Score (IOP-Score) to quantify each predicate's tendency to lead toward the inlier or outlier class. The approach is demonstrated on two synthetic datasets and the Annthyroid benchmark, with qualitative visual and score-based interpretations. The central claim is that the method provides a 'comprehensive global explanation' of iForest, including 'feature boundaries' and the direction in which features drive outlier identification.","tokens_in":14048,"tokens_out":3670,"duration_ms":39508,"significance":"The paper addresses a real gap: most iForest explanation methods produce feature-importance vectors rather than structural descriptions of the model's logic. If the proposed graph representation were faithful, it would offer a new, visually interpretable modality for understanding iForest. The method is clearly specified, the equations for the weighting and IOP-Score are explicit, and the authors state that the full implementation is publicly available, which supports reproducibility. However, the central promise of explaining 'feature boundaries' is not actually delivered by the representation, because split values are deliberately removed. The experimental section is entirely qualitative and does not compare against existing methods such as SHAP, DIFFI, or ExIFFI, despite the introduction framing the work as overcoming their limitations. The strength of the graph-based idea is real, but the evidence provided is insufficient to support the breadth of the claimed contribution.","major_comments":[{"comment":"The removal of split values from each predicate triple, reducing (f,σ,v) to (f,σ), is load-bearing for the paper's central claim of providing 'insights into decision boundaries' and 'details on feature boundaries for both inliers and outliers.' Without the threshold value, the representation cannot distinguish between, for example, 'X <= 10' and 'X <= 100', which are completely different decision regions in the original iForest. A concrete failure case: if inliers are clustered near 0 and near 1 and an outlier lies near 0.5, different random splits will sometimes produce 'X <= v' and sometimes 'X > v' as the predicate that isolates the outlier; the graph would contain both directions with comparable frequencies, IOP-Scores near zero, and no way to tell the user that the outlier is in the middle gap. Thus the proposed representation, as defined, cannot substantiate the advertised 'feature boundary' explanations.","section":"Section 3.1, abstract, contribution list"},{"comment":"The step 'Remove paths of predicates exceeding treed max for outliers' rests on the statement that 'iForest classifies observations that reach the maximum tree depth as inliers.' This is not correct for the standard Isolation Forest. In the scoring formula given in Section 2.1, the anomaly score is s(x,n)=2^{−E(h(x))/c(n)}, where E(h) is averaged over all trees; an instance whose path length equals the maximum depth can still have s(x,n)>0.5, so it would be classified as an outlier. For example, with n=200 and default max depth 8, c(200)≈9.75 and s=2^{-8/9.75}≈0.566>0.5. Discarding all outlier traversals that reach the maximum depth therefore drops real outlier paths and biases the graph toward shallow, easy isolations, misrepresenting the model's actual decision process.","section":"Section 3.1, Algorithm 1, line 10"},{"comment":"The experimental evaluation is qualitative only: the authors inspect node colors, edge thicknesses, and IOP-Score tables on three datasets, and no quantitative comparison is made with SHAP, DIFFI, ExIFFI, or any other baseline, nor are fidelity or faithfulness metrics reported. The manuscript itself acknowledges in Section 5 that 'a more in-depth comparison with these techniques is necessary to establish the specific advantages and trade-offs of DPG.' Since the framing of the paper is that DPG overcomes the limitations of existing feature-importance methods, the absence of any comparative or quantitative validation leaves the central explanatory claim unsubstantiated.","section":"Section 4, Section 5"}],"minor_comments":[{"comment":"The keywords list contains 'Food quality · Food engineering', which is unrelated to the paper's content and appears to be a copy-paste error from another manuscript.","section":"Front matter"},{"comment":"The statement that exact triples (f,σ,v) 'are typically unique to individual trees and are not shared or reused across trees' is imprecise: split values drawn from the same feature range can coincide across trees, and the argument for aggregation does not require uniqueness.","section":"Section 3.1"},{"comment":"The IOP-Score definition refers to 'the frequency of the edge connecting node v to the Inlier class' and 'to the Outlier class', but the graph construction described earlier creates edges between predicates, not between predicates and terminal class nodes. Please clarify how class nodes are incorporated into the graph and whether the frequencies in Eq. (2) are the weighted frequencies introduced in Eq. (1) or the raw frequencies.","section":"Section 3.1, Eq. (2)"},{"comment":"Line 3 extracts predicates as triples (f,σ,v), but line 11 converts them to pairs (f,σ) only after the graph has already been described as having predicate pairs as nodes; a reader cannot tell when the value removal happens relative to graph construction.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising, but the advertised 'comprehensive explanation including feature boundaries' is not supported by the representation once split values are removed. The max-depth filtering step is not just a heuristic choice; it is based on an incorrect understanding of iForest's scoring function and can remove true outlier paths. I would encourage the authors to either restore threshold information (possibly through a more compact representation, e.g., approximate ranges) or reframe the method as a 'global feature-direction importance' technique rather than a boundary-explanation technique. A quantitative comparison against existing iForest explanation methods would also greatly strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on anomaly detection explainability. The paper adapts the authors' earlier Decision Predicate Graph (DPG) to Isolation Forest, producing a global graph of (feature, direction) predicates plus a per-node score. The code is public, the construction is clearly specified, and the IOP-Score is simple and internally consistent. The synthetic examples and the Annthyroid demo show the graph can highlight the right features in easy cases. That is a genuine, if modest, addition to the XAI toolbox.\n\nThe problem is the 'comprehensive' claim. In Section 3.1 they delete the split value v from each predicate, keeping only pairs like (F4, >). The abstract and contributions promise 'insights into decision boundaries' and 'details on feature boundaries'. But a boundary is a threshold. Without v, the graph cannot distinguish 'F4 greater than 3.2' from 'F4 greater than 9.7'. The stress-test counterexample is concrete: inliers near 0 and 1, one outlier at 0.5. Some trees split above 0.5 (outlier goes left, F<=), some below (outlier goes right, F>). The graph will contain both F<= and F> predicates, IOP-Scores will cancel toward zero, and the explanation is silent about the actual gap where the outlier lives. That is not a boundary explanation.\n\nA second, smaller distortion: outlier paths that reach maximum depth are deleted (Section 3.1). That biases the graph toward short paths, though iForest's anomaly score averages all paths. The authors give a rationale, but it remains a selection on the thing being explained.\n\nThe experiments are sanity checks, not validations. There's no quantitative comparison with SHAP or DIFFI, and the Limitations section concedes comparison is still needed. So the 'surpassing' language in the introduction is not supported.\n\nNone of this is fatal. The core graph construction is sound as a feature-direction summary. The fix is to either restore thresholds at some granularity (e.g., per-feature quantile buckets) or change the claims from 'boundaries' to 'directional feature usage'. With that, plus an actual comparison against existing explainers, this becomes a solid niche contribution.\n\nWho it's for: practitioners who want a visual, global summary of iForest behavior and are okay with losing exact thresholds. It deserves serious peer review; send it out, but flag the threshold issue to the authors.","headline":"A clearly specified, code-public XAI extension for Isolation Forest that loses the threshold information its 'boundary' claims depend on; worth a revise-and-resubmit, not a desk reject.","tokens_in":14577,"tokens_out":1759,"would_cite":false,"duration_ms":16387,"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":"Turning an Isolation Forest into a graph of direction-annotated predicates, scored by the Inlier-Outlier Propagation Score, yields a global explanation of which features and in which direction drive outlier classification.","keywords":["Explainable AI","Isolation Forest","Decision Predicate Graph","global explanation","outlier detection","IOP-Score","feature importance","model interpretability"],"falsifier":"Train two Isolation Forests on the same data with different random threshold ranges so the same (feature, direction) predicates occur with matching frequencies while thresholds sit in very different places; if the DPGs and IOP-Scores are nearly identical but the models disagree on which points are outliers, the deleted threshold values carry information the explanation omits.","tokens_in":13536,"feed_emoji":"🌲","tokens_out":11471,"duration_ms":104018,"temperature":0.7,"pith_summary":"This paper tries to establish that Isolation Forest, an ensemble outlier detector, can be given a global explanation by converting its trees into a Decision Predicate Graph. Each graph node is a predicate pairing a feature with a direction ($\\le$ or $>$), with split values removed so predicates can be compared across trees. The paper adds the Inlier-Outlier Propagation Score, which measures each predicate's tendency to push samples toward the outlier or inlier class. If the method works, a user can see which features separate anomalies, in which direction the separation goes, and whether a feature acts directly or only through longer decision paths. This matters because existing explanations of Isolation Forest mostly reduce to feature-importance vectors, which say which features matter but not how the model uses them.","feed_headline":"Graph reveals which features push Isolation Forest outliers","feed_subtitle":"A new score shows not just which features matter, but whether high or low values signal an anomaly.","key_machinery":"The central object is the Decision Predicate Graph (DPG) specialized to Isolation Forest. Each internal split rule of an isolation tree is first written as a triple (feature, sign, value), then the value is deleted so nodes become pairs such as $F_0 >$ or $TSH \\le$, making predicates comparable across trees that randomly chose different thresholds. Nodes are connected by directed edges weighted by how often training samples satisfy the two predicates consecutively, with outlier and inlier frequencies reweighted to offset class imbalance. The Inlier-Outlier Propagation Score is defined per node as $(f_i(v)-f_o(v))/f_{in}(v)$, where $f_i$ and $f_o$ are the weighted frequencies of transitions from $v$ toward inlier and outlier classes and $f_{in}$ is the total weighted frequency entering $v$. This score is what turns the graph from a visualization into a quantitative statement about which predicates are discriminative and in which direction.","core_discovery":"The central claim is that representing a trained Isolation Forest as a weighted directed graph of direction-annotated feature predicates yields a global explanation of the model's outlier logic. Predicates of the form (feature, sign) become nodes; an edge from one predicate to the next records how often training samples satisfy them consecutively while traversing the trees, with frequencies reweighted so the rare outlier class is not drowned out by inliers. The IOP-Score for a node is the normalized difference between the frequency of transitions heading to the inlier class and the frequency heading to the outlier class; a score near $-1$ marks a predicate that strongly characterizes outliers, a score near $+1$ marks an inlier-defining predicate, and zero marks a neutral one. The paper argues that the graph plus these scores shows the features that isolate anomalies, the direction of the constraints (outliers above or below a split), and whether a predicate separates outliers directly or only when combined with further splits.","pith_inferences":["My inference: the same predicate-pair abstraction could be applied to other tree ensembles whose splits use random thresholds, wherever a global explanation of boundary direction is wanted.","My inference: because split values are discarded, the same predicate pair can merge splits with very different thresholds; whether this loses fidelity is testable by comparing graphs from models trained with different threshold ranges.","My inference: the class-reweighting scheme assumes the model's own outlier and inlier labels are reliable, so noisy labels would propagate into the edge weights and the IOP-Score.","My inference: the IOP-Score could also work as an unsupervised feature-ranking signal, with predicates near $-1$ flagging candidate anomaly-relevant variables."],"forward_implications":["Users of an Isolation Forest can inspect one graph to see which features the model relies on to mark points as outliers, rather than only a ranked list of importances.","The sign attached to each predicate reveals the direction of the boundary, for instance that high TSH values, not low ones, drive the outlier class in the Annthyroid example.","Because edges carry reweighted frequencies, the graph also shows whether a predicate separates outliers immediately or only as part of a longer chain of conditions.","The IOP-Score gives a bounded, comparable value for every predicate, letting practitioners tell which features are discriminative for anomalies and which are neutral.","As a global explanation of the whole ensemble, the method supports auditing and fine-tuning of Isolation Forest rather than explaining only individual predictions."],"supporting_citations":[{"why":"defines Isolation Forest, the algorithm whose decisions this paper explains.","marker":"[7]"},{"why":"introduces the Decision Predicate Graph technique that this paper extends to Isolation Forest.","marker":"[12]"},{"why":"provides the additive feature-contribution explanation approach whose local, vector-based nature motivates the gap toward graph-based global explanations.","marker":"[10]"},{"why":"presents a depth-based feature importance method for Isolation Forest, the main model-specific alternative the paper contrasts with.","marker":"[11]"},{"why":"supplies the Isolation Forest implementation used in the experiments.","marker":"[19]"},{"why":"offers the comparative evaluation of anomaly detection algorithms that frames the real-world case study.","marker":"[21]"},{"why":"introduces the benchmark collection from which the Annthyroid dataset is drawn.","marker":"[20]"}],"fun_headline_variants":["Graph and IOP-Score reveal why Isolation Forest flags outliers","New score shows which features and directions mark outliers","Isolation Forest explained: feature-predicate graph and IOP-Score","Global explanation for Isolation Forest via directed feature graph","How anomalies are isolated: graph-based score exposes feature logic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the idea that dropping the actual threshold values from each split leaves a graph that still faithfully represents how the model isolates outliers.","fun_headline_variants_meta":{"raw":{"variants":["Graph and IOP-Score reveal why Isolation Forest flags outliers","New score shows which features and directions mark outliers","Isolation Forest explained: feature-predicate graph and IOP-Score","Global explanation for Isolation Forest via directed feature graph","How anomalies are isolated: graph-based score exposes feature logic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1604,"prompt_tokens":971,"completion_tokens":633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":551}},"tokens_in":587,"tokens_out":633,"duration_ms":6469,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:39:34.276264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two Isolation Forests on the same data with different random threshold ranges so the same (feature, direction) predicates occur with matching frequencies while thresholds sit in very different places; if the DPGs and IOP-Scores are nearly identical but the models disagree on which points are outliers, the deleted threshold values carry information the explanation omits.","supporting_citations":[{"cited_title":"Decision predicate graphs: Enhancing interpretability in tree ensembles","cited_arxiv_id":null,"evidence_quote":"introduces the Decision Predicate Graph technique that this paper extends to Isolation Forest."},{"cited_title":"Lundberg and Su-In Lee","cited_arxiv_id":null,"evidence_quote":"provides the additive feature-contribution explanation approach whose local, vector-based nature motivates the gap toward graph-based global explanations."}],"review_version":1}