{"id":"7add94db-c383-431d-8557-a2ab663ba9a7","arxiv_id":"2502.03895","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ANFIS-PCA-BPSO applies PCA to normalized firing strengths and selects components with BPSO, reducing ANFIS rule count and training time at a small accuracy cost.","lead":"This paper proposes a rule-reduction method for ANFIS by applying PCA to normalized firing strengths and selecting components with binary particle swarm optimization. The method is reported to cut rule count and training time dramatically with only small accuracy changes, but the write-up has inconsistencies and no code is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 'fewer rules' counts PCA components, not fuzzy rules; a PCA component is a dense linear mix of all firing strengths, so the claimed interpretable rule reduction is not established.","rationale":"The load-bearing condition for the paper's central claim is that reducing 'rules' preserves interpretability. That condition fails at the PCA stage, before the reader's concern about frozen premise parameters even arises. Even if normalized firing strengths were perfectly reliable, PCA linearly mixes all rule firing strengths, so selecting a principal component is not selecting a rule. This is visible in the equations and in the reported rule counts. The dimensional mismatch in Eq. 19 reinforces that the pipeline is not fully specified. I would keep a conditional verdict: the contribution could survive if the authors demonstrate that the selected components have sparse, interpretable loadings or if they explicitly reframe the method as component-based ANFIS compression rather than fuzzy-rule reduction. As written, the abstract's 'fewer rules' claim should not be accepted at face value, and the conditionality should be tied to the loadings test above.","tokens_in":26665,"tokens_out":8728,"duration_ms":94205,"concrete_test":"Re-run the Iris experiment and save (i) the PCA loading matrix V_k applied to the normalized firing strengths and (ii) the fitted layer-4 consequent parameters. For each selected component r, compute sparsity = fraction of the M original rule weights with |a_{rj}| > 1e-6. Then compare three models on the same test split: (a) the published ANFIS-PCA-BPSO with 2 selected components; (b) a genuine 2-rule ANFIS whose antecedents are the two original rules with the largest absolute loadings in the selected components; (c) a negative control using the same 2 PCA components but with loadings replaced by one-hot vectors. If (a) and (b) differ by more than the reported accuracy gap, or if sparsity is below 80%, the rule count in Table 2 does not describe a fuzzy rule base, and the paper must report loadings and reframe the method or stop calling these components rules.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.3.1 (Eqs. 18-19) is internally inconsistent: MAT_Nfs is N×M, keys is M×D, so the Hadamard product is undefined unless D=M. More importantly, even under the intended pipeline—PCA on MAT_Nfs followed by BPSO on the D_reduced components—each retained 'rule' is a principal component score, i.e., a dense linear combination of all M original normalized firing strengths. In the consequent layer, the network output is Σ_{r∈S} (Σ_j a_{rj} wbar_j) f_r(x) = Σ_j wbar_j (Σ_{r∈S} a_{rj} f_r(x)). This is not a set of M or fewer fuzzy rules: every original rule still contributes through the loadings a_{rj}, and the antecedents are no longer conjunctions of input membership functions. Therefore Tables 2, 6, and 7, and the abstract's '2 rules' on Iris, report the number of selected principal components, not interpretable fuzzy rules. The claimed rule reduction is a category error unless the loadings are shown to be one-hot/sparse, which PCA generically does not produce. This undermines the central claim even before considering the reader's concern about frozen premise parameters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ANFIS-PCA-BPSO, a rule-reduction technique that inserts PCA and binary particle swarm optimization between layers 3 and 4 of ANFIS. The normalized firing strengths are first transformed by PCA; BPSO then selects a subset of components using a fitness function based on training error. The authors report experiments on eight classification and four regression benchmarks plus an ischemic stroke dataset, comparing accuracy/RMSE, number of rules, and training time against standard ANFIS and several published rule-reduction methods. Their main claims are fewer rules, shorter training, and preserved accuracy.","tokens_in":26897,"tokens_out":6068,"duration_ms":62761,"significance":"If correct, this would be a practically useful pruning strategy for low-dimensional ANFIS, and the paper has strengths: it evaluates across multiple benchmarks, reports standard deviations, includes ablation studies, and applies the method to a real medical dataset. However, the central claim is undermined by the fact that PCA components are dense linear combinations of firing strengths rather than fuzzy rules, so the reported 'rule counts' do not count interpretable fuzzy rules. The dimensional inconsistency in the formal definition and the feature-selection leakage in the real-data experiment further weaken the empirical support.","major_comments":[{"comment":"The formal definition of the proposed integration is inconsistent. MAT_Nfs is defined as N x M and keys is defined as M x D, so the Hadamard product in Eq. (19) is undefined unless N = M = D. The displayed result also claims a size of M x D, which contradicts the row dimension N of MAT_Nfs. Even under the natural correction of a matrix product, MAT_Nfs times keys would be N x D, not M x D. Because this product is the mathematical core of the proposed pipeline, the method is not reproducible as written.","section":"Section 4.3.1, Eqs. (18)-(19)"},{"comment":"The central claim that the method reduces the number of fuzzy rules is not supported by the described procedure. PCA on the normalized firing strengths produces principal components that are dense linear combinations of all M firing strengths. Selecting K of these components gives an output of the form sum over selected components of (sum over all original rules of loadings times firing strengths) times a consequent function; every original rule still contributes through the loadings, and the selected objects are not fuzzy rules with conjunctive antecedents. Reporting '2 rules' for Iris in Table 2, or the rule counts in Tables 6 and 7, therefore counts selected principal components, not fuzzy rules. Without evidence that the PCA loadings are one-hot or sparse—which PCA generically does not provide—the rule-reduction and interpretability claims in the abstract and contributions section are a category error.","section":"Abstract, Section 4.3, Tables 2, 6, and 7"},{"comment":"The optimization loop operates exclusively between layers 3 and 4, and only the consequent parameters are updated after the PCA-BPSO selection. The firing strengths used for PCA are therefore those of the initial grid-partitioned FIS, whose membership functions are never re-estimated. If the initial membership functions are poorly centered or scaled, the variance ranking from PCA may not reflect the true contribution of each rule to the final prediction. The paper does not provide an ablation that varies the initialization or re-tunes premise parameters, so the robustness of the selection step is unexamined.","section":"Algorithm 1, Section 4.3"},{"comment":"For the real-world ischemic stroke experiment, the input features are selected using p-values and Pearson correlations computed on the entire dataset (Table 8) before the 5-fold cross-validation is run. This is a selection-before-validation procedure and can optimistically bias the reported results. A nested cross-validation, or feature selection performed inside each training fold, is needed to support the real-data claim.","section":"Section 9, Table 8"}],"minor_comments":[{"comment":"The statement that the consequent parameter vector 'equals (n_inputs+1) x j' is dimensionally unclear; it should say that each rule j has a consequent parameter vector of length n_inputs+1.","section":"Section 2, Eq. (5)"},{"comment":"The training time of the baseline ANFIS has very large standard deviations (for example, 3.53e+03 +/- 6.65e+03 seconds on Iris); the source of this variance and the convergence criterion used for the baseline should be stated.","section":"Section 6.1, Table 2"},{"comment":"The Airfoil dataset is listed with 4 features, but the standard UCI Airfoil self-noise dataset has 5 input features; please verify the entry.","section":"Section 5.1, Table 1"},{"comment":"Reference [62] and reference [65] are the same paper and appear twice in the bibliography; the duplicate should be removed.","section":"References"},{"comment":"The conclusion states that the model's efficacy diminishes with datasets having a large number of features; this limitation restricts the scope to low-dimensional data and should be stated explicitly in the abstract as well.","section":"Section 10"}],"recommendation":"reject","confidential_remarks":"The empirical results may be salvageable if the work is reframed as a component-selection or feature-aggregation method, but the current manuscript's main interpretability claim is not supported by the described pipeline. The dimensional inconsistency in Eq. (19) and the feature-selection leakage in Section 9 are additional load-bearing issues. I recommend rejection in the present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the engineering, not for the stated claim. The pipeline—PCA on normalized firing strengths followed by BPSO selection between ANFIS layers 3 and 4—is a reasonable way to shrink the layer-3-to-4 computation, and I haven't seen this exact pairing in the literature. The paper does a wider-than-usual empirical job: 12 UCI/KEEL benchmarks, an ablation of PCA-only and BPSO-only, and a private stroke dataset, with standard deviations and comparisons to several rule-reduction baselines. On training time the gains are real and large (Iris 3530s to 13s, for example).\n\nThe soft spots are not minor. The stress-test note holds up: what BPSO selects after PCA are principal components, not fuzzy rules. A PCA component is a dense linear combination of all the normalized firing strengths, so every original rule still contributes through the loadings. Reporting '2 rules' for Iris is a category error unless the loadings are one-hot/sparse, which PCA generically is not. That directly undercuts the paper's interpretability claim, which is the main selling point. Related, Eq. (19) has a dimensional mismatch: MAT_Nfs is N×M and keys is M×D, so the Hadamard product is undefined unless D=M. I think the intended operation is clearer than the text says, but as written it's wrong.\n\nA few other things need attention. The BPSO fitness function is not fully specified—training error on what split?—and no code or data are provided, so the numbers can't be checked. The premise parameters are frozen after initialization, which the reader's report flags; that's a real limitation if the initial membership functions are poor. On the stroke dataset, feature selection uses p-values/correlations computed on the same data before model evaluation, which is leakage. And accuracy is not uniformly 'high': TAE, THY and BAL drop several points, and on the real dataset the baseline has better RMSE and correlation. The paper acknowledges the trade-off, which is honest, but the abstract overstates.\n\nWho's it for: people working on ANFIS rule explosion who want a computational shortcut, not people who need interpretable fuzzy rules. As written, I wouldn't cite the rule counts as evidence of interpretability. But the problem is real, the experiments are extensive, and the core idea is worth a serious referee if the authors can re-scope the claims, fix the equations, and release enough detail to reproduce the method. Send it to review with major-revision expectations.","headline":"A workable computational shortcut for shrinking ANFIS layer 3-to-4, but the 'fewer rules' claim conflates principal components with fuzzy rules and needs major rework before it supports interpretability.","tokens_in":27460,"tokens_out":2193,"would_cite":false,"duration_ms":22404,"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":"The paper claims that applying PCA to normalized firing strengths and selecting components with BPSO drastically reduces ANFIS rule counts and training time while preserving accuracy on low-dimensional data.","keywords":["ANFIS","rule reduction","Binary Particle Swarm Optimization","Principal Component Analysis","firing strengths","interpretable fuzzy rules","low-dimensional data","classification and regression"],"falsifier":"On a dataset with highly skewed or poorly initialized membership functions, run ANFIS-PCA-BPSO and then re-train the premise parameters of the surviving rules; if accuracy jumps substantially, the initial firing strengths were ranking rules by importance incorrectly. Alternatively, compare the selected rule subset against a randomly chosen subset of the same size: if random subsets match its accuracy across several datasets, the firing-strength-based selection is not doing the work.","tokens_in":26424,"feed_emoji":"📉","tokens_out":7666,"duration_ms":73606,"temperature":0.7,"pith_summary":"This paper proposes a rule-reduction method for ANFIS, a five-layer neuro-fuzzy network whose grid-partitioned rule base grows exponentially with input dimensionality. The authors insert a two-stage selector between the normalization layer and the defuzzification layer: PCA first decorrelates the normalized firing strengths of all generated rules, and Binary Particle Swarm Optimization then picks a small subset of the resulting components that minimizes prediction error. Their claim is that on low-dimensional classification and regression benchmarks this keeps accuracy within a few points of the full ANFIS while shrinking rule counts from dozens or hundreds to roughly two to four rules and cutting training time by orders of magnitude. The practical payoff is that a transparent, rule-based model becomes fast enough and simple enough for settings where interpretability matters and computing time is limited.","feed_headline":"Fuzzy AI rule explosion tamed: 81 rules down to 2","feed_subtitle":"Two-stage PCA and swarm selection prune ANFIS rules while keeping accuracy and slashing training time.","key_machinery":"The load-bearing mechanism is the two-stage PCA-BPSO selector interposed between layer 3 (normalized firing strengths $\\bar{w}_j$) and layer 4 (weighted consequents $\\bar{w}_j f_j$) of ANFIS. PCA converts the matrix of normalized firing strengths, $\\text{MAT}_{Nfs}$, into linearly uncorrelated principal components, and BPSO maintains a swarm of binary position vectors that act as a keys matrix, element-wise multiplying the components; the fitness function is the prediction error of the masked rule set on training data. The inertia weight and acceleration coefficients of BPSO are linearly updated each iteration so that exploration gives way to exploitation, and because the update loop touches only layers 3 and 4, the antecedent membership-function parameters stay fixed at their initialization throughout pruning.","core_discovery":"On the paper's own terms, the central discovery is that the normalized firing strengths of a fully generated ANFIS rule base contain enough information to decide which rules matter, and that this decision can be made by an embedded optimizer rather than by expert-set thresholds. The method forms a matrix of normalized firing strengths over samples and rules, applies PCA to obtain uncorrelated components (keeping those explaining about 95% of variance), and lets BPSO generate a binary keys matrix that selects components by evaluating the resulting rule set's error. Only the consequent parameters are updated during the forward pass, so the optimization loop runs entirely between layers 3 and 4 without backpropagation through the antecedent part. The reported outcome is a rule base reduced to between about 2 and 4 rules on most benchmarks, training times reduced from thousands of seconds to under a minute in most cases, and accuracy that matches or slightly exceeds the full ANFIS on several datasets, with only a small accuracy drop on a few others.","pith_inferences":["This reader's inference: because the PCA-BPSO step operates on normalized firing strengths rather than raw features, the same selector could be dropped into other rule-based systems that normalize rule activations, such as TSK fuzzy classifiers outside the ANFIS architecture.","This reader's inference: the fixed-premise assumption means the method's success depends on initial membership functions being reasonable; a natural follow-up is to re-tune the surviving rules' premise parameters after pruning and measure the accuracy gain.","This reader's inference: replacing the accuracy-only BPSO fitness with a multi-objective score that penalizes rule count could yield even smaller rule bases, since the current results already achieve near-baseline accuracy with very few rules."],"forward_implications":["Rule counts drop from 9-256 in the full ANFIS baseline to roughly 2-4 rules across the tested benchmarks, so the pruned model is far cheaper to store, inspect, and execute.","Training time drops from thousands of seconds to under a minute on most classification and regression benchmarks, making ANFIS practical for rapid retraining or real-time deployment.","Accuracy stays within a few points of the full ANFIS on most datasets and improves on Iris, Banana, Haberman, and Monk2, so the main cost of pruning is small and occasional.","On the ischemic-stroke application, the pruned model keeps statistically significant predictions (p-value 0.00007) with a small rule base, suggesting that the reduction does not destroy predictive signal in a real clinical task.","Compared with published rule-reduction baselines, the method achieves the fewest rules on 7 of 8 classification datasets and the best or near-best RMSE on 3 of 4 regression datasets, locating it on the efficient end of the accuracy-complexity frontier."],"supporting_citations":[{"why":"Supplies the binary particle swarm optimization mechanism that selects rule components.","marker":"[60]"},{"why":"Defines the five-layer ANFIS architecture and hybrid learning loop that the pruning stage is inserted into.","marker":"[11]"},{"why":"Provides the linearly decreasing inertia-weight strategy used to adapt BPSO parameters during optimization.","marker":"[68]"},{"why":"The authors' earlier ANFIS-BPSO model, which this work extends by adding PCA on firing strengths.","marker":"[85]"},{"why":"Previous PCA-plus-ANFIS application to the same ischemic-stroke dataset, used as the real-world comparison.","marker":"[87]"},{"why":"Classification baseline that reports accuracy and rule counts on the shared benchmarks.","marker":"[49]"},{"why":"Evolutionary fuzzy-rule baseline used for accuracy and rule-count comparison.","marker":"[38]"},{"why":"Fuzzy association-rule classification baseline used for accuracy and rule-count comparison.","marker":"[53]"},{"why":"Regression baseline (clustering-based extreme learning ANFIS) used for RMSE and rule-count comparison.","marker":"[43]"},{"why":"Regression baseline with fractional Tikhonov regularization used for RMSE and rule-count comparison.","marker":"[52]"}],"fun_headline_variants":["ANFIS rule count slashed to 2 via PCA and swarm selection","PCA + BPSO trim ANFIS to 2 rules, keep accuracy","From 81 rules to 2: PCA and swarm prune ANFIS","Swarm-optimized PCA shrinks ANFIS to 2 rules","ANFIS rule base cut from 81 to 2 with BPSO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the normalized firing strengths computed from the initial, un-tuned membership functions faithfully rank which rules matter, even though those membership functions are never updated after the BPSO pruning step.","fun_headline_variants_meta":{"raw":{"variants":["ANFIS rule count slashed to 2 via PCA and swarm selection","PCA + BPSO trim ANFIS to 2 rules, keep accuracy","From 81 rules to 2: PCA and swarm prune ANFIS","Swarm-optimized PCA shrinks ANFIS to 2 rules","ANFIS rule base cut from 81 to 2 with BPSO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2941,"prompt_tokens":982,"completion_tokens":1959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1870}},"tokens_in":598,"tokens_out":1959,"duration_ms":14275,"temperature":1.0,"reasoning_tokens":1870,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:18:55.938597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with highly skewed or poorly initialized membership functions, run ANFIS-PCA-BPSO and then re-train the premise parameters of the surviving rules; if accuracy jumps substantially, the initial firing strengths were ranking rules by importance incorrectly. Alternatively, compare the selected rule subset against a randomly chosen subset of the same size: if random subsets match its accuracy across several datasets, the firing-strength-based selection is not doing the work.","supporting_citations":[{"cited_title":"Kennedy, R","cited_arxiv_id":null,"evidence_quote":"Supplies the binary particle swarm optimization mechanism that selects rule components."},{"cited_title":"Haznedar, M","cited_arxiv_id":null,"evidence_quote":"Defines the five-layer ANFIS architecture and hybrid learning loop that the pruning stage is inserted into."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the linearly decreasing inertia-weight strategy used to adapt BPSO parameters during optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The authors' earlier ANFIS-BPSO model, which this work extends by adding PCA on firing strengths."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Previous PCA-plus-ANFIS application to the same ischemic-stroke dataset, used as the real-world comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classification baseline that reports accuracy and rule counts on the shared benchmarks."},{"cited_title":"Fazzolari, R","cited_arxiv_id":null,"evidence_quote":"Evolutionary fuzzy-rule baseline used for accuracy and rule-count comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fuzzy association-rule classification baseline used for accuracy and rule-count comparison."},{"cited_title":"Pramod, G","cited_arxiv_id":null,"evidence_quote":"Regression baseline (clustering-based extreme learning ANFIS) used for RMSE and rule-count comparison."},{"cited_title":"Tomasiello, W","cited_arxiv_id":null,"evidence_quote":"Regression baseline with fractional Tikhonov regularization used for RMSE and rule-count comparison."}],"review_version":1}