acceptodds
Under review as a conference paper at ICLR 2027

Z-Score Rule Induction: An Interpretable Lightweight Classifier

Abstract

Extracting interpretable rules to retrieve rare target instances from large unlabeled populations is a critical challenge when confirmed positive cases are extremely scarce. While decision trees offer interpretability, they overfit to the small set of positive examples and fail to generalize to unseen positive instances. To address this, we propose Z-Score Rule Induction (ZSRI), a lightweight algorithm specifically designed to induce simple, interpretable rules from tiny positive seeds in a single pass. ZSRI ranks features by their positive-class deviations using univariate -scores and induces decision boundaries via extremum- or quantile-based strategies to encompass target positive instances. Experimental evaluations on Wisconsin Breast Cancer, Pima Indians Diabetes, and synthetic benchmarks demonstrate that ZSRI achieves competitive predictive performance while executing up to faster than decision trees. Furthermore, on the Wisconsin Breast Cancer dataset, ZSRI trained with as few as positive seed instances recovers up to of unseen positive cases while maintaining a high precision of .

Primary area:
interpretability and explainable AI
Keywords:
Few-shot learning, Knowledge discovery, Positive-unlabeled learning, Rule induction, Rule-based classification

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ipUYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ZSRI proposed in this paper learns an interpretable conjunctive rule of kk rules, i.e. an axis-aligned hyperrectangle. ZSRI standardizes features, ranks them by the mean positive-class zz-score, keeps the top kk, and computes each threshold directly from the positive class using either of two proposed strategies (recall-oriented or precision-oriented). The kk conditions are then ANDed directly. An optional smoothing step converts rule violations into probabilities while keeping the same decision inside the rule region.

Experiments on 3 datasets show competitive F1 and faster training than decision trees; with 10 positive examples, ZSRI retrieves high recall at good precision.

Critical strengths
  • Simple, fast, interpretable. The zz-score filter plus direct positive-class thresholding gives a single-pass O(nm)O(nm) rule learner with a readable AND-of-thresholds output, avoiding recursive partitioning and iterative covering.
  • Strong few-shot retrieval. Under severe positive scarcity, ZSRI retains useful recall and precision while decision trees, OneR, and RIPPER collapse.
Critical weaknesses
  • Thin methodological novelty. The core is essentially zz-score feature filtering followed by positive-extremum/quantile thresholds. Rule induction is incredibly mature (RIPPER, CN2, PRIM, CORELS, Bayesian Rule Lists, ILP rules), and the paper does not adequately position ZSRI against PRIM or hyperbox classifiers.
  • Continuous features only. zz-scores, and some other components such as quantiles presume real-valued features. Categorical, ordinal, count, and mixed-type features are neither handled nor discussed.
  • Too few datasets in the experiments. Three datasets, one synthetic and linearly separable, cannot support general claims. Besides, they are quite small. There is no interaction-heavy, high-dimensional, multi-class, or modern ILP/PRIM comparison.
  • No ablation. The paper does not isolate the effects of zz-score ranking, the buffer α\alpha, the quantile qq, the smoothing method, or kk. It is unclear which components matter.
  • No code. No reproducibility artifact is released, weakening confidence in runtimes and few-shot curves.
Questions for authors

See weaknesses.

Reviewer_3x7gOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a simple and straightforward method to derive a decision rule as a conjunction of K comparison conditions on a subset of numerical variables. The variables are selected in order of largest absolute average z-score of the positive examples. The direction of the comparisons and thresholds are derived respectively from the signs of these z-scores and quantiles or extrema of these variables.

Critical strengths
  • Very simple and straightforward idea
  • Paper is very clearly written, with quality figures explaining the method
Critical weaknesses
  • The method is perhaps too simple and obvious. It almost seems like a formalization of what a domain expert would use to derive a rule by hand.
  • My main concern is applicability. This seems like it should only work on particularly simple (almost toy) data. The experiments on only two real world, tiny datasets do not inspire confidence that this is an algorithm with wide applicability. Even on these datasets, the proposed method does not outperform other simple methods like decision trees on the only reported non-gameable metric (F1).
  • Most of the analysis on few-shot performance looks only at recall in isolation. This is simply not a good choice of a single metric to compare different methods since a trivial solution can always achieve 100% recall at the cost of low precision/specificity. For example, lowering the threshold applied to a decision tree would be able to arbitrarily boost the recall. A metric like F1 or balanced accuracy should be used instead.
  • While the writing is good, I find the presentation overly formal for such a simple algorithm. Theorem 2, for example, stands out as immediately obvious and could be replaced by a single sentence.
  • The motivation behind the probability smoothing seems strange to me. First I would object calling this probability smoothing since it is simply a way to provide a continuous score that does not draw on any concrete probabilistic motivation. Furthermore, if the goal is to provide a continuous sense of how close a "negative example" is to the frontier, why not apply the same principle to positive examples instead of prescribing a score of 1 to everything that satisfies the rule. Having an unclamped score would also allow a user to apply whichever threshold they choose, thus relaxing or tightening the original rule using a single parameter.
Questions for authors

How does this method fare on a wider selection of datasets of different sizes?

Reviewer_zCAWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Z-Score Rule Induction (ZSRI), an efficient and interpretable algorithm for binary classification and positive instance retrieval. The main idea is to identify relevant features using the average z-scores of positive samples and then construct a conjunction of threshold-based rules. Two strategies are proposed to determine the thresholds, one based on extrema and one based on quantiles, to control the trade-off between precision and recall. The method is evaluated on three datasets, considering standard classification, few-shot positive retrieval, and computational efficiency.

Critical strengths
  • The paper is well written, clearly organized, and easy to follow. The proposed method is well explained.
  • The problem addressed is relevant, particularly in scenarios where positive examples are scarce and model interpretability is important.
  • I appreciate the simplicity of the proposed approach, which efficiently produces human-readable rules.
  • The experimental results are encouraging, particularly in the few-shot setting, where ZSRI achieves substantially higher recall than several of the considered baselines.
Critical weaknesses
  • I believe the experimental evaluation, although encouraging, is one of the weakest points. Only three datasets are considered, two of which are relatively small medical datasets. Moreover, I would like to see a broader comparison with existing methods, particularly in the few-shot setting. For instance, the authors compare ZSRI with standard Decision Trees and RIPPER, but these algorithms may not be particularly effective under extreme class imbalance without appropriate configurations. Including class-weighted or cost-sensitive variants could provide a fairer comparison.

  • I would also suggest considering approaches specifically designed for positive-unlabeled learning, as well as additional rule-induction algorithms. For example, FIND-RS [1], could be an interesting baseline, since it also learns interpretable rules by generalizing from positive examples.

  • The paper often frames the problem as positive-unlabeled learning or retrieval from unlabeled populations. However, in Section 5.1.1, the background dataset used during training consists of negative examples, while the remaining positive samples are withheld for evaluation. This differs from a standard PU-learning setting, where the unlabeled population may contain both positive and negative instances. I believe the authors should clarify this distinction and explain whether their method is intended for genuine PU learning or for supervised classification with very few positive examples.

  • For the few-shot experiments, the authors state that the results are averaged over 30 independent random seed draws. However, only mean values are reported, without standard deviations or confidence intervals. I would expect at least some indication of the variability of the reported performance, together with an appropriate statistical comparison between methods. A similar concern applies to the standard classification experiments, which appear to rely on a single train-test split.

  • I am not entirely convinced by the probabilistic interpretation of the smoothing mechanism presented in Section 4.2. The proposed estimator assigns a probability of 1 to every sample satisfying all the induced rules. However, satisfying the rules does not necessarily imply that a sample belongs to the positive class, as false positives may still occur. Moreover, using differences between standard normal CDF values to construct the penalties does no, justify interpreting the resulting scores as posterior probabilities. I think the authors should clarify this aspect and possibly refer to the output as a confidence score rather than a posterior probability estimate. If the intention is to provide actual probability estimates, an empirical evaluation of their calibration would be necessary.

  • The mathematical formulation is generally clear and correct. However, the presented theoretical results are rather straightforward and mostly formalize properties that directly follow from the definition of the algorithm.

Ref

[1] L. Bergamin, M. Polato, F. Aiolli. Improving rule-based classifiers by Bayes point aggregation. Neurocomputing, 613, 128699, 2025. https://doi.org/10.1016/j.neucom.2024.128699

Questions for authors

Q1. Could the authors clarify the precise learning setting considered in the few-shot experiments? In particular, why is the background population constructed using known negative examples, and how would the proposed approach behave if the unlabeled data also contained unknown positive instances?

Q2. Could the authors extend the experimental comparison by including class-balanced or cost-sensitive variants of the existing baselines, together with methods designed for PU learning? (see suggestion above)

Q3. Since the few-shot experiments are repeated 30 times, could the authors report standard deviations or confidence intervals and provide some statistical analysis of the differences between the methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.