Pith. sign in

REVIEW 3 major objections 4 minor 46 references

CLARK: Closed-loop Learning for Adaptive Reasoning over Knowledge Graphs

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CLARK claims that a closed loop of rule mining, probabilistic calibration, and graph refinement improves balanced accuracy by +3.36 points on two medical classification datasets.

desk verdict A plausible synthesis of KG extraction, LP^MLN weight learning, and symbolic rule mining, but the central generalization claim rests on a Table 3 with no described test split, so the empirical payoff is unverified. read the letter →

arxiv 2607.19996 v1 pith:V3UYLNJV submitted 2026-07-22 cs.AI

Yousef Khan , Luca Gherardini , Marco Maratea , Joel Arrais , Jose Sousa This is my paper
classification cs.AI
keywords knowledgegraphsprobabilisticanswersetprogrammingLP^MLNruleminingclassificationdistributionshiftmedicaldecisionsupportclosed-looplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CLARK proposes that classification over a knowledge graph improves when the graph itself is treated as a learnable object: translate the graph into a probabilistic logic program, mine candidate rules, learn their weights by maximum likelihood, prune unstable ones, and write accepted rules back as new edges. The authors demonstrate this on two medical datasets, where the best configuration (a structural learner combined with a knowledge-base rule miner) raises balanced accuracy by +3.36 points over the static CACTUS graph baseline. Their evidence also suggests that likelihood-based calibration aligns heterogeneous rule sources into a common scale, whereas frequency-based rule mining alone fails to improve accuracy. A sympathetic reader would care because this is a concrete path toward adaptive, interpretable medical classifiers that can update as new patient data arrives.

What carries the argument

The central engine is LP^MLN (Logic Programs with Markov Logic Networks), a probabilistic extension of Answer Set Programming that attaches log-linear weights to rules and defines a probability distribution over stable models. CLARK's loop works through an explicit mapping from a CACTUS knowledge graph to a weighted LP^MLN base program: nodes become soft prior atoms, edges become soft Horn rules, and cardinality constraints enforce exactly one state per attribute. A Structural Learner then proposes candidate rules constrained to connected components of the graph, scores them by held-out log-likelihood gain, and feeds them into maximum-likelihood weight learning; bootstrap pruning removes uns

What would settle it

Run CLARK with a strict three-way split: training worlds for weight learning, a model-selection set for rule scoring and bootstrap pruning, and a final test set that is never touched by Algorithm 1; recompute Table 3's balanced-accuracy deltas on that final test set. If the +3.36 combined gain or the individual positive deltas do not reproduce on untouched patient worlds, the central claim of more generalisable inference is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a knowledge graph becomes a better substrate for classification when its structure is refined in a closed loop: translate the CACTUS graph into an LP^MLN program, mine candidate rules, learn their weights by maximum likelihood on training worlds, prune unstable rules by bootstrap resampling, and write accepted rules back as new edges. In the authors' experiments on the Cleveland Heart Disease and Wisconsin Breast Cancer datasets, this loop takes the CACTUS-only baseline and improves balanced accuracy by +3.36 points in the best configuration (Structural Learner + AMIE3), with per-setting gains of +1.22 and +0.55 on HDC and +1.19 and +0.40 on WDBC. The author

Load-bearing premise

The accuracy gains assume the balanced-accuracy numbers were computed on patient worlds not used for learning rule weights or selecting rules; the paper does not explicitly describe any test split, resampling, or cross-validation for the downstream classification evaluation.

Editorial extensions

If this is right

  • On the authors' measurements, pairing the structural learner with AMIE3 produces the largest gain (+3.36 balanced-accuracy points), larger than either learner alone, which they attribute to a consensus effect.
  • LP^MLN weight learning calibrates rules from different miners onto a common numerical scale, so the final model is less sensitive to which miner proposed a rule.
  • Adding probabilistic reasoning to the original graph without new rules can hurt accuracy (e.g., -4.81 on HDC under degree-based inference), indicating that calibration alone is not enough; structural enrichment is needed.
  • Rules selected by frequency heuristics alone (CMAR) fail to improve and can degrade performance, supporting the paper's claim that likelihood-aligned rule selection is the active ingredient.
  • Because new patient records enter as additional worlds and trigger weight relearning, CLARK is framed as a decision-support system that adapts its knowledge graph over time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the result holds on a proper held-out test set, it would imply that a model can track distribution shift without full retraining, since new patient worlds can enter as evidence and trigger weight updates.
  • The consensus gain from combining SL and AMIE3 suggests a testable extension: generating diverse rule sets from complementary learners and merging them by likelihood-weighted agreement may produce further superadditive gains; the paper does not test this directly.
  • The paper uses graph centrality as a classification readout; replacing it with direct marginal probabilities from the learned LP^MLN program is a natural next test of whether the probabilistic semantics, not just the graph topology, carry the gain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces CLARK, a closed-loop framework that starts from a CACTUS-derived knowledge graph, translates it into an LP^MLN program, iteratively enriches it with rules from symbolic learners (SL, CMAR, AMIE3), learns rule weights by maximum likelihood, prunes unstable rules via bootstrap, and writes accepted rules back into the KG. The central empirical claim is that this closed-loop refinement improves downstream classification balanced accuracy relative to the CACTUS-only baseline, with the best configuration (SL+AMIE3+LP^MLN) reporting a combined +3.36 percentage-point gain in Table 3. The paper also compares rule learners on rule-quality metrics (weight, support, confidence, Δ-LL) using non-parametric tests.

Significance. If the central claim were established, CLARK would be a meaningful step toward interpretable, knowledge-driven classification that adapts its graph structure through probabilistic rule learning. The combination of CACTUS-derived structure, LP^MLN weight learning, and iterative enrichment is well motivated and the experimental design isolates the contribution of rule learners. The paper contributes a concrete algorithm (Algorithm 1), a principled translation from weighted KGs to LP^MLN programs, and a 'Structural Learner' that scores rules by held-out log-likelihood. However, the empirical evidence as presented does not yet support the advertised generalization claim because the evaluation protocol leaves undetermined whether the final balanced-accuracy numbers are computed on data not used at any stage of the pipeline. The rule-quality analysis is technically sound but does not compensate for the missing test-split guardrail.

major comments (3)
  1. [§5.4, Table 3, Algorithm 1] The central claim of improved and generalisable inference rests entirely on Table 3, yet the manuscript never specifies the data partition used to compute the balanced-accuracy numbers. Algorithm 1 splits W into W_train and W_held (line 1), but W_held is used inside the loop only for scoring candidate rules via Δ(r) (line 5); there is no separate test split, cross-validation, or repeated resampling described in §5.4. The text says predictions are 'derived from the enriched KG' without stating whether they come from W_train, W_held, or the full W. If the numbers are computed on W_train, the +3.36 ΔBA is an in-sample fit; if on W_held, that set has already influenced rule selection; if on the full W, then both the graph and the weights have seen all records. In all scenarios the reported improvement does not demonstrate generalisation as claimed.
  2. [§4.1 and §5.1: Construction of G^(0)] Section 4.1 states that G^(0) is constructed from the full dataset D ("Given D ... and G^(0) = (V,E)"), and §5.1 explains that CACTUS transforms the datasets into the structured representation. This means the initial KG already encodes information from all records before any split is performed. Even if a final test partition were introduced, the base graph would carry information from that test set, so the evaluation would still not be clean. A proper evaluation must build G^(0) from training data only (or use a nested CV scheme that includes graph construction), and report performance on a fully held-out partition.
  3. [§5.4, Table 3: Statistical reliability] Table 3 reports balanced-accuracy changes of +0.04 to +1.79 percentage points, but no confidence intervals, repeated runs, or significance tests are given. With 303 and 569 patient worlds, differences of a few tenths of a percentage point on a balanced-accuracy metric are within sampling noise. The claim of a consistent +3.36 gain cannot be interpreted without a measure of variability. Given that the rule-quality comparisons in Figures 2-3 do report Mann-Whitney U tests with Holm-Bonferroni correction, the absence of any inferential test for the primary Table 3 results is a notable omission that should be addressed.
minor comments (4)
  1. [§5.4, Table 3 heading] The table formatting is inconsistent: the column header "HDC PageRank Degree WDBC PageRank Degree ΔBA" is ambiguous, and the table body mixes values like "+0.04" and "+0.0" without a clear legend for the baseline. Please restructure the table to show each dataset and inference strategy separately, with a clear indication of which entries are deltas relative to G*.
  2. [§4.2, line 5 of Algorithm 1] Algorithm 1 passes W_held into the candidate generation step L(G, W_train, W_held), but §4.2 describes SL's held-out log-likelihood gain Δ(r) as the criterion. This is a design choice worth explaining: using the held-out set for rule selection is legitimate only if a second, untouched test set is reserved for final evaluation. The paper should state this explicitly and implement it.
  3. [Supplemental Material Statement] The statement that source code cannot be made available due to commercialisation purposes limits reproducibility. Given that the primary empirical claim is a small performance improvement, making at least the evaluation splits and centrality-decoding code available would substantially strengthen the paper.
  4. [§5.4, last paragraph] The sentence "This suggests that agreement between the learners on the relevance of certain rules can amplify their impact" is speculative. No evidence of agreement is shown; the plus-3.36 result could also arise from the larger rule set, overlapping rules, or noise. Please either provide direct evidence for consensus effects or soften the interpretation.

Circularity Check

1 steps flagged · score 7.0 of 10

Central 'improved classification/generalisability' claim rests on Table 3, but the paper never defines a test split distinct from W_train/W_held; reported balanced-accuracy gains appear to be in-sample fits.

  1. fitted input called prediction [Algorithm 1 (lines 1, 5, 8); Section 5.4 (Table 3)]
    "Split W into W_train and W_held. ... C ← L(G, W_train, W_held). ... Run LPMLN weight learning on (Π, W_train) to obtain Π̂. ... In this setting, each patient corresponds to a world, and predictions are derived from the enriched KG using centrality-based scoring."

    The only sample split is W_train/W_held; W_held is consumed in line 5 to score candidate rules via Δ(r), and all LP^MLN weights are fit on W_train. Section 5.4 then reports balanced accuracy for predictions 'derived from the enriched KG' without introducing any separate test partition. Thus the Table 3 numbers, and the headline ΔBA of +3.36, are computed on worlds that either contributed to the fitted graph/weights or influenced rule selection. The advertised 'improved classification performance and more generalisable inference' is therefore a restatement of in-sample fit, not an out-of-sample prediction. By the paper's own procedure, the reported predictions are functions of the fitted values.

full rationale

CLARK's empirical contribution is the claim that enriching a CACTUS-derived KG with LP^MLN-calibrated rules improves downstream classification and generalization. The evidence is Table 3's balanced-accuracy deltas. Tracing the pipeline in Algorithm 1, the only data split is W_train/W_held; W_held is used directly in candidate-rule selection (line 5: 'C ← L(G, W_train, W_held)'), and weight learning is explicitly on W_train (line 8). Section 5.4 describes predictions as 'derived from the enriched KG' but never states whether the reported accuracies come from W_train, W_held, or the full W. Since the enriched KG and its edge weights are outputs of fitting to these same worlds, the 'predictions' are not independent of the fitted parameters. Moreover, Section 5.1 indicates the CACTUS graph is built from the full dataset before any split ('Using CACTUS, the datasets are transformed into a structured representation'), so even the base graph already encodes information from all records. Without a defined test split, the +3.36 ΔBA cannot support the generalization claim; it is, by construction, an in-sample or leakage-tainted evaluation. This is the load-bearing circularity: the central empirical result reduces to a fit on the data it claims to predict. The paper's use of external LP^MLN and CACTUS references is not itself circular; the issue is the missing out-of-sample guardrail in the evaluation chain.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework rests on fitted quantities at every stage: CACTUS derives node and edge scores from the same data, LP^MLN learns rule weights by maximum likelihood, and several selection thresholds are unspecified. No new theoretical entities are introduced; the main assumptions are that the CACTUS abstraction is faithful, that one-state-per-attribute constraints are valid, and that graph centrality is a sound decoding of the probabilistic program.

free parameters (7)
  • CACTUS node prior scores p_v = per-node, estimated by CACTUS on the same dataset
    Converted to log-linear weight w_v = log(p_v/(1-p_v)) in §4.1; estimated from the data used for evaluation.
  • CACTUS edge association strengths p_{v->u} = per-edge, estimated by CACTUS
    Transformed to log-linear weights w_{v->u} for edge-induced rules in §4.1; fitted to the same data.
  • LP^MLN rule weights \hat{w}_k = learned by maximum likelihood on W_train for all base and candidate rules
    Algorithm 1 lines 7-8; these are the fitted parameters of the final probabilistic program.
  • SL held-out log-likelihood threshold = unspecified; 'positive (optionally regularised) gain'
    Hand-selected rule inclusion criterion in §4.2; no numeric value or regularisation strength given.
  • Bootstrap stability threshold = unspecified; 'consistently nonzero and of the same sign'
    Algorithm 1 line 9; no number of replicates or acceptance threshold stated.
  • Top-K candidate rule selection = unspecified
    Algorithm 1 line 6 selects a subset \tilde{C} of candidates; K is not specified.
  • Centrality-based decoding of class predictions = unspecified
    §5.4 derives predictions via Degree/PageRank over the enriched KG; the exact mapping from scores to class labels is not specified.
assumptions (5)
  • standard math LP^MLN semantics with log-linear weights over stable models is a valid probabilistic model
    Section 3.2 Definition 1; the entire pipeline relies on LP^MLN inference and weight learning.
  • domain assumption CACTUS-derived directed edges represent conditional dependencies that are meaningful for classification
    Sections 3.3 and 4.1; if the KG abstraction is not faithful, all downstream rules inherit its bias.
  • domain assumption Each attribute takes exactly one state in every stable model (cardinality constraint)
    §4.1 enforces 1{val(a,sigma):sigma in S(a)}1; this totality/mutual-exclusivity assumption may fail for missing data.
  • domain assumption Rules restricted to a single connected component of the KG capture relevant structure
    §4.2 Structural Learner; the connectivity bias is a heuristic restriction of the hypothesis space.
  • ad hoc to paper Graph centrality (degree/PageRank) over the enriched KG is a valid predictor of the class
    §5.4 uses centrality-based scoring without derivation or comparison to other decoding schemes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLARK: Closed-loop Learning for Adaptive Reasoning over Knowledge Graphs." pith.science (2026). https://pith.science/paper/V3UYLNJV

@misc{pith2026260719996,
  author       = {Pith},
  title        = {Pith review of: CLARK: Closed-loop Learning for Adaptive Reasoning over Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3UYLNJV}},
  note         = {Machine review of arXiv:2607.19996}
}
abstract

Machine Learning models are widely used for automating classification tasks by extracting statistical patterns from data. However, their performance deteriorates if the data distribution changes, making them ill-suited to handle uncertain and evolving information. Moreover, they provide limited support for integrating prior knowledge. To address these limitations, we present CLARK (Closed-loop Learning for Adaptive Reasoning over Knowledge Graphs), a framework that integrates knowledge graphs, symbolic rule mining, and probabilistic reasoning under the Logic Programs with Markov Logic Networks (LP$^{\text{MLN}}$) formalism. Starting from CACTUS-derived KGs, CLARK translates graph structure into an LP$^{\text{MLN}}$ program and iteratively enriches it with candidate rules proposed by symbolic learners. These rules are calibrated through probabilistic weight learning, enabling reasoning under uncertainty and refinement of the underlying graph structure. We evaluate CLARK on two medical datasets, analysing both rule quality and downstream classification performance. Results demonstrate that CLARK leads to improved classification performance and more generalisable inference. Overall, CLARK provides a principled approach to constructing adaptive, interpretable, knowledge-driven models for classification.

Figures

Figures reproduced from arXiv: 2607.19996 by the authors.

Figure 1
Figure 1. CLARK pipeline. (1) Inputs: abstracted tabular data (e.g., feature Radius with value High) and an initial KG G (0) . (2) Symbolic construction: (2.1) KG G (0) is translated into an LPMLN base program (e.g., val(radius, high)); (2.2) candidate rules are mined and (2.3) the base program augmented with candidate rules undergoes weight learning via MC-SAT, assigning log-linear weights. (3) Outputs: a weighted LPMLN prog… view at source ↗
Figure 2
Figure 2. Distribution of rule-quality metrics (weight, support, confidence, [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Distribution of rule-quality metrics (weight, support, confidence, [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 3 canonical work pages

  1. [1]

    https://github.com/azreasoners/lpmln (2020), commit 4ccf89d, accessed 2025-12-03

    azreasoners: Lpmln. https://github.com/azreasoners/lpmln (2020), commit 4ccf89d, accessed 2025-12-03

    2020

  2. [2]

    Azzolini, D., Mazzotta, G., Ricca, F., Riguzzi, F.: A Novel Framework for Reason- ing over Optimization Problems in Probabilistic Answer Set Programming. In: Pro- ceedings of the 22nd International Conference on Principles of Knowledge Repre- sentationandReasoning.pp.67–77(102025).https://doi.org/10.24963/kr.2025/7, https://doi.org/10.24963/kr.2025/7

    work page doi:10.24963/kr.2025/7 2025

  3. [3]

    Journal of Machine Learning Research18(109), 1–67 (2017)

    Bach, S.H., Broecheler, M., Huang, B., Getoor, L.: Hinge-loss markov random fields and probabilistic soft logic. Journal of Machine Learning Research18(109), 1–67 (2017)

    2017

  4. [4]

    Bach, S.H., Broecheler, M., Huang, B., Getoor, L.: Hinge-loss markov random fields and probabilistic soft logic. J. Mach. Learn. Res.18(1), 3846–3912 (Jan 2017)

    2017

  5. [5]

    Cambridge University Press (2003)

    Baral, C.: Knowledge Representation, Reasoning and Declarative Problem Solving. Cambridge University Press (2003)

    2003

  6. [6]

    International Journal of Approximate Reasoning125, 218–239 (2020)

    Cozman, F.G., Mauá, D.D.: The joy of probabilistic answer set programming: Se- mantics, complexity, expressivity, inference. International Journal of Approximate Reasoning125, 218–239 (2020)

    2020

  7. [7]

    Transactions on Graph Data and Knowledge1(1), 8:1–8:35 (2023)

    d’Amato, C., Mahon, L., Monnin, P., Stamou, G.: Machine Learning and Knowl- edge Graphs: Existing Gaps and Future Research Challenges. Transactions on Graph Data and Knowledge1(1), 8:1–8:35 (2023). https://doi.org/10.4230/ TGDK.1.1.8, https://drops.dagstuhl.de/entities/document/10.4230/TGDK.1.1.8

    work page doi:10.4230/tgdk.1.1.8 2023

  8. [8]

    https://github.com/dig-team/ amie (2026), gitHub repository, accessed January 30, 2026

    dig-team: Amie: Rule mining in knowledge graphs. https://github.com/dig-team/ amie (2026), gitHub repository, accessed January 30, 2026

    2026

Show all 46 references
  1. [9]

    In: International Symposium on Practical Aspects of Declarative Languages

    Dundua, B., Kutsia, T., Reisenberger-Hagmayer, K.: An overview of p log. In: International Symposium on Practical Aspects of Declarative Languages. pp. 34–

  2. [10]

    The VLDB Journal24(6), 707–730 (2015)

    Galárraga, L., Teflioudi, C., Hose, K., Suchanek, F.M.: Fast rule mining in onto- logical knowledge bases with amie+. The VLDB Journal24(6), 707–730 (2015). https://doi.org/10.1007/s00778-015-0394-1

    2015 doi

  3. [11]

    Theory and Practice of Logic Programming19(1), 27–82 (2019)

    Gebser, M., Kaminski, R., Kaufmann, B., Schaub, T.: Multi-shot asp solving with clingo. Theory and Practice of Logic Programming19(1), 27–82 (2019)

    2019

  4. [12]

    Nature Machine In- telligence2(11), 665–673 (2020)

    Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine In- telligence2(11), 665–673 (2020)

    2020

  5. [13]

    CoRR abs/2004.07780(2020), https://arxiv.org/abs/2004.07780

    Geirhos, R., Jacobsen, J., Michaelis, C., Zemel, R.S., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. CoRR abs/2004.07780(2020), https://arxiv.org/abs/2004.07780

    2004 arXiv

  6. [14]

    In: Kowalski, R., Bowen, Kenneth (eds.) Proceedings of International Logic Pro- gramming Conference and Symposium

    Gelfond, M., Lifschitz, V.: The stable model semantics for logic programming. In: Kowalski, R., Bowen, Kenneth (eds.) Proceedings of International Logic Pro- gramming Conference and Symposium. pp. 1070–1080. MIT Press (1988), http: //www.cs.utexas.edu/users/ai-lab?gel88

    1988

  7. [15]

    MIT press (2007) A Probabilistic Framework for Knowledge-Driven Classification 17

    Getoor, L., Taskar, B.: Introduction to statistical relational learning. MIT press (2007) A Probabilistic Framework for Knowledge-Driven Classification 17

    2007

  8. [16]

    Gherardini, L., Lengyel, I., Peto, T., Klaverd, C.C.W., Meester-Smoord, M.A., Colijnd, J.M., Consortium, E.R., Consortium, E., Sousa, J.: Cactus as a reliable tool for early classification of age-related macular degeneration (2025), https:// arxiv.org/abs/2506.14843

    2025 arXiv

  9. [17]

    In: Bellazzi, R., Juarez Herrero, J.M., Sacchi, L., Zupan, B

    Gherardini, L., Tworek, P., Szczypka, M., Khan, Y., Mikołajczyk, M., Lewandowski, R., Sousa, J.: Explainable ai for allergy diagnosis: A cactus frame- work to address doctor variability. In: Bellazzi, R., Juarez Herrero, J.M., Sacchi, L., Zupan, B. (eds.) Artificial Intelligen...

    2025

  10. [18]

    ACM Trans- actions on Intelligent Systems and Technology15(3), 1–23 (2024)

    Gherardini, L., Varma, V.R., Capała, K., Woods, R., Sousa, J.: Cactus: a compre- hensive abstraction and classification tool for uncovering structures. ACM Trans- actions on Intelligent Systems and Technology15(3), 1–23 (2024)

    2024

  11. [19]

    BMC Medical Research Methodology18(1), 81 (2018)

    Giacalone, M., Agata, Z., Cozzucoli, P.C., Alibrandi, A.: Bonferroni-holm and permutation tests to compare health data: methodological and applicative issues. BMC Medical Research Methodology18(1), 81 (2018). https://doi.org/10.1186/ s12874-018-0540-8, https://doi.org/10.1186/...

    2018 doi

  12. [20]

    In: International Joint Conference on Rules and Reasoning

    Hahn, S., Janhunen, T., Kaminski, R., Romero, J., Rühling, N., Schaub, T.: Plingo: a system for probabilistic reasoning in clingo based on lp mln. In: International Joint Conference on Rules and Reasoning. pp. 54–62. Springer (2022)

    2022

  13. [21]

    Theory and Prac- tice of Logic Programming25(2), 134–167 (2025)

    Hahn, S., Janhunen, T., Kaminski, R., Romero, J., Rühling, N., Schaub, T.: Plingo: A system for probabilistic reasoning in answer set programming. Theory and Prac- tice of Logic Programming25(2), 134–167 (2025)

    2025

  14. [22]

    ACM Computing Surveys (2021)

    Hogan, A., Blomqvist, E., Cochez, M., d’Amato, C., Melo, G.D., Gutierrez, C., Kirrane, S., Labra Gayo, J.E., Navigli, R., Neumaier, S., Ngomo, A.C.N., Polleres, A., Rashid, S.M., Rula, A., Schmelzeisen, L., Sequeda, J., Staab, S., Zimmermann, A.: Knowledge graphs. ACM Computin...

    2021 doi

  15. [23]

    Scandinavian Journal of Statistics6(2), 65–70 (1979)

    Holm, S.: A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics6(2), 65–70 (1979)

    1979

  16. [24]

    UCI Machine Learning Repository (1989), DOI: https://doi.org/10.24432/C52P4X

    Janosi Andras, Steinbrunn William, P.M., Robert, D.: Heart disease. UCI Machine Learning Repository (1989), DOI: https://doi.org/10.24432/C52P4X

    1989 doi

  17. [25]

    In: Proceedings of the NIPS workshop on probabilistic programming: foundations and applications

    Kimmig, A., Bach, S., Broecheler, M., Huang, B., Getoor, L., Mansinghka, V., Roy, D., Goodman, N.: A short introduction to probabilistic soft logic. In: Proceedings of the NIPS workshop on probabilistic programming: foundations and applications. pp. 1–4 (2012)

    2012

  18. [26]

    Nature communications10(1), 1096 (2019)

    Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., Müller, K.R.: Unmasking clever hans predictors and assessing what machines really learn. Nature communications10(1), 1096 (2019)

    2019

  19. [27]

    In: ICLP (Technical Communications) (2015)

    Lee, J., Meng, Y., Wang, Y.: Markov logic style weighted rules under the stable model semantics. In: ICLP (Technical Communications) (2015)

    2015

  20. [28]

    In: Proceed- ings of the Fifteenth International Conference on Principles of Knowledge Repre- sentation and Reasoning

    Lee, J., Wang, Y.: Weighted rules under the stable model semantics. In: Proceed- ings of the Fifteenth International Conference on Principles of Knowledge Repre- sentation and Reasoning. p. 145–154. KR’16, AAAI Press (2016)

    2016

  21. [29]

    Lee, J., Wang, Y.: Weight learning in a probabilistic extension of answer set pro- grams (2018), https://arxiv.org/abs/1808.04527

    2018 arXiv

  22. [30]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Lee, J., Yang, Z.: Lpmln, weak constraints, and p-log. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)

    2017

  23. [31]

    In: Proceedings 2001 IEEE International Conference on Data Mining

    Li, W., Han, J., Pei, J.: Cmar: accurate and efficient classification based on multiple class-association rules. In: Proceedings 2001 IEEE International Conference on Data Mining. pp. 369–376 (2001). https://doi.org/10.1109/ICDM.2001.989541 18 Y. Khan et al

    2001

  24. [32]

    Lifschitz, V.: Answer set programming, vol. 3. Springer Cham (2019)

    2019

  25. [33]

    The Annals of Mathematical Statistics18(1), 50–60 (1947)

    Mann, H.B., Whitney, D.R.: On a test of whether one of two random variables is stochastically larger than the other. The Annals of Mathematical Statistics18(1), 50–60 (1947). https://doi.org/10.1214/aoms/1177730491

    1947

  26. [34]

    Proceedings of the IEEE104(1), 11–33 (2015)

    Nickel, M., Murphy, K., Tresp, V., Gabrilovich, E.: A review of relational machine learning for knowledge graphs. Proceedings of the IEEE104(1), 11–33 (2015)

    2015

  27. [35]

    In: European Conference on Logics in Artificial Intelligence

    Nickles,M.:Atoolforprobabilisticreasoningbasedonlogicprogrammingandfirst- order theories under stable model semantics. In: European Conference on Logics in Artificial Intelligence. pp. 369–384. Springer (2016)

    2016

  28. [36]

    Machine learning62(1), 107–136 (2006)

    Richardson, M., Domingos, P.: Markov logic networks. Machine learning62(1), 107–136 (2006)

    2006

  29. [37]

    Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019), https://arxiv.org/abs/1811. 10154

    2019

  30. [38]

    Journal of Web Semantics79, 100806 (2023)

    Schramm, S., Wehner, C., Schmid, U.: Comprehensible artificial intelligence on knowledge graphs: A survey. Journal of Web Semantics79, 100806 (2023). https://doi.org/https://doi.org/10.1016/j.websem.2023.100806, https:// www.sciencedirect.com/science/article/pii/S1570826823000355

    2023

  31. [39]

    Biometrika52(3-4), 591–611 (1965)

    Shapiro, S.S., Wilk, M.B.: An analysis of variance test for normality (complete samples). Biometrika52(3-4), 591–611 (1965)

    1965

  32. [40]

    Shimodaira, H., Sugiyama, M., Storkey, A., Gretton, A., David, S.B., Quinone- roCandela, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.: Dataset Shift in Machine Learning, pp. 201–205. Neural Information Processing Series, Yale Uni- versity Press (2009)

    2009

  33. [41]

    In: Proceed- ings of the AAAI Conference on Artificial Intelligence

    Wang, Y., Lee, J.: Handling uncertainty in answer set programming. In: Proceed- ings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015)

    2015

  34. [42]

    UCI Machine Learning Repository (1993)

    William, W., Olvi, M., Nick, S., W., S.: Breast cancer wisconsin (diagnostic). UCI Machine Learning Repository (1993). https://doi.org/10.24432/C5DW2B

    1993 doi

  35. [43]

    Educa- tional and Psychological Measurement36(2), 297–300 (1976)

    Willson, V.L.: Critical values of the rank-biserial correlation coefficient. Educa- tional and Psychological Measurement36(2), 297–300 (1976). https://doi.org/10. 1177/001316447603600207

    1976

  36. [44]

    Transactions on Graph Data and Knowledge1(1), 7:1–7:23 (Dec 2023)

    Wu, H., Wang, Z., Wang, K., Omran, P.G., Li, J.: Rule learning over knowledge graphs: A review. Transactions on Graph Data and Knowledge1(1), 7:1–7:23 (Dec 2023). https://doi.org/10.4230/TGDK.1.1.7, https://d-nb.info/1367508878/34

    2023

  37. [45]

    Ye, W., Jiang, L., Xie, E., Zheng, G., Ma, Y., Cao, X., Guo, D., Qi, D., He, Z., Tian, Y., Coffee, M., Zeng, Z., Li, S., Ting-hao, Huang, Wang, Z., Rehg, J.M., Kautz, H., Zhang, A.: The clever hans mirage: A comprehensive survey on spurious correlations in machine learning (20...

    2025

  38. [46]

    Proceedings of the ACM on Human-Computer Interaction 3(CSCW), 1–23 (Nov 2019)

    Zhang, Z., Singh, J., Gadiraju, U., Anand, A.: Dissonance between human and machine understanding. Proceedings of the ACM on Human-Computer Interaction 3(CSCW), 1–23 (Nov 2019). https://doi.org/10.1145/3359158, http://dx.doi.org/ 10.1145/3359158

    2019 doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.