h t t p s : / / w w w . c s . u t e x a s . e d u / u s e r s / m l / p a p e r s / a b d u c e - i n d u c e - i j c a i w - 9 7 . p d f

Ingest: Integrating Abduction and Induction in Machine Learning

Type: kb/sources/types/ingest-report.md

Classification

A 1997 IJCAI workshop paper reviewing a research program, with logical definitions, a worked repair example, and summaries of previously published experiments. Author Raymond J. Mooney, a University of Texas computer-science researcher and coauthor of several reviewed systems, provides direct knowledge of their mechanisms. The paper is an overview by a participant rather than an independent comparative evaluation.

Summary

Mooney distinguishes two integrations: abduction can locate faults and suggest repairs to an imperfect theory, and induction can learn a knowledge base used for abductive diagnosis. In theory refinement, assumptions that make an unprovable positive case derivable identify candidate repair points; coverage across positive cases and exclusion of negative cases then guide rule deletion or induction of replacement rules. The target is a consistent theory with minimal syntactic change, approximated heuristically because exact optimization is intractable. In the second integration, LAB adds disorder-to-symptom rules to improve minimum-cover diagnostic accuracy on training cases. The paper reports successful applications of the broader program in biology, medicine, and tutoring, but refers readers elsewhere for detailed experiments. It is useful for understanding the division of labor between explanation construction and general rule learning, including how constraints on abduction affect which repairs are considered.

Quotes

No source quotes have been retained yet.

Connections Found

The source supplies a concrete technical comparison for candidacy evidence versus acceptance evidence: an assumption that explains one failed case nominates a repair, while checks across positive and negative cases discriminate revisions. This illustrates a separation of roles without establishing the note's general prohibition. It also provides a useful boundary for reach-assessment: the paper evaluates consistency, simplicity, and predictive accuracy, but does not assess whether a proposed mechanism supports its claimed scope beyond the tested cases.

Extractable Value

  1. Separate a case explanation from a retained rule change. An abductive assumption supplies what would make a particular observation provable; induction constructs a general rule from cases. The proposal is not itself an accepted revision, and an accepted revision under the paper's task criteria is not thereby an established causal explanation. This gives the candidacy note a formal comparison beyond its current workflow cases. [quick-win]
  2. Negative cases discriminate equally simple explanations. In the worked example, both Q(X) :- V(X) and Q(X) :- W(X) cover the positive cases, but the latter also covers a negative case. Deleting the antecedent Q(X) is another proposed repair that fails negative-case checks. Simplicity and positive coverage therefore leave choices unresolved that counterexamples can settle. [just-a-reference]
  3. Extend evaluation at the revision-selection stage. Training consistency, minimal syntactic change, and independent-test accuracy answer different questions. A Commonplace extension could add mechanism-discriminating interventions or structured shifts, with proof obligations where the domain is formalized, before granting a revision a wider scope. These would implement reach-assessment as defined in the KB; they are proposed additions, not criteria demonstrated by this paper. [experiment]
  4. Expose the assumptions that restrict repair search. Single-fault versus multiple-assumption abduction and preferences for more specific assumptions affect the repair points available or preferred. A revision process can miss a needed change before its evaluator sees any candidate. The paper makes these search restrictions concrete without establishing that one is generally superior. [deep-dive]
  5. Keep the two integrations distinct. LAB learns rules whose downstream use is abductive minimum-cover diagnosis; this differs from using abduction to guide repairs of a deductively evaluated theory. Its example supports evaluating learned knowledge through its intended inference procedure rather than equating all rule learning with deductive classification. [just-a-reference]

Limitations (our opinion)

The empirical claims summarize earlier publications without quantitative result tables, detailed splits, or controlled comparisons in this paper. LAB's claimed advantage is explicitly tied to one evaluation metric, whose definition is not supplied here. The proposed Bayesian-network refinement work is ongoing, so it cannot serve as demonstrated evidence for that extension. Predictive gains could reflect useful initial knowledge and inductive bias without establishing that the learned rules capture the right explanatory mechanism.

The effective update space is also bounded. Logical refinement receives encoded case descriptions, labels, and an initial theory; it can compose clause and literal additions or removals into revised logical mappings. The overview does not show these systems learning new observation representations, repair operators, or abduction policies. LAB receives symptom sets with correct disorder sets and adds individual propositional disorder-to-symptom rules; its minimum-cover diagnostic procedure and greedy addition strategy remain fixed. As learning inside a fixed decomposition cautions, improvement within these choices does not validate the choices themselves or show that a required correction outside them is reachable.

The capture contains the full paper, but PDF extraction loses mathematical glyphs and interleaves columns in the worked example. Exact formal transcription requires checking the primary PDF. The readable prose supports the role distinctions above; the capture should not be used as a clean formula source.

Review Candidacy evidence licenses escalation to assessment, not acceptance against the paper's worked repair example to decide whether it warrants a third comparison case, explicitly separating proposal, training-case selection, and any additional assessment of explanatory reach.