DRKE: Domain-Rule-based Knowledge Enhancement

• DRKE is a family of methodologies that leverages domain-specific rules, heuristics, and constraints to enhance learning beyond purely data-driven approaches.
• It operationalizes techniques like search-space pruning, virtual sample generation, and soft constraint integration to improve model performance.
• DRKE has proven effective in boosting efficiency, transferability, and reliability across tasks such as knowledge graph completion and LLM augmentation.

Domain-Rule-based Knowledge Enhancement (DRKE) denotes, across the cited literature, the use of domain knowledge expressed as rules, heuristics, constraints, memberships, or other rule-like structures to improve learning and reasoning beyond purely data-driven inference. The enhancement may occur by restricting a search space, refining a training set, adding soft probabilistic constraints, penalizing predictions that violate learned domain structure, eliciting executable rules from experts, routing anonymized tasks to domain-specialized agents, or injecting domain knowledge into prompts and differentiable logic modules (Singh et al., 2010, Iqbal, 2011, Jiang et al., 2015, Wang et al., 2019, Nikitin et al., 2021, Yang et al., 12 Sep 2025, Srinivas et al., 30 Apr 2026). In the 2010–2026 works surveyed here, DRKE is therefore not a single algorithm but a family of mechanisms for making models more efficient, more transferable under shift, more consistent under editing, or more reliable in specialized domains.

1. Conceptual scope and lineage

An early operational form of DRKE appears in Vikram Singh and Sapna Nagpal’s Interactive Association Rule algorithm, where user preference over attribute(s) of interest defines a reduced working database $D'$ and thereby narrows the mining scope before frequent-itemset discovery begins (Singh et al., 2010). A second early line appears in RASCAL, which treats propositional rules as a data-level resource: rules are unfolded into operational form, turned into virtual samples, and also used to prune training instances likely to be flawed, so that any downstream inductive learner can benefit from domain knowledge without architecture-specific modification (Iqbal, 2011).

Later work broadens DRKE from search guidance and preprocessing to structured inference over richer representations. Ontology matching with knowledge rules adds a third signal beyond terminology and ontology structure by preferring alignments that preserve local rule patterns such as ordering constraints, decision-tree relations, or association rules (Jiang et al., 2015). Domain Representation for Knowledge Graph Embedding models relation-specific head and tail domains as hyper-ellipsoids in embedding space, so that candidate entities outside the plausible semantic domain receive an inference-time penalty (Wang et al., 2019). A loosely coupled framework for knowledge graph completion then combines embeddings with ontology-based reasoning by iteratively exchanging inferred triples rather than forcing both components into a single joint objective (Kaoudi et al., 2022).

A distinct human-in-the-loop branch makes the rules themselves the feedback channel. Decision Rule Elicitation for Domain Adaptation asks experts not only for labels, but also for the decision rule used on misclassified instances, with the rules represented as Boolean predicates or shallow decision trees and treated as weak learners that can improve lifelong learning and domain adaptation under distribution shift (Nikitin et al., 2021). More recent LLM-era work extends the same logic to prompt-based knowledge plugins, domain KG augmentation services, privacy-preserving multi-agent systems, natural-language rule authoring, rule-aware knowledge editing, and differentiable fuzzy-rule modulation in vision models (Ding et al., 2022, Yao et al., 2023, Wang et al., 2023, Cheng et al., 2024, Yang et al., 12 Sep 2025, Wang, 3 Mar 2026, Srinivas et al., 30 Apr 2026).

2. Formal mechanisms and computational patterns

The cited literature instantiates DRKE through several recurring computational patterns.

Mechanism family	Representative operation	Representative papers
Search-space pruning	user-specified attributes define a working database $D'$	(Singh et al., 2010)
Data refinement	virtual sample creation and flawed-instance deletion	(Iqbal, 2011)
Weak-rule fusion	combine historical classifier and elicited feedback rules	(Nikitin et al., 2021)
Soft rule constraints	prefer alignments or edits that preserve rule structure	(Jiang et al., 2015, Cheng et al., 2024)
Domain geometry	penalize candidates outside relation-specific domains	(Wang et al., 2019)
Modular symbolic-statistical coupling	iterate between learned predictions and deductive inference	(Kaoudi et al., 2022)
Prompt/agent enhancement	select, verbalize, retrieve, or route domain knowledge	(Yao et al., 2023, Yang et al., 12 Sep 2025, Wang, 3 Mar 2026)
Differentiable logic update	rule-derived logit modulation via fuzzy inference	(Srinivas et al., 30 Apr 2026)

In rule-fusion settings, the canonical formalization is the feedback classifier of decision rule elicitation. Each elicited rule $f_i$ is treated as a weak learner, weighted by a learned coefficient $\theta_i$ and modulated by a similarity term to the test example that triggered the feedback: $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$ This rule-based component is then mixed with the historical classifier through

$$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$

with a possible extension to input-dependent $\alpha(x)$ (Nikitin et al., 2021). The same paper explicitly treats the rules as imperfect but informative, which is central to DRKE: enhancement need not rely on globally correct rule bases.

In data-preprocessing settings, RASCAL operationalizes each rule, scores it by correctness, scope, and generalization capability, and combines these into

$U(r)=C(r)+E(r)+G(r).$

The utility controls both how many virtual samples a rule generates and whether examples inconsistent with strong rules should be deleted (Iqbal, 2011). DRKE here is neither post-hoc prompting nor architectural modification; it is transformation of the data distribution.

In ontology matching, the rule signal enters as soft probabilistic constraints inside a Markov logic network. The method rewards proposed matches when corresponding entities participate in corresponding rule patterns, and for numerical attributes it relaxes exact matching into a distance-based soft constraint rather than requiring identical thresholds (Jiang et al., 2015). In knowledge graph completion, loose coupling takes a different route: a KGE model predicts high-confidence triples, a reasoner infers further triples from ontologies or expert rules, and the augmented graph is fed back into the embedding stage until a fixpoint is reached (Kaoudi et al., 2022).

Geometric domain modeling provides yet another formalization. For each relation $r$, DRE defines head and tail domains and represents each as a hyper-ellipsoid

$(x-a)^T M (x-a)=1,$

with a test-time penalty

$D'$0

The enhancement is rule-like rather than symbolic: a candidate that lies outside the learned domain is penalized even if the baseline embedding score is strong (Wang et al., 2019).

3. Eliciting and authoring domain rules

The most explicit human-centered DRKE mechanism is decision rule elicitation. A model $D'$1 is trained on historical data and deployed on a sequence of test sets in a lifelong-learning setting. When the system misclassifies a sample, the expert is asked not merely to relabel it but to provide a decision rule describing how they would classify it. The rules are expected to be expressible as Boolean predicates or shallow decision trees, and the authors connect this design to fast-and-frugal heuristics (Nikitin et al., 2021). In simulation, experts are modeled as limited-depth CART trees, where depth controls rule complexity and experience controls how much of the data distribution the expert has seen.

The empirical evidence in the same work is notable because it distinguishes rule feedback from label feedback under distribution shift. In the synthetic switching-distribution experiment, decision rule feedback yields Train distr. $D'$2, Test distr. $D'$3, and Train+test $D'$4; labels feedback with 40 experts yields Train distr. $D'$5, Test distr. $D'$6, and Train+test $D'$7; no feedback yields Train distr. $D'$8, Test distr. $D'$9, and Train+test $f_i$0 (Nikitin et al., 2021). In the real-user study on the same kind of switching distribution, with $f_i$1 and SGD, the decision-rule-feedback model improves test performance from $f_i$2 to $f_i$3 while not harming training performance significantly. The study also reports that more experienced participants tended to create simpler and more precise rules, whereas less experienced participants often produced overcomplicated rules and revised them more frequently.

KALMRA addresses the complementary problem of authoring rules and actions for KRR systems in a form accessible to domain experts. It lets users write factual English, converts those inputs into frame-based semantic representations and Unique Logical Representations, and reasons with DLV rather than XSB so that disjunctive heads, action semantics, and temporal reasoning become available (Wang et al., 2023). For rules, it supports safe forms of the type “If $f_i$4, then $f_i$5,” and for actions it uses the Simplified Event Calculus with constructs such as initiates/2, terminates/2, and holdsAt/2. On the UTI guidelines benchmark it achieves $f_i$6 correctness on rule authoring, and on bAbI it achieves $f_i$7 average accuracy for authoring and reasoning with actions (Wang et al., 2023).

A common misconception is that DRKE requires a fully formal knowledge-engineering workflow from the outset. The decision-rule-elicitation work instead relies on imperfect shallow heuristics, while KALMRA shows that relatively natural factual English can be compiled into logical rules and action representations (Nikitin et al., 2021, Wang et al., 2023). This suggests that DRKE spans both lightweight heuristic extraction and high-fidelity knowledge authoring.

4. Data-centric, ontology-centric, and graph-centric DRKE

In association-rule mining, DRKE appears as search-space control. Interactive Apriori begins by asking the user to specify attribute(s) of interest, scrutinizes the database to retain only transactions containing those attributes, and mines the resulting working database $f_i$8 instead of the entire database $f_i$9 (Singh et al., 2010). On artificial datasets with 7 attributes, minimum support levels of $\theta_i$0 and $\theta_i$1, and sizes from $\theta_i$2 to $\theta_i$3, IAR consistently reduces the number of candidates, frequent itemsets, T-tree nodes, T-tree updates, storage of the T-tree in bytes, and runtime. The paper states that IAR always takes less time than Apriori and that the advantage becomes more noticeable as data size increases.

RASCAL shifts DRKE from search control to training-set refinement. Rules are unfolded into operational rules whose antecedents are conjunctions of atomic literals and whose consequents are the class label. Virtual examples are then created by satisfying the rule antecedent and filling unspecified feature values randomly, while a pruning stage removes samples that conflict with high-utility rules (Iqbal, 2011). On promoter recognition and splice-junction recognition, RASCAL-refined datasets improve accuracy across feed-forward neural networks, SVM, C4.5, kNN, and KBANN. The gain is stronger in the $\theta_i$4 data setting, virtual samples alone perform surprisingly well, KBANN still improves even though it already uses the same rules, and performance improves up to roughly $\theta_i$5–$\theta_i$6 virtual samples before too many virtual samples begin to hurt performance (Iqbal, 2011).

Ontology matching with knowledge rules generalizes DRKE to schema alignment. KAOM defines a knowledge rule as a local relation $\theta_i$7 over ontology entities and adds soft constraints that favor alignments preserving these rules, whether the rules arise from Bayesian networks, decision trees, association rules, or expert knowledge (Jiang et al., 2015). The method is particularly effective for complex correspondences. On NBA, it successfully matches all numerical and nominal attributes without name similarity; on Census, at $\theta_i$8, it finds 6 out of 8 value correspondences for adult:workclass and income:class_of_worker, while MLOM finds none; and on OntoFarm it achieves better recall and much better performance on complex correspondences than baselines such as CODI and logmap2 (Jiang et al., 2015).

Graph-centric DRKE appears in two complementary forms. DRE defines relation-specific head and tail domains and learns a hyper-ellipsoid for each. On WN18, DRE(STransE) improves filtered MR from $\theta_i$9 to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$0 and Hits@10 from $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$1 to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$2; on FB15K, DRE(STransE) improves MR from $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$3 to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$4 and Hits@10 from $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$5 to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$6 (Wang et al., 2019). The loosely coupled KGE-reasoning framework then combines any pluggable KGE model with any monotonic forward-chaining reasoner, exchanging inferred triples iteratively and improving MRR by up to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$7 over vanilla embeddings and up to $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$8 over hybrid baselines such as pLogicNet (Kaoudi et al., 2022).

KnowledgeDA turns a domain KG into a unified augmentation service for PLM fine-tuning. It localizes domain entities in text via embedding similarity, generates augmented samples from both KG-based and training-data-based entity retrieval, and selects high-quality augmented samples via confidence-based assessment around $$C_{feedback}(x) = \sigma\!\left(\sum_{i=1}^{F} f_i(x)\, sim(X_{test}^{i}, x)\, \theta_i\right).$$9 (Ding et al., 2022). The method improves accuracy by about $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$0 over direct fine-tuning on CMID and KUAKE-QIC, yields the highest novel entity coverage, improves RoBERTa-large and DeBERTa-large on TRANS by $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$1 and $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$2, improves SO-PLC by about $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$3 accuracy and $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$4 F1 over no augmentation, and also improves CMedQA and PubMedQA (Ding et al., 2022).

5. DRKE in LLMs, multi-agent systems, and knowledge editing

Prompt-based DRKE appears explicitly in DOKE, where a domain knowledge extractor prepares effective knowledge, customizes it to the current sample, and expresses it in an LLM-understandable way without fine-tuning the model parameters (Yao et al., 2023). In recommender systems, the knowledge includes item attributes, item-to-item collaborative filtering, user-to-item relevance, and knowledge-graph reasoning paths. The strongest effects come from sample-specific CF knowledge: on ML1M, ChatGPT rises from NDCG@1 $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$5 to $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$6 and from NDCG@10 $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$7 to $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$8; on Beauty, NDCG@1 rises from $$\mathcal{C}(x) = \alpha \mathcal{C}_{hist}(x, \theta_{hist}) + (1 - \alpha) \mathcal{C}_{feedback}(x, \theta_{f}),$$9 to $\alpha(x)$0 and NDCG@10 from $\alpha(x)$1 to $\alpha(x)$2; on Online Retail, NDCG@1 rises from $\alpha(x)$3 to $\alpha(x)$4 and NDCG@10 from $\alpha(x)$5 to $\alpha(x)$6, with the gains reported as statistically significant at $\alpha(x)$7 (Yao et al., 2023).

GSI Agent applies a related logic to a regulation-heavy engineering domain. It combines LoRA-based SFT on a 10,955-example GSI Dataset, RAG over an internal corpus of municipal manuals, plans, regulations, inspection forms, and related documents, and an agent layer governed by three soft rules: use retrieved passages when relevant, do not invent regulations or technical standards, and ask concise clarification questions if critical information is missing (Wang, 3 Mar 2026). On the GSI dataset, BLEU-4 improves from $\alpha(x)$8 to $\alpha(x)$9, while on the common knowledge dataset it remains essentially stable at $U(r)=C(r)+E(r)+G(r).$0 versus $U(r)=C(r)+E(r)+G(r).$1. The ablation with G-Eval reports $U(r)=C(r)+E(r)+G(r).$2 for LLM+RAG, $U(r)=C(r)+E(r)+G(r).$3 for LLM+Fine-tuning, and $U(r)=C(r)+E(r)+G(r).$4 for LLM+RAG+Fine-tuning (Wang, 3 Mar 2026).

GAMA places DRKE inside a privacy-preserving multi-agent system. After AMPP anonymizes sensitive entities and separates private and public spaces, DRKE operates in the public space by using a Domain-Analyzing Agent, Word2Vec- and TF-IDF-based domain relation estimation, top-$U(r)=C(r)+E(r)+G(r).$5 domain selection, dynamic creation of Domain Expert Agents, and semantic fusion of their outputs (Yang et al., 12 Sep 2025). The domain set used in experiments includes Entertainment, Finance, History, Law, Sociology, Arts, Literature, Medicine, Politics, Sports, Technology, Business, Communication, Science, Education, Environment, Agriculture, Philosophy, Transportation, and Psychology. On public QA datasets, GAMA reports $U(r)=C(r)+E(r)+G(r).$6 on TCW $U(r)=C(r)+E(r)+G(r).$7, $U(r)=C(r)+E(r)+G(r).$8 on TCW $U(r)=C(r)+E(r)+G(r).$9, and $r$0 on LGP; on privacy-aware datasets it reports $r$1 on KPP and $r$2 on LPP (Yang et al., 12 Sep 2025). The ablation table reports GAMA $r$3, $r$4 $r$5, $r$6 $r$7, and $r$8 $r$9 on KPP/LPP, while the authors’ discussion still describes DRKE as particularly important for knowledge-heavy tasks; this tension is part of the paper’s own presentation (Yang et al., 12 Sep 2025).

KLUE pushes DRKE into differentiable neuro-symbolic vision. It learns latent support concepts without concept labels, constructs Horn-like implication rules between those concepts and classes, and uses a Differentiable Knowledge Unit to compute a logic-based correction

$(x-a)^T M (x-a)=1,$0

which modulates class logits before supervision is applied to the refined probabilities (Srinivas et al., 30 Apr 2026). On COCO with WRN-101, the best KLUE variant improves mAP by $(x-a)^T M (x-a)=1,$1 and AUC by $(x-a)^T M (x-a)=1,$2 over the baseline; on PASCAL VOC 2012, it improves by $(x-a)^T M (x-a)=1,$3 mAP and $(x-a)^T M (x-a)=1,$4 AUC; on ChestMNIST, WRN-101 baseline at $(x-a)^T M (x-a)=1,$5 mAP / $(x-a)^T M (x-a)=1,$6 AUC rises to $(x-a)^T M (x-a)=1,$7 mAP / $(x-a)^T M (x-a)=1,$8 AUC for KLUE(v2) with the stronger SAT setting (Srinivas et al., 30 Apr 2026). In domain generalization from PASCAL VOC to COCO, WRN-101 mAP rises from $(x-a)^T M (x-a)=1,$9 to $D'$00, and AUC from $D'$01 to $D'$02 (Srinivas et al., 30 Apr 2026).

RULE-KE extends DRKE to knowledge editing for multi-hop QA. It mines logical rules from the base knowledge graph with AMIE-3, selects rules whose body relations are semantically correlated with a requested edit using a threshold $D'$03, infers correlated knowledge by forward and backward tracking, and converts the inferred facts into additional edits that augment either parameter-based or memory-based knowledge editors (Cheng et al., 2024). On RKe-eval, RULE-KE boosts parameter-based methods by up to $D'$04 and memory-based methods by up to $D'$05; on MQuAKE-CF-3K the gains are smaller because the benchmark is more rule-sparse (Cheng et al., 2024).

6. Empirical regularities, misconceptions, and open issues

Several empirical regularities recur across the literature. First, DRKE is repeatedly most valuable when the observed data are incomplete, shifted, anonymized, or otherwise semantically impoverished. Decision-rule elicitation outperforms label feedback under switching distributions, DRE improves link prediction by filtering out semantically implausible entities, loose KGE-reasoning improves sparse-graph completion, GAMA uses DRKE to compensate for anonymization-induced semantic loss, and KLUE improves both hard-sample behavior and domain transfer (Nikitin et al., 2021, Wang et al., 2019, Kaoudi et al., 2022, Yang et al., 12 Sep 2025, Srinivas et al., 30 Apr 2026). This suggests that DRKE often functions as an invariance or consistency mechanism rather than merely an accuracy booster.

Second, DRKE should not be reduced to hand-written symbolic rules. The surveyed methods include user-specified attribute preferences, propositional rules unfolded into virtual data, elicited Boolean predicates, Bayesian-network and decision-tree relations for ontology matching, learned relation domains in embedding space, collaborative-filtering signals verbalized into prompts, KG-derived entity substitutions, mined logical rules for knowledge editing, and latent support concepts learned without concept labels (Singh et al., 2010, Iqbal, 2011, Jiang et al., 2015, Yao et al., 2023, Ding et al., 2022, Cheng et al., 2024, Srinivas et al., 30 Apr 2026). A related misconception is that rule enhancement requires tight neuro-symbolic coupling; the loosely coupled KGE framework shows that iterative triple exchange is sufficient for large gains, and GSI Agent shows that soft prompt-level constraints can already enforce domain-grounded behavior (Kaoudi et al., 2022, Wang, 3 Mar 2026).

Third, the literature is equally clear that more rule-based augmentation is not always better. RASCAL reports that too many virtual samples hurt performance beyond roughly $D'$06–$D'$07;