Semantics-Aware Caching for Concept Learning
| Authors | Louis Mozart Kamdem Teyou et al. |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2603.06506 |
| Download | |
| Categories | stat.ML, cs.LG |
Abstract
Concept learning is a form of supervised machine learning that operates on knowledge bases in description logics. State-of-the-art concept learners often rely on an iterative search through a countably infinite concept space. In each iteration, they retrieve instances of candidate solutions to select the best concept for the next iteration. While simple learning problems might require a few dozen instance retrieval calls to find a fitting solution, complex learning problems might necessitate thousands of calls. We alleviate the resulting runtime challenge by presenting a semantics-aware caching approach. Our cache is essentially a subsumption-aware map that links concepts to a set of instances via crisp set operations. Our experiments on 5 datasets with 4 symbolic reasoners, a neuro-symbolic reasoner, and 5 popular pagination policies demonstrate that our cache can reduce the runtime of concept retrieval and concept learning by an order of magnitude while being effective for both symbolic and neuro-symbolic reasoners.
Engineering Breakdown
Plain English
This paper addresses a performance bottleneck in concept learning—a supervised ML approach that searches through infinite concept spaces defined in description logics. The authors found that complex learning problems require thousands of instance retrieval calls to find a good solution, making iterative search prohibitively slow. They introduce a semantics-aware caching layer that stores mappings between concepts and their instances using subsumption-aware operations, dramatically reducing redundant retrievals. Experiments across 5 datasets with 4 symbolic reasoners and neuro-symbolic systems show significant speedup without sacrificing solution quality.
Core Technical Contribution
The key innovation is a subsumption-aware caching mechanism that exploits semantic relationships between concepts in description logic spaces. Unlike standard caching that treats concepts as opaque keys, this approach leverages subsumption (the 'is-more-specific-than' relation) to reuse instance computations across semantically related concepts. The cache works as a semantics-preserving map where one concept's cached instances can inform decisions about related concepts through crisp set operations, effectively amortizing the cost of instance retrieval across the entire iterative search. This is novel because prior concept learners did not exploit these logical relationships to reduce redundant lookups.
How It Works
The system operates within an iterative concept learning loop where each iteration generates candidate concepts and retrieves their instances from a knowledge base. When a candidate concept is evaluated, the cache first checks if the concept or semantically related concepts already have cached instances stored. If found, set operations (intersection, union, subsumption checks) are applied to determine if cached results can satisfy the current query without hitting the underlying knowledge base. If not cached, the reasoner performs instance retrieval and the result is stored with semantic annotations marking its position in the subsumption hierarchy. Subsequent iterations that encounter concepts subsuming or subsumed by cached concepts can reuse or efficiently refine those cached instance sets, avoiding expensive re-queries to the symbolic reasoner.
Production Impact
For production systems using concept learning (common in knowledge graph completion, ontology learning, and semantic web applications), this cache reduces wall-clock runtime by orders of magnitude on complex problems by cutting instance retrieval calls from thousands to dozens. Integration is straightforward: the cache sits transparently between the concept learning algorithm and the symbolic reasoner, requiring no changes to the core learner logic. The main trade-off is memory overhead proportional to the number of distinct concepts explored and their instance sets, which is acceptable since the speedup typically pays for itself within the first few iterations. This is particularly valuable in interactive or time-constrained settings (e.g., real-time ontology refinement systems, automated ML pipeline optimization) where reducing reasoning latency directly improves user experience.
Limitations and When Not to Use This
The approach assumes subsumption relationships are well-defined and computable, which holds in classical description logics but may degrade with fuzzy or probabilistic ontologies where subsumption becomes approximate. The cache effectiveness depends heavily on the structure of the concept space and how exploration orders interact with subsumption hierarchies—adversarial orderings or very flat hierarchies may show minimal benefit. The paper does not address scalability to extremely large knowledge bases where the instance sets themselves become prohibitively large to cache, nor does it handle dynamic knowledge bases where instance membership changes between iterations. Further work is needed on cache invalidation strategies, memory budgeting policies, and adaptation to non-subsumption-based concept relationships common in neural-symbolic systems.
Research Context
This work builds on decades of research in inductive logic programming (ILP) and description logic learning, where concept learning has been a core task. It specifically targets a performance problem identified in modern symbolic reasoners and neuro-symbolic systems that use iterative refinement for concept discovery. The paper extends prior optimization work (e.g., heuristic pruning, early stopping) by introducing a semantic caching dimension that hasn't been systematically explored. By bridging symbolic reasoning efficiency with practical ML scalability, this work opens research directions in caching strategies for other iterative symbolic reasoning tasks like abductive reasoning, planning, and query optimization.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::