[論文レビュー] Semantics-Aware Caching for Concept Learning

Louis Mozart Kamdem Teyou, Caglar Demir+1 more|arXiv (Cornell University)|Mar 6, 2026

Advanced Graph Neural Networks被引用数 0

ひとこと要約

意味論を意識したキャッシュ層が概念取得を加速し、記号論理推論器における概念学習パイプラインをシンボリックおよびニューロ・シンボリック設定の双方で劇的に高速化する。

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.

研究の動機と目的

• Identify the runtime bottleneck in concept learning (repeated reasoning for candidate concepts).
• Propose a semantics-aware cache that exploits subsumption to reuse previously computed instance sets.
• Evaluate cache impact across multiple datasets, reasoners, and cache eviction policies.
• Demonstrate that cache integration yields substantial runtime savings and generalizes to neuro-symbolic reasoning.

提案手法

• Introduce a subsumption-aware cache that maps concepts to instance sets using crisp set operations.
• Exploit AL C semantics to prune candidate retrieval via identified subsumption relationships with lightweight syntactic heuristics.
• Cache initialization precomputes Ret(A), Ret(¬A), and Ret(∃r.A) for common primitives to accelerate warm starts.
• Implement a fetchInstances procedure that uses semantic decomposition and cached results to minimize reasoner calls.
• Evaluate five eviction policies (FIFO, LIFO, LRU, MRU, RP) and compare semantic vs non-semantic caching.
• Test across four datasets with four symbolic reasoners and one neuro-symbolic reasoner (Ebr).

(a) Carcinogenesis

実験結果

リサーチクエスチョン

• RQ1 Does the semantics-aware cache reduce concept retrieval time across different reasoners?
• RQ2 How do cache size and eviction policy influence runtime and hit rate in concept retrieval?
• RQ3 Is the cache effective for both symbolic and neuro-symbolic concept learning tasks?
• RQ4 What is the impact of cache initialization on performance on small vs. large datasets?

主な発見

• Cache reduces concept retrieval time by up to 80% for slower reasoners and up to 20% for faster ones when cache capacity is sufficient.
• In concept learning, the cache can reduce overall runtimes by up to three orders of magnitude.
• LRU consistently provides the best trade-off between runtime reduction and cache efficiency across reasoners and datasets.
• Warm (initialized) caches yield more stable and faster improvements as cache size grows, despite higher startup cost.
• Non-semantic caching provides little or no improvement, highlighting the importance of semantic structure exploitation.
• On large datasets, faster reasoners still benefit (e.g., Ebr: from ~700ks to ~100ks in Carcinogenesis), and memory usage remains manageable with proper eviction.

(b) Mutagenesis