Ilmu Komputer & AI editorial
GRAFT: Graph-Matched Retrieval and Fusion of Tables in Data Lakes
The core problem
Autonomous data agents answer analytical queries by retrieving and reasoning over evidence scattered across tabular data lakes. Existing retrieval methods score each table independently against the query and therefore ignore the joinability and unionability relationships that link tables together. As a result, they return fragmented evidence that downstream agents cannot integrate into a coherent answer.
GRAFT (Graph-matched Retrieval and Fusion of Tables) addresses this gap with two principal contributions. First, it casts table retrieval as a graph matching problem between a query-derived intent graph and a heterogeneous data lake graph, and introduces IGMS, a log-determinant reward that couples semantic relevance, structural compatibility, and evidence diversity in a single objective. Second, it recasts subgraph generation as a Markov decision process and learns a value function via implicit Q-learning on self-generated trajectories produced by a canonical compression operator that inverts the homomorphism. A three-stage online pipeline then exploits anchor reachability, predicate admissibility, and reward monotonicity to prune the candidate space before exact IGMS evaluation.
Innovation
GRAFT was evaluated on Spider and BIRD adapted to the tabular data lake setting. The baselines include point-wise scoring, greedy-expansion, and structure-aware retrieval methods. Across both benchmarks, GRAFT achieves the best Recall, Precision, , and Sufficiency.
Relative to the strongest baseline, GRAFT reports a 7.8% improvement in and a 10.6% improvement in Sufficiency. Sufficiency measures whether the retrieved evidence is complete enough for a downstream agent to answer the query, so the gain indicates that GRAFT returns more integrable evidence rather than merely more relevant tables. The method also maintains high search efficiency, meaning the three-stage pruning pipeline keeps the exact IGMS evaluation tractable despite the combinatorial subgraph space.
The consistent gains across point-wise, greedy-expansion, and structure-aware baselines suggest that jointly modeling semantic relevance, structural compatibility, and evidence diversity is the key driver of performance, rather than any single factor in isolation.
Why it matters
The results support the central claim that table retrieval in data lakes should not be treated as independent per-table scoring. By casting retrieval as graph matching, GRAFT explicitly captures joinability and unionability, which are precisely the relations that allow downstream agents to fuse evidence across tables. The IGMS reward's log-determinant term is particularly important: it encourages diversity in the selected subgraph, preventing the retrieval from returning many near-duplicate tables that would inflate relevance scores without improving answerability.
The implicit Q-learning component addresses the combinatorial explosion of subgraph generation. The canonical compression operator inverts the homomorphism between the intent graph and the lake graph, which allows the value function to be learned on self-generated trajectories without requiring expensive expert demonstrations. This is a practical advantage for deployment in dynamic data lakes where new tables are added frequently.
The three-stage online pipeline provides the efficiency needed for real-world use. Anchor reachability, predicate admissibility, and reward monotonicity each prune the candidate space using cheap checks, so exact IGMS evaluation is reserved for a small set of promising candidates. The reported high search efficiency confirms that this pruning does not sacrifice retrieval quality.
Limitations include the reliance on a well-defined intent graph derived from the query, which may be challenging for ambiguous natural-language queries. Future work could explore adaptive graph construction and extending the framework to cross-modal evidence beyond tables.
Who should read this
Opening member content…