Research map / Learning theory and algorithms / Prediction and decision losses
Multipartite Entity Resolution: Motivating a K-Tuple Perspective (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Argues that multipartite entity resolution should be evaluated with k-tuple precision/recall rather than pairwise decomposition, showing on synthetic 4-dataset person records that k-tuple scores are far lower than pairwise scores.
Problem
Evaluation of multipartite entity resolution (M-ER) over k>2 datasets: standard ER metrics assume two datasets, and the common pairwise decomposition into bipartite tasks can yield inconsistent (transitivity-violating) predictions and overstated performance.
Approach
Proposes evaluating M-ER with k-tuple precision and recall, where a predicted k-tuple is correct only if the exact same tuple appears in the ground truth, and shows this differs substantially from pairwise and bipartite-average evaluation.
Key results
- Pairwise and bipartite-average baselines reach F-measures around 0.96 with TF-IDF, while k-tuple evaluation with connected components gives 0.864 F-measure (TF-IDF) and drops to 0.609 with Levenshtein.
- Cliques-based k-tuple evaluation gives F-measures of 0.785 (TF-IDF), 0.692 (Jaro-Winkler), and 0.573 (Levenshtein).
- The result of M-ER depends on the combination method (connected components vs cliques), suggesting the pipeline should operate on k-tuples rather than pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate multipartite predictions as k-tuples instead of aggregated pairs | TF-IDF with connected components · Synthetic M-ER, k=4 | F-measure: 0.961 -> 0.864 | Hurt |
| Use cliques instead of connected components for clustering | k-Tuple evaluation with TF-IDF · Synthetic M-ER, k=4 | F-measure: 0.864 -> 0.785 | Hurt |
Surprising findings
- Pairwise evaluation, which is the current literature standard, appears to overstate M-ER performance: near-solved pairwise scores (~0.96) collapse to 0.57-0.86 F-measure under k-tuple evaluation.
Limitations
- Preliminary experiments on a single synthetic dataset with k=4 and a simple algorithm.
- Does not propose a new M-ER method or a new k-tuple metric.
- Blocking evaluation for M-ER is not addressed.
Experimental setup
Datasets: GeCo-generated synthetic corrupted person records (k=4 datasets)
Benchmarks: synthetic 4-source person record M-ER data
Metrics: precision recall F-measure
Baselines: Bipartite Avg Pair-wise evaluation FAMER
Builds on: FAMER GeCo
Closest papers by meaning
- A Theoretical Analysis of First Heuristics of Crowdsourced Entity Resolution (AAAI 2017 · similarity 0.86)
- Accelerating Column Generation via Flexible Dual Optimal Inequalities with Application to Entity Resolution (AAAI 2020 · similarity 0.86)
- MESED: A Multi-Modal Entity Set Expansion Dataset with Fine-Grained Semantic Classes and Hard Negative Entities (AAAI 2024 · similarity 0.83)
- Unaligned Message-Passing and Contextualized-Pretraining for Robust Geo-Entity Resolution (AAAI 2025 · similarity 0.82)
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.81)
- NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.