Beyond Row-Level Prediction: Evaluating and Learning Composition for Reusable Table Representations
Organizations: SAP SE, Berlin, Germany · TU Darmstadt, Darmstadt, Germany · Hessian AI, Darmstadt, Germany
Abstract
Tables are compositional, with content spread across a schema, cell values, rows, and columns. Large-scale operations such as retrieval, overlap estimation, deduplication, and data curation need these local signals composed into one reusable embedding vector per table. Table representation methods, however, are mostly evaluated through row-level prediction or supervised downstream tasks, which do not measure table-level embedding quality directly. We introduce a benchmark that evaluates table-level embeddings through four desiderata: consistency under partial views, discriminability across label granularities, robustness to benign perturbations, and efficiency. Across twelve methods on synthetic and real corpora, lightweight schema- and text-based fingerprints are among the strongest. The sharpest contrast comes from two tabular foundation models, TabPFN and ConTextTab: although both are strong row-level predictors, their states, composed by mean, max, or three further parameter-free pooling rules, trail the strongest fingerprints. We therefore ask whether the gap lies in the encoder or in the compositional rule. Keeping ConTextTab frozen, we learn only the compositional rule over its cell states, using support-conditioned distribution matching (SCDM), a self-supervised loss we propose that matches the distribution of embedding similarities to the distribution of partition overlaps within and across tables. Across all fifteen head configurations, the learned rule raises the Spearman correlation between embedding similarity and partial-view overlap from 0.04 to 0.43-0.50 on synthetic data and improves triplet ranking, clustering, and retrieval, with smaller but systematic gains on real data. A downstream parent-table retrieval task confirms these findings. How local states are composed, and with what objective, substantially affects the quality of table-level embeddings.
Figures & tables
| Family | Methods | Main signal / representation |
| Structural & statistical | HashingSchema, SchemaContent, TableStatistics, StatisticalSummary, MatrixFactorization | Schema tokens, metadata, summary statistics, and numeric spectrum; direct table-level embeddings. |
| Serialization-based text | HashingText, LLMText | CSV-style serialization encoded by feature hashing (HashingText) or by the all-MiniLM-L6-v2 sentence-transformer (LLMText). |
| Specialized representation | TableVectorizer, HyTREL, Armadillo | Mixed-type row vectorization, hypergraph encoding, or graph neural encoding with pooled table-level embeddings. |
| Foundation-model-derived | TabPFN, ConTextTab | Row- or cell-level hidden states of pretrained TFMs, composed into one table-level embedding by a fixed pooling rule (mean or max in the main tables). |
| Embedder | Synthetic | Real | ||||||
| D1 Spearman | TR-Avg | CL-Avg | R@5 | D1 Spearman | TR-Avg | CL-Avg | R@5 | |
| HashingSchema | 0.59 0.01 | 0.53 0.01 | 0.54 0.02 | 0.71 0.01 | 0.31 0.03 | 0.59 0.03 | 0.40 0.03 | 0.96 0.01 |
| HashingText | 0.54 0.04 | 0.47 0.01 | 0.46 0.03 | 1.00 0.00 | 0.41 0.03 | 0.51 0.02 | 0.28 0.03 | 0.98 0.00 |
| TabPFN (mean) | -0.08 0.05 | 0.17 0.00 | 0.25 0.02 | 0.73 0.01 | 0.06 0.04 | 0.13 0.01 | 0.20 0.01 | 0.60 0.02 |
| TabPFN (max) | -0.17 0.04 | 0.21 0.01 | 0.25 0.01 | 0.71 0.01 | 0.07 0.04 | 0.16 0.01 | 0.18 0.01 | 0.60 0.01 |
| TabPFN (sum) | -0.08 0.05 | 0.17 0.00 | 0.25 0.02 | 0.73 0.01 | 0.06 0.04 | 0.13 0.01 | 0.20 0.01 | 0.60 0.02 |
| Data | Embedder | D1 Consistency | D2 Label-Based - Avg | D3 Robustness | |||||
| Spearman | TR-Avg | CL-Avg | R@5 | Perm | Noise | Mask | |||
| Synthetic | HashingSchema | – | 0.59 0.01 | 0.53 0.01 | 0.54 0.02 | 0.71 0.01 | 1.00 0.00 | 0.98 0.00 | 0.99 0.00 |
| HashingText | – | 0.54 0.04 | 0.47 0.01 | 0.46 0.03 | 1.00 0.00 | 1.00 0.00 | 0.65 0.02 | 0.97 0.00 | |
| ConTextTab (pooled) | – | 0.04 0.03 | 0.18 0.00 | 0.27 0.02 | 0.70 0.01 | 0.97 0.00 | 0.97 0.00 | 0.97 0.00 | |
| ConTextTab -R Attention | 0.75 | 0.50 0.02 | 0.34 0.01 | 0.46 0.02 | 0.88 0.01 | 1.00 0.00 | 1.00 0.00 | 0.99 0.00 | |
| ConTextTab -R GatedAttention | 1.00 | 0.48 0.02 | 0.31 0.01 | 0.46 0.03 | 0.89 0.01 | 1.00 0.00 | 1.00 0.00 | 0.99 0.00 | |
| Embedder | Synthetic | Real | ||||||
| R@1 | R@5 | R@10 | MRR | R@1 | R@5 | R@10 | MRR | |
| HashingSchema | 0.10 0.00 | 0.47 0.01 | 0.78 0.02 | 0.28 0.00 | 0.67 0.02 | 0.97 0.01 | 0.98 0.00 | 0.80 0.01 |
| SchemaContent | 0.01 0.00 | 0.05 0.00 | 0.10 0.00 | 0.05 0.00 | 0.10 0.01 | 0.28 0.02 | 0.38 0.02 | 0.20 0.01 |
| TableStatistics | 0.04 0.00 | 0.19 0.01 | 0.33 0.01 | 0.13 0.00 | 0.04 0.01 | 0.16 0.01 | 0.25 0.02 | 0.11 0.01 |
| StatisticalSummary | 0.32 0.02 | 0.69 0.02 | 0.82 0.02 | 0.48 0.02 | 0.25 0.02 | 0.51 0.03 | 0.63 0.03 | 0.38 0.02 |
| MatrixFactorization | 0.19 0.01 | 0.49 0.03 | 0.61 0.04 | 0.33 0.02 | 0.10 0.01 | 0.27 0.02 | 0.38 0.03 | 0.19 0.01 |
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| Embedder | Dim | Key Hyperparameters |
| HashingSchema | 1024 | n_features =1024 (feature hashing on schema) |
| SchemaContent | 256 | n_features =256 |
| TableStatistics | 14 | output_dim =14 (fixed statistical features) |
| StatisticalSummary | 22 | Numeric: min, max, mean, std, skew, median; Categorical: cardinality, entropy |
| MatrixFactorization | 16 | n_components =16, center =True |
| HashingText | 1024 | n_features =1024 (feature hashing on text) |
| Desideratum | Metric | Configuration |
| D1: Consistency | Pearson correlation | IoU vs. cosine similarity |
| Spearman correlation | IoU vs. cosine similarity | |
| MSE | IoU vs. cosine similarity | |
| D2: Triplet | TR-R (Random) | Random negative, margin=0.01 |
| TR-H (Hard) | Hardest negative mining | |
| TR-CH (Cluster-Hard) | Hard negatives within clusters |
| Parameter | Value | Parameter | Value | Parameter | Value |
| Tables | 100 | Partitions/table | 10 | Overlap ratio | 0.5 |
| Row fraction | [0.2, 0.5] | Column fraction | [0.2, 0.5] | Min partition | |
| Seeds | 10 | Rows (synthetic) | 100 | Rows (real) | varies |
| Synthetic – Semantic | Synthetic – Stat | Real – Stat | ||||
| Embedder | Label | Group | Label | Group | Label | Group |
| HashingSchema | 0.88 0.02 | 0.88 0.02 | 1.00 0.00 | 1.00 0.00 | 0.96 0.01 | 0.84 0.02 |
| SchemaContent | 0.88 0.02 | 0.87 0.02 | 1.00 0.00 | 0.99 0.00 | 0.97 0.01 | 0.96 0.01 |
| TableStatistics | 0.62 0.03 | 0.58 0.03 | 0.99 0.00 | 0.99 0.00 | 0.99 0.00 | 0.99 0.00 |
| StatisticalSummary | 0.71 0.03 | 0.68 0.03 | 0.99 0.01 | 0.99 0.00 | 0.92 0.02 | 0.90 0.01 |
| MatrixFactorization | 0.53 0.03 | 0.51 0.02 | 0.92 0.02 | 0.92 0.01 | 0.65 0.02 | 0.65 0.02 |
| Embedder | Synthetic | Real |
| HashingSchema | 0.77 0.01 | 0.92 0.01 |
| SchemaContent | 0.77 0.02 | 0.95 0.01 |
| TableStatistics | 0.63 0.02 | 0.64 0.02 |
| StatisticalSummary | 0.69 0.01 | 0.58 0.01 |
| MatrixFactorization | 0.61 0.02 | 0.30 0.01 |
| HashingText | 0.87 0.03 | 0.94 0.01 |
| Embedder | Input channel changed | D1 Spearman | D2 Avg | Summary | |||
| Anon. | Anon. | D3 Avg | Rank | ||||
| HashingSchema | Column names + dtypes | 0.29 | -0.29 | 0.49 | -0.12 | 0.00 | +2.68 |
| SchemaContent | Names/types + metadata | 0.26 | -0.02 | 0.36 | -0.05 | 0.00 | +0.25 |
| TableStatistics | No names; size/type stats | 0.32 | -0.02 | 0.42 | 0.00 | 0.00 | -0.38 |
| StatisticalSummary | No names; value summaries | 0.37 | -0.02 | 0.57 | 0.00 | 0.00 | -0.33 |
| MatrixFactorization | No names; numeric values | 0.39 | +0.02 | 0.45 | 0.00 | 0.00 | -0.92 |
| Embedder | Input channel changed | D1 Spearman | D2 Avg | Summary | |||
| Anon. | Anon. | D3 Avg | Rank | ||||
| HashingSchema | Column names + dtypes | 0.25 | -0.07 | 0.50 | -0.18 | -0.01 | +1.04 |
| SchemaContent | Names/types + metadata | 0.25 | -0.01 | 0.37 | -0.12 | 0.00 | +0.38 |
| TableStatistics | No names; size/type stats | 0.21 | -0.01 | 0.41 | -0.01 | 0.00 | +0.17 |
| StatisticalSummary | No names; value summaries | 0.21 | +0.01 | 0.50 | +0.01 | 0.00 | -0.33 |
| MatrixFactorization | No names; numeric values | 0.19 | +0.02 | 0.33 | -0.01 | 0.00 | -0.29 |
| Real data | |||||
| Pooling | D1 Spear. | Triplet avg | Cluster avg | R@5 | D2 avg |
| mean | 0.19 0.03 | 0.25 0.01 | 0.22 0.01 | 0.76 0.01 | 0.41 |
| max | -0.06 0.03 | 0.14 0.01 | 0.23 0.02 | 0.72 0.02 | 0.36 |
| sum | 0.19 0.03 | 0.25 0.01 | 0.22 0.01 | 0.76 0.01 | 0.41 |
| min | -0.08 0.03 | 0.15 0.01 | 0.22 0.01 | 0.69 0.01 | 0.35 |
| std | -0.05 0.03 | 0.03 0.01 | 0.27 0.03 | 0.77 0.01 | 0.36 |
| Real data | |||||
| Pooling | D1 Spear. | Triplet avg | Cluster avg | R@5 | D2 avg |
| mean | 0.06 0.04 | 0.13 0.01 | 0.20 0.01 | 0.60 0.02 | 0.31 |
| max | 0.07 0.04 | 0.16 0.01 | 0.18 0.01 | 0.60 0.01 | 0.31 |
| sum | 0.06 0.04 | 0.13 0.01 | 0.20 0.01 | 0.60 0.02 | 0.31 |
| min | 0.06 0.03 | 0.16 0.01 | 0.19 0.01 | 0.57 0.01 | 0.31 |
| std | 0.03 0.03 | 0.20 0.00 | 0.18 0.01 | 0.54 0.01 | 0.31 |
| Embedder | Triplet (Direct) | Clustering (Direct) | Retrieval/Probe (Direct) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.89 0.01 | 0.00 0.00 | 0.00 0.00 | 0.30 0.00 | 4.75 | 0.12 0.01 | 0.57 0.02 | 0.10 0.01 | 0.26 0.01 | 2.00 | 0.23 0.01 | 0.12 0.02 | 12.50 | 6.42 |
| SchemaContent | 0.02 0.00 | 0.00 0.00 | 0.00 0.00 | 0.01 0.00 | 13.50 | 0.07 0.01 | 0.34 0.01 | 0.03 0.00 | 0.15 0.01 | 10.50 | 0.35 0.01 | 0.11 0.02 | 11.00 | 11.67 |
| TableStatistics | 0.38 0.02 | 0.00 0.00 | 0.00 0.00 | 0.13 0.01 | 12.50 | 0.10 0.01 | 0.48 0.01 | 0.07 0.01 | 0.22 0.01 | 5.25 | 0.30 0.01 | 0.12 0.02 | 11.50 | 9.75 |
| StatisticalSummary | 0.80 0.01 | 0.00 0.00 | 0.01 0.00 | 0.27 0.00 | 4.25 | 0.11 0.01 | 0.51 0.01 | 0.07 0.00 | 0.23 0.01 | 4.25 | 0.75 0.01 | 0.18 0.02 | 4.50 | 4.33 |
| MatrixFactorization | 0.69 0.03 | 0.00 0.00 | 0.00 0.00 | 0.23 0.01 | 8.75 | 0.10 0.01 | 0.45 0.01 | 0.04 0.01 | 0.19 0.01 | 7.00 | 0.50 0.02 | 0.15 0.02 | 7.50 | 7.75 |
| Embedder | Triplet (Semantic) | Clustering (Semantic) | Retrieval/Probe (Semantic) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.96 0.01 | 0.49 0.03 | 0.50 0.02 | 0.65 0.02 | 1.25 | 0.79 0.02 | 0.85 0.02 | 0.70 0.04 | 0.78 0.03 | 2.00 | 0.99 0.00 | 0.88 0.02 | 2.50 | 1.92 |
| SchemaContent | 0.02 0.00 | 0.00 0.00 | 0.00 0.00 | 0.01 0.00 | 13.50 | 0.41 0.02 | 0.45 0.02 | 0.21 0.02 | 0.36 0.02 | 9.50 | 0.98 0.00 | 0.87 0.02 | 3.50 | 8.83 |
| TableStatistics | 0.40 0.02 | 0.00 0.00 | 0.00 0.00 | 0.13 0.01 | 12.50 | 0.53 0.02 | 0.61 0.03 | 0.42 0.03 | 0.52 0.02 | 5.25 | 0.92 0.01 | 0.58 0.03 | 9.50 | 9.08 |
| StatisticalSummary | 0.83 0.01 | 0.02 0.01 | 0.03 0.02 | 0.30 0.01 | 5.25 | 0.59 0.03 | 0.66 0.02 | 0.45 0.02 | 0.57 0.02 | 3.75 | 0.97 0.00 | 0.68 0.03 | 5.50 | 4.83 |
| MatrixFactorization | 0.71 0.03 | 0.00 0.00 | 0.00 0.00 | 0.24 0.01 | 8.25 | 0.56 0.03 | 0.56 0.02 | 0.29 0.04 | 0.47 0.03 | 5.75 | 0.89 0.03 | 0.51 0.02 | 12.00 | 8.67 |
| Embedder | Triplet (Semantic+Difficulty) | Clustering (Semantic+Difficulty) | Retrieval/Probe (Semantic+Difficulty) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.91 0.00 | 0.01 0.01 | 0.01 0.01 | 0.31 0.00 | 4.25 | 0.40 0.02 | 0.70 0.02 | 0.35 0.02 | 0.48 0.02 | 2.00 | 0.62 0.03 | 0.44 0.03 | 9.00 | 5.08 |
| SchemaContent | 0.02 0.00 | 0.00 0.00 | 0.00 0.00 | 0.01 0.00 | 13.38 | 0.21 0.01 | 0.38 0.02 | 0.12 0.01 | 0.24 0.02 | 9.50 | 0.79 0.02 | 0.43 0.03 | 6.00 | 9.62 |
| TableStatistics | 0.38 0.02 | 0.00 0.00 | 0.00 0.00 | 0.13 0.01 | 12.38 | 0.27 0.01 | 0.52 0.02 | 0.21 0.02 | 0.33 0.02 | 5.25 | 0.64 0.01 | 0.31 0.02 | 12.00 | 9.88 |
| StatisticalSummary | 0.81 0.01 | 0.01 0.01 | 0.01 0.01 | 0.28 0.01 | 4.25 | 0.31 0.01 | 0.59 0.01 | 0.24 0.01 | 0.38 0.01 | 4.00 | 0.91 0.01 | 0.40 0.03 | 4.50 | 4.25 |
| MatrixFactorization | 0.70 0.03 | 0.00 0.00 | 0.00 0.00 | 0.23 0.01 | 8.25 | 0.30 0.02 | 0.50 0.01 | 0.15 0.02 | 0.32 0.02 | 5.75 | 0.78 0.02 | 0.39 0.03 | 7.50 | 7.17 |
| Embedder | Triplet (Stat) | Clustering (Stat) | Retrieval/Probe (Stat) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.96 0.01 | 0.80 0.02 | 0.80 0.02 | 0.86 0.02 | 1.00 | 0.96 0.01 | 0.55 0.03 | 0.42 0.04 | 0.64 0.02 | 1.25 | 1.00 0.00 | 1.00 0.00 | 1.00 | 1.08 |
| SchemaContent | 0.02 0.00 | 0.00 0.00 | 0.00 0.00 | 0.01 0.00 | 14.00 | 0.78 0.02 | 0.07 0.05 | 0.05 0.03 | 0.30 0.03 | 12.50 | 1.00 0.00 | 0.99 0.00 | 2.50 | 9.67 |
| TableStatistics | 0.24 0.03 | 0.07 0.02 | 0.07 0.02 | 0.13 0.02 | 9.75 | 0.77 0.03 | 0.18 0.02 | -0.06 0.04 | 0.30 0.02 | 11.75 | 0.99 0.00 | 0.99 0.00 | 5.00 | 8.83 |
| StatisticalSummary | 0.86 0.03 | 0.45 0.03 | 0.47 0.04 | 0.59 0.03 | 3.00 | 0.93 0.01 | 0.44 0.03 | 0.38 0.05 | 0.59 0.02 | 3.25 | 0.99 0.00 | 0.99 0.00 | 6.00 | 4.08 |
| MatrixFactorization | 0.44 0.03 | 0.12 0.02 | 0.12 0.02 | 0.23 0.03 | 6.25 | 0.79 0.02 | 0.14 0.02 | -0.02 0.02 | 0.30 0.01 | 10.50 | 0.97 0.01 | 0.92 0.01 | 11.50 | 9.42 |
| Embedder | Triplet (Avg over Labels) | Clustering (Avg over Labels) | Retrieval/Probe (Avg over Labels) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.93 0.01 | 0.33 0.01 | 0.33 0.01 | 0.53 0.01 | 1.50 | 0.56 0.02 | 0.67 0.02 | 0.39 0.03 | 0.54 0.02 | 1.50 | 0.71 0.01 | 0.61 0.02 | 8.50 | 3.83 |
| SchemaContent | 0.02 0.00 | 0.00 0.00 | 0.00 0.00 | 0.01 0.00 | 14.00 | 0.37 0.02 | 0.31 0.03 | 0.10 0.02 | 0.26 0.02 | 10.50 | 0.78 0.01 | 0.60 0.02 | 6.50 | 10.33 |
| TableStatistics | 0.35 0.02 | 0.02 0.00 | 0.02 0.00 | 0.13 0.01 | 10.75 | 0.42 0.02 | 0.45 0.02 | 0.16 0.02 | 0.34 0.02 | 5.75 | 0.71 0.01 | 0.50 0.02 | 10.00 | 8.83 |
| StatisticalSummary | 0.83 0.02 | 0.12 0.01 | 0.13 0.02 | 0.36 0.01 | 3.75 | 0.48 0.02 | 0.55 0.02 | 0.29 0.02 | 0.44 0.01 | 4.00 | 0.91 0.01 | 0.56 0.02 | 4.50 | 4.08 |
| MatrixFactorization | 0.64 0.03 | 0.03 0.01 | 0.03 0.01 | 0.23 0.01 | 6.50 | 0.43 0.02 | 0.41 0.02 | 0.12 0.02 | 0.32 0.02 | 6.75 | 0.79 0.02 | 0.49 0.02 | 9.00 | 7.42 |
| Embedder | Triplet (Direct) | Clustering (Direct) | Retrieval/Probe (Direct) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.95 0.01 | 0.26 0.03 | 0.27 0.03 | 0.49 0.02 | 3.00 | 0.12 0.00 | 0.49 0.01 | 0.05 0.00 | 0.22 0.01 | 3.50 | 0.92 0.02 | 0.73 0.03 | 2.50 | 3.00 |
| SchemaContent | 0.35 0.02 | 0.00 0.00 | 0.00 0.00 | 0.12 0.01 | 13.12 | 0.08 0.00 | 0.32 0.02 | 0.02 0.00 | 0.14 0.01 | 11.50 | 0.64 0.01 | 0.71 0.03 | 5.00 | 9.88 |
| TableStatistics | 0.29 0.02 | 0.00 0.00 | 0.00 0.00 | 0.10 0.01 | 13.88 | 0.11 0.01 | 0.45 0.01 | 0.07 0.00 | 0.21 0.01 | 4.75 | 0.59 0.02 | 0.31 0.03 | 8.50 | 9.04 |
| StatisticalSummary | 0.82 0.01 | 0.05 0.01 | 0.06 0.01 | 0.31 0.01 | 5.00 | 0.11 0.01 | 0.45 0.02 | 0.07 0.01 | 0.21 0.01 | 3.50 | 0.74 0.02 | 0.29 0.02 | 8.00 | 5.50 |
| MatrixFactorization | 0.71 0.01 | 0.00 0.00 | 0.00 0.00 | 0.24 0.00 | 9.50 | 0.10 0.01 | 0.38 0.01 | 0.04 0.00 | 0.18 0.01 | 9.25 | 0.45 0.02 | 0.09 0.01 | 12.00 | 10.25 |
| Embedder | Triplet (Semantic) | Clustering (Semantic) | Retrieval/Probe (Semantic) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.95 0.01 | 0.45 0.03 | 0.46 0.03 | 0.62 0.02 | 3.00 | 0.21 0.01 | 0.52 0.01 | 0.09 0.01 | 0.27 0.01 | 4.00 | 0.97 0.01 | 0.97 0.01 | 2.50 | 3.17 |
| SchemaContent | 0.36 0.02 | 0.00 0.00 | 0.00 0.00 | 0.12 0.01 | 13.12 | 0.13 0.01 | 0.32 0.02 | 0.04 0.01 | 0.16 0.01 | 11.00 | 0.76 0.02 | 0.94 0.02 | 5.50 | 9.88 |
| TableStatistics | 0.28 0.02 | 0.00 0.00 | 0.00 0.00 | 0.09 0.01 | 13.88 | 0.18 0.01 | 0.47 0.01 | 0.11 0.01 | 0.25 0.01 | 5.25 | 0.71 0.02 | 0.47 0.03 | 8.50 | 9.21 |
| StatisticalSummary | 0.83 0.01 | 0.12 0.01 | 0.12 0.01 | 0.36 0.01 | 5.00 | 0.20 0.01 | 0.48 0.02 | 0.12 0.02 | 0.27 0.01 | 3.50 | 0.83 0.01 | 0.42 0.02 | 8.00 | 5.50 |
| MatrixFactorization | 0.71 0.01 | 0.00 0.00 | 0.00 0.00 | 0.24 0.00 | 9.50 | 0.17 0.01 | 0.40 0.01 | 0.07 0.01 | 0.21 0.01 | 9.25 | 0.55 0.02 | 0.12 0.02 | 12.50 | 10.42 |
| Embedder | Triplet (Stat) | Clustering (Stat) | Retrieval/Probe (Stat) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.78 0.04 | 0.57 0.05 | 0.60 0.05 | 0.65 0.04 | 1.00 | 0.91 0.03 | 0.58 0.09 | 0.67 0.10 | 0.72 0.07 | 1.00 | 1.00 0.00 | 0.84 0.02 | 4.50 | 2.17 |
| SchemaContent | 0.24 0.02 | 0.00 0.01 | 0.01 0.01 | 0.09 0.01 | 11.50 | 0.62 0.03 | 0.04 0.03 | -0.02 0.01 | 0.21 0.01 | 12.75 | 1.00 0.00 | 0.96 0.01 | 2.00 | 8.75 |
| TableStatistics | 0.17 0.02 | 0.00 0.00 | 0.00 0.00 | 0.06 0.01 | 14.00 | 0.62 0.02 | 0.08 0.03 | 0.03 0.02 | 0.25 0.02 | 9.00 | 0.98 0.00 | 0.99 0.00 | 5.00 | 9.33 |
| StatisticalSummary | 0.67 0.03 | 0.06 0.02 | 0.06 0.02 | 0.26 0.02 | 3.25 | 0.78 0.02 | 0.23 0.06 | 0.31 0.06 | 0.44 0.05 | 3.00 | 0.99 0.00 | 0.90 0.02 | 5.50 | 3.92 |
| MatrixFactorization | 0.47 0.01 | 0.00 0.00 | 0.00 0.00 | 0.16 0.00 | 9.00 | 0.64 0.02 | 0.01 0.01 | 0.01 0.01 | 0.22 0.01 | 11.00 | 0.96 0.00 | 0.65 0.02 | 13.50 | 11.17 |
| Embedder | Triplet (Avg over Labels) | Clustering (Avg over Labels) | Retrieval/Probe (Avg over Labels) | Overall | ||||||||||
| TR-R | TR-H | TR-CH | TR-Avg | Rank | Purity | NMI | ARI | CL-Avg | Rank | R@5 | LP | Rank | Avg Rank | |
| HashingSchema | 0.89 0.02 | 0.43 0.04 | 0.44 0.03 | 0.59 0.03 | 1.00 | 0.41 0.02 | 0.53 0.04 | 0.27 0.04 | 0.40 0.03 | 1.00 | 0.96 0.01 | 0.84 0.02 | 3.50 | 1.83 |
| SchemaContent | 0.32 0.02 | 0.00 0.00 | 0.00 0.00 | 0.11 0.01 | 12.50 | 0.28 0.01 | 0.23 0.02 | 0.01 0.01 | 0.17 0.01 | 13.50 | 0.80 0.01 | 0.87 0.02 | 4.00 | 10.00 |
| TableStatistics | 0.25 0.02 | 0.00 0.00 | 0.00 0.00 | 0.08 0.01 | 14.00 | 0.30 0.01 | 0.34 0.02 | 0.07 0.01 | 0.24 0.01 | 6.00 | 0.76 0.01 | 0.59 0.02 | 9.00 | 9.67 |
| StatisticalSummary | 0.77 0.02 | 0.08 0.01 | 0.08 0.01 | 0.31 0.01 | 4.75 | 0.37 0.01 | 0.39 0.03 | 0.16 0.03 | 0.31 0.02 | 3.25 | 0.85 0.01 | 0.54 0.02 | 8.00 | 5.33 |
| MatrixFactorization | 0.63 0.01 | 0.00 0.00 | 0.00 0.00 | 0.21 0.00 | 9.50 | 0.30 0.01 | 0.26 0.01 | 0.04 0.01 | 0.20 0.01 | 10.00 | 0.65 0.01 | 0.28 0.02 | 12.50 | 10.67 |
| Phase | Data source | Tables / partitions | Partition policy | Notes |
| Training | Real (training) | 16 tables per batch 8 partitions per table | Row/column fractions: Minimum partition: Overlap ratio: | Train the head on top of a fixed encoder IoU-based SOFTCE listwise objective Same-table support defined by Direct identity |
| Validation | Real (validation) | 20 tables 5 partitions per table | Row/column fractions sampled from Minimum partition: Overlap ratio: | Evaluator-based val_score Mean of D1 Spearman and averaged D2 sections Computed on the validation pass Used to select checkpoints and seeds |
| Benchmark | Real + synthetic (evaluation) | 100 tables 10 partitions per table | Row/column fractions sampled from Minimum partition: Overlap ratio: | Full D1–D3 benchmark reporting Evaluation-time partitions Synthetic results reflect transfer from heads trained on real tables |
| Head | D1 Consistency | D2 Label-Based - Avg | D3 Robustness | |||||
| Spearman | TR-Avg | CL-Avg | R@5 | Perm | Noise | Mask | ||
| ConTextTab (pooled) | – | 0.04 0.03 | 0.18 0.00 | 0.27 0.02 | 0.70 0.01 | 0.97 0.00 | 0.97 0.00 | 0.97 0.00 |
| Attention | 0.00 | 0.44 0.02 | 0.35 0.01 | 0.46 0.03 | 0.87 0.01 | 1.00 0.00 | 0.99 0.00 | 0.98 0.00 |
| 0.25 | 0.46 0.02 | 0.35 0.01 | 0.44 0.01 | 0.87 0.01 | 1.00 0.00 | 0.99 0.00 | 0.99 0.00 | |
| 0.50 | 0.45 0.02 | 0.34 0.01 | 0.47 0.02 | 0.88 0.01 | 1.00 0.00 | 1.00 0.00 | 0.99 0.00 | |
| 0.75 | 0.50 0.02 | 0.34 0.01 | 0.46 0.02 | 0.88 0.01 | 1.00 0.00 | 1.00 0.00 | 0.99 0.00 | |
| Head | D1 Consistency | D2 Label-Based - Avg | D3 Robustness | |||||
| Spearman | TR-Avg | CL-Avg | R@5 | Perm | Noise | Mask | ||
| ConTextTab (pooled) | – | 0.19 0.03 | 0.25 0.01 | 0.22 0.01 | 0.76 0.01 | 0.97 0.00 | 0.97 0.00 | 0.97 0.00 |
| Attention | 0.00 | 0.26 0.03 | 0.40 0.02 | 0.29 0.03 | 0.96 0.01 | 1.00 0.00 | 0.99 0.01 | 0.98 0.00 |
| 0.25 | 0.27 0.03 | 0.41 0.02 | 0.30 0.03 | 0.96 0.01 | 1.00 0.00 | 0.99 0.00 | 0.98 0.00 | |
| 0.50 | 0.25 0.03 | 0.42 0.02 | 0.31 0.04 | 0.96 0.00 | 1.00 0.00 | 0.99 0.00 | 0.98 0.00 | |
| 0.75 | 0.23 0.03 | 0.37 0.02 | 0.27 0.02 | 0.95 0.01 | 1.00 0.00 | 0.99 0.00 | 0.99 0.00 | |
| Embedder | D1 Consistency | D2 Label-Based - Avg | D3 Robustness | Overall | |||||||
| Spearman | Rank | TR-Avg | CL-Avg | R@5 | Rank | Perm | Noise | Mask | Rank | Avg Rank | |
| ConTextTab -R Attention ( ) | 0.50 | 3.00 | 0.34 | 0.46 | 0.88 | 5.00 | 1.00 | 1.00 | 0.99 | 5.67 | 4.56 |
| HashingSchema | 0.59 | 1.00 | 0.53 | 0.54 | 0.71 | 5.33 | 1.00 | 0.98 | 0.99 | 7.67 | 4.67 |
| ConTextTab -R GatedAttention ( ) | 0.48 | 4.50 | 0.31 | 0.46 | 0.89 | 4.67 | 1.00 | 1.00 | 0.99 | 5.67 | 4.94 |
| HashingText | 0.54 | 2.00 | 0.47 | 0.46 | 1.00 | 2.67 | 1.00 | 0.65 | 0.97 | 10.33 | 5.00 |
| ConTextTab -R GatedMLP ( ) | 0.48 | 4.50 | 0.30 | 0.43 | 0.88 | 7.00 | 1.00 | 1.00 | 0.99 | 5.67 | 5.72 |
| Embedder | D1 Consistency | D2 Label-Based - Avg | D3 Robustness | Overall | |||||||
| Spearman | Rank | TR-Avg | CL-Avg | R@5 | Rank | Perm | Noise | Mask | Rank | Avg Rank | |
| HashingText | 0.41 | 1.00 | 0.51 | 0.28 | 0.98 | 3.00 | 1.00 | 0.73 | 0.98 | 9.33 | 4.44 |
| ConTextTab -R Attention ( ) | 0.27 | 3.00 | 0.41 | 0.30 | 0.96 | 3.83 | 1.00 | 0.99 | 0.98 | 6.83 | 4.56 |
| HashingSchema | 0.31 | 2.00 | 0.59 | 0.40 | 0.96 | 2.00 | 1.00 | 0.94 | 0.95 | 10.17 | 4.72 |
| ConTextTab -R GatedAttention ( ) | 0.26 | 4.50 | 0.38 | 0.27 | 0.95 | 5.67 | 1.00 | 0.99 | 0.99 | 6.00 | 5.39 |
| ConTextTab -R GatedMLP ( ) | 0.24 | 7.00 | 0.31 | 0.27 | 0.95 | 6.50 | 1.00 | 1.00 | 0.99 | 4.83 | 6.11 |
| Embedder | Numeric-S | Numeric-L | Text-S | Text-L | Mixed-S | Mixed-L | GeoMean Tbl/s | Rank |
| HashingSchema | 9793.76 112.47 | 2618.63 28.77 | 11675.20 77.04 | 3213.25 6.29 | 11694.96 911.74 | 3206.73 56.35 | 5748.43 104.02 | 1.00 |
| SchemaContent | 2293.53 17.80 | 928.61 2.98 | 1653.73 7.43 | 161.07 1.08 | 1765.45 31.58 | 270.29 4.26 | 804.30 3.58 | 2.00 |
| TableStatistics | 1471.32 15.68 | 201.71 4.73 | 1227.06 79.31 | 53.83 1.73 | 1201.78 10.75 | 87.43 3.62 | 356.70 8.17 | 3.00 |
| StatisticalSummary | 529.07 1.23 | 163.42 5.61 | 116.95 0.31 | 13.66 0.17 | 161.30 2.37 | 24.64 0.20 | 90.48 0.73 | 6.00 |
| MatrixFactorization | 368.22 11.62 | 72.43 0.26 | 360.33 9.88 | 73.04 0.91 | 355.93 1.53 | 74.27 0.50 | 162.71 0.96 | 5.00 |
| HashingText | 715.00 1.51 | 6.41 0.05 | 1265.68 1.79 | 12.26 0.04 | 900.99 3.22 | 8.34 0.13 | 90.10 0.34 | 7.00 |
Tabular Representation Learning
Tabular representation learning builds useful embeddings of structured rows, columns, and tables for prediction, retrieval, clustering, and generation.
See all 59 papers