PaperlayerView exact revision on arXiv

Source-linked AI summary

Revisiting Simple Baselines for Relational Deep Learning

Viet The Nguyen, Vijay Prakash Dwivedi, Gunther Gust

arXiv:2610.20534v1cs.LG

TL;DR

The effectiveness of simple relational deep learning baselines remains underexplored, limiting assessment of how much architectural complexity is needed. This paper adapts fixed feature propagation and Correct-and-Smooth to relational prediction, finding that feature propagation improves most RelBench tasks and brings classification performance close to a supervised heterogeneous GNN. It also identifies task-specific, temporally valid meta-graph construction as a challenge for label propagation.

  • Problem

    Simple propagation baselines are not directly applicable to relational entity graphs, which are heterogeneous and temporal and have labels only for task target entities.

    Source: §1 · p. 2

  • Method

    The method trains an MLP on target attributes, enriches features with fixed temporal neighborhood propagation, and optionally refines predictions with Correct-and-Smooth using a task-specific meta-graph.

    Source: §1 · p. 2, §2.1 · p. 2, §2.2 · p. 2

  • Results

    Feature propagation improves 20 of 21 tasks and brings average classification AUROC within two points of the supervised heterogeneous GNN baseline.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Fixed feature propagation recovers substantial relational signal without learned aggregation, providing a simple baseline for assessing gains from more complex RDL models.

    Source: §4 · p. 4

    The relational schema does not specify which connections support label propagation, motivating task-aware meta-graphs that respect temporal visibility.

    Source: §4 · p. 4

Abstract

from arXiv · show

Relational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Transformers have achieved strong performance across diverse domains. However, the effectiveness of simpler relational baselines remains underexplored, making it difficult to assess how much architectural complexity is needed to achieve strong predictive performance. Motivated by the observation that predictive signals arise from target attributes, relational attributes, and historical labels, we revisit how effectively these signals can be exploited using simple propagation methods. We adapt fixed low-pass feature propagation to relational entity graphs and train an MLP for prediction. The resulting predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that incorporates historical labels over task-specific meta-graphs connecting target entities. Across classification and regression tasks on RelBench, an MLP trained on neighborhood-filtered features outperforms one using only target-entity features on most tasks and recovers a substantial portion of the performance of heterogeneous GNNs. Correct-and-Smooth provides further gains from historical labels, although applying it requires a separately constructed meta-graph because connections among target entities are not directly specified by the relational entity graph. Together, our studies revisit the effectiveness of established propagation methods as simple RDL baselines while identifying canonical meta-graph construction as a key challenge for effective label propagation.

1 Introduction

Relational deep learning models tables as temporal heterogeneous graphs, but the effectiveness of simpler baselines remains underexplored. This work adapts feature and label propagation to relational tasks, finding that fixed feature propagation improves performance on most RelBench tasks and bringing classification AUROC close to a supervised heterogeneous GNN.

1 Introduction

  • Relational deep learning represents table rows as nodes in temporal heterogeneous graphs, with edges encoding primary–foreign key relationships.GNNs and Transformer-based models aggregate information across related entities [3] [4] [6] [7] [8].Source: §1 · p. 1
  • Entity-level predictive signals include target attributes, related-entity attributes, and historical labels.Existing models aggregate temporal-neighborhood features, while historical labels can add signal to zero-shot predictions [6].Source: §1 · p. 2
  • Simple MLP, fixed-propagation, and Correct-and-Smooth methods are not directly applicable to relational entity graphs because they are heterogeneous, temporal, and label-sparse.The database-schema graph may also be unsuitable for task-label propagation.Source: §1 · p. 2
  • Feature propagation improves 20 of 21 tasks and brings average classification AUROC within two points of the supervised heterogeneous GNN baseline.The authors also identify canonical, temporally valid meta-graph construction as a key challenge for label propagation.Source: §1 · p. 2

2 Simple Propagation for Relational Deep Learning

The method separates relational prediction into fixed feature propagation followed by optional label-based correction and smoothing. It respects temporal visibility, using a separate task-specific meta-graph for historical labels because target entities may be connected only indirectly in the relational graph.

2 Simple Propagation for Relational Deep Learning

  • The method incorporates three sources of predictive signal into the baseline using feature and label propagation; Appendices B and C provide the feature construction, filter definition, and C&S updates.Source: §2 · p. 2

2.1 Relational Entity Graphs

  • The relational entity graph has table rows as nodes and primary–foreign key links as edges; task predictions may use only information available by each seed time.The graph is used for fixed feature propagation before MLP training, rather than learned relation-specific aggregation [4].Source: §2.1 · p. 2

2.2 Feature Propagation

  • For each seed time, the method removes later-timestamped rows, applies a fixed low-pass filter with symmetric degree normalization, and trains an MLP on the resulting features.The main experiments use K = 2 propagation steps; K = 0 uses target attributes alone.Source: §2.2 · p. 2

2.3 Label Propagation with Correct-and-Smooth

  • Correct-and-Smooth propagates labeled-node residuals to adjust MLP predictions, then smooths predictions using observed labels as anchors.Because target entities may be connected only indirectly in the relational graph, label propagation uses a separate meta-graph over prediction rows and historical labels, admitting only labels observed before the seed time.Source: §2.3 · pp. 2–3

3 Experiments

The evaluation compares target-only and neighborhood-filtered MLPs, then applies Correct-and-Smooth to filtered predictions across RelBench classification and regression tasks. Feature propagation improves all reported classification tasks and most regression tasks, while substantially reducing trainable parameters relative to HeteroGNN.

3 Experiments

  • The evaluation spans RelBench-v1 and RelBench-v2 entity-level classification and regression tasks, averaging results over ten random seeds and comparing target-attribute MLP, K=2 filtered MLP, filtered MLP + C&S, and HeteroGNN.Classification uses AUROC and regression uses MAE; rel-f1 and rel-event are excluded because temporal leakage has been reported [6] [13].Source: §3 · p. 3

3.1 Task Performance

  • All 12 classification tasks improve with feature propagation, raising average AUROC from 55.95 to 73.65 on RelBench-v1 and from 65.93 to 82.38 on RelBench-v2.The filtered MLP comes within 1.8 and 1.6 points of HeteroGNN, respectively; C&S adds about 0.51 and 0.02 average points.Source: §3.1 · pp. 3–4
  • Feature propagation lowers regression MAE on eight of nine tasks and matches or exceeds HeteroGNN on avito/ad-ctr and arxiv/author-publication.The exception is trial/site-success, where MAE rises from 0.426 to 0.437; C&S provides almost no regression improvement.Source: §3.1 · p. 4

3.2 Parameter Reduction

  • Training only the MLPs reduces trainable parameters relative to HeteroGNN because feature propagation and C&S use fixed operators.Figure 2 reports trainable parameters and MLP parameter reduction relative to HeteroGNN.Source: §3.2 · p. 4, Figure 2 · p. 4

4 Discussion

Fixed feature propagation recovers substantial relational signal without learned aggregation, providing a simple baseline for assessing more complex RDL models. Historical-label propagation remains more challenging because the schema does not specify which connections support label propagation.

4 Discussion

  • Fixed feature propagation recovers substantial relational signal without learned aggregation, offering a simple baseline for assessing gains from more complex RDL models.The paper identifies task-aware, temporally valid meta-graphs and more efficient seed-time filtering as needs for further work.Source: §4 · p. 4
  • The database schema connects related entities but does not specify which connections support task-label propagation.The discussion calls for task-aware constructions that capture useful label dependencies while respecting temporal visibility.Source: §4 · p. 4
  • Filtering each seed-time snapshot also motivates more efficient preprocessing.This is raised alongside the need for task-aware meta-graphs.Source: §4 · p. 4

A Benchmark Tasks and Constructed Graphs

The section describes database-level graph and feature representations, then outlines task splits and evaluation scope. Tasks share a database graph but require snapshots for their distinct seed times.

Databases

  • Database representations vary substantially in graph rows, feature dimensions, feature-type distributions, and resulting MLP size.rel-avito has 622 input dimensions, while rel-trial has 15,145, mainly from text, and an MLP roughly 24× larger.Source: A Benchmark Tasks and Constructed Graphs · p. 5, Table 4 · p. 6
  • The graph and feature representation is constructed once per database, independently of the task evaluated on it.Tasks defined on the same database share the underlying graph, but require separate snapshots.Source: A Benchmark Tasks and Constructed Graphs · pp. 5–6

Tasks

  • Tasks use chronological train, validation, and test splits, and each distinct seed time requires a filtered snapshot; rel-event and rel-f1 tasks are excluded.The exclusions reflect reported temporal leakage for rel-event and a likely compromised status for rel-f1.Source: A Benchmark Tasks and Constructed Graphs · p. 6

B Feature Propagation and Training Details

The method encodes table attributes into a shared feature matrix, adds snapshot-specific structural features, and applies fixed propagation over the temporally visible relational graph before MLP prediction.

B Feature Propagation and Training Details

  • Numerical, text, categorical, and timestamp columns receive distinct encodings, while primary and foreign keys define graph structure rather than model features.Text uses 300-dimensional GloVe sentence embeddings; categorical columns use indicators for six frequent values and the row value’s log-frequency.Source: B Feature Propagation and Training… · p. 7
  • Table feature matrices occupy disjoint row and column blocks, leaving each entity’s attributes in its source-table block and zeros elsewhere.The stacked representation is formed across the database’s tables.Source: B Feature Propagation and Training… · p. 7
  • Omitting row normalization and symmetric adjacency reduced test AUROC by up to 41 percentage points.For nonzero, non-negative features, the normalization agrees with signed-sum normalization; zero rows remain zero and empty column blocks do not affect the norm.Source: B Feature Propagation and Training… · p. 7
  • After row normalization, two structural features are appended and recomputed for each visible snapshot: degree excluding self-loops and timestamp age capped at 50 years.Rows without timestamps receive age zero; appending after normalization preserves the structural features’ absolute scale.Source: B Feature Propagation and Training… · p. 7

B.1 Filter Definition

  • The filter uses the symmetric adjacency induced by rows visible at seed time, adds self-loops, and propagates normalized attributes with snapshot-specific structural features.Setting K = 0 and both structural columns to zero yields the target-attribute baseline.Source: §B.1 · p. 7

B.2 Graph Filtering Cost

Filtering is precomputed by task and snapshot, with cost shaped by seed-time count, feature dimensionality, and dense snapshot storage.

B.2 Graph Filtering Cost

  • Filtering is precomputed once per distinct task seed time, so its cost grows with the number of seed times.The collapsed graph and stacked features are shared by tasks on a database, while each task computes snapshots for its seed times.Source: §B.2 · p. 8
  • Filtering took 5.5 minutes on rel-avito and 47 minutes on rel-stack for one representative task per database; on rel-avito, deepening the filter to K = 10 increased time to 113 minutes.Source: §B.2 · p. 8
  • Dense design matrices dominate memory, with serial peak memory ranging from 44 GB on rel-avito to 214 GB on rel-amazon.The timing comparison also reports MLP training and one run of HeteroGNN training.Source: §B.2 · p. 8, Table 5 · p. 8

B.3 Training Details

The experiments use a shared MLP training setup and compare it with a pipeline that precomputes filtered features. Correct-and-Smooth applies tuned label-propagation stages to base predictions over a task-entity graph.

B.3 Training Details

  • The MLP has one 256-unit hidden layer with ReLU and 0.5 dropout; classification uses binary cross-entropy, while regression uses ℓ1 loss with predictions clipped to training-target percentiles 2–98.Batch normalization is removed because training-period running statistics can degrade under temporal test shift; on one task, test AUROC changed from 0.886 to 0.224 without that removal.Source: §B.3 · p. 8
  • Results use official RelBench splits without modification and average ten fixed seeds controlling MLP initialization and batch shuffling.The relational filter is deterministic on the entity graph.Source: §B.3 · p. 9
  • RelBench-style models sample temporal neighborhoods and jointly train the relational model and prediction head, whereas this pipeline filters the visible graph once per seed time and reuses the features.Filtering moves relational computation into preprocessing.Source: §B.3 · p. 9

C Label Propagation with Correct-and-Smooth

  • Correct-and-Smooth post-processes base predictions in two label-propagation stages over a graph concerning the task entities.Correct initializes labeled-node residuals as observed labels minus base predictions, with zeros elsewhere.Source: C Label Propagation with… · p. 9
  • The Smooth stage initializes labeled nodes with observed labels and other nodes with corrected predictions; both stages assume positively correlated errors and labels along edges.The configuration, including normalizations, scaling, and stage selection, is tuned on validation; seeds without labeled neighbors retain base predictions.Source: C Label Propagation with… · p. 9

C.1 Graph Constructions

The paper constructs task-specific meta-graphs to propagate historical labels from the target and related entities. Variants differ in which histories they include and whether labels attach directly, through relays, or along temporal chains.

C.1 Graph Constructions

  • Related entities are found by composing foreign-key walks back to the task table, and up to 100 highest-weight entities are selected using records observed before the seed time.Shared-record weights are inversely scaled by the number of entities sharing each record.Source: §C.1 · pp. 9–10
  • Every variant excludes labels observed at or after the seed time, so construction differences concern eligible histories and their connections to the prediction row.The figure distinguishes direct connections from two-hop relay connections and temporal chains.Source: §C.1 · p. 10, Figure 3 · p. 11
  • SELF connects the target to its past labels with age-decreasing weights, CROSS uses each selected related entity’s newest label, and CROSS+SELF combines both sources.CROSS+SELF uses unit weights for own-label edges and co-occurrence weights for related-label edges.Source: §C.1 · pp. 10, 12
  • CROSS+HIST uses a relay carrying a related entity’s base prediction and up to three newest labels, while ALL adds relays for the target’s full history.Each target-history label gets a separate relay, with copies of up to three newest own labels attached.Source: §C.1 · p. 12
  • SELF-CHAIN places the target’s full history on a newest-first path, and CHAIN+CROSS adds a weighted direct connection to each selected related entity’s newest label.Older own labels are farther from the prediction row; related entities contribute one label at one hop.Source: §C.1 · p. 12
  • CHAIN+HIST-3 forms newest-first paths for the target and related entities with up to three labels each, while CHAIN+HIST removes that limit.Related paths begin with co-occurrence weights, and subsequent edges have weight 1.Source: §C.1 · p. 12

D Experimental Results

Label-propagation results depend on meta-graph construction and task: no construction consistently wins, and selected variants can help some regression tasks while failing to improve others. The extended results also examine model comparisons, structural features, filtering depth, and runtime.

D Experimental Results

  • Meta-graph gains on the filtered base are generally small and construction-dependent, with no consistent winner; the self-label star lowers avito/user-clicks AUROC by 2.7 points.Related-entity labels improve trial/study-outcome AUROC by up to +1.67 points, while stack/user-engagement favors the temporal chain.Source: D Experimental Results · p. 12, Table 7 · p. 13
  • The best meta-graph varies by task, so the mean gain from selecting each task’s best variant is an oracle-style upper bound, while any fixed variant gains less.For regression, validation-selected meta-graphs lower MAE on six of seven tasks; own-history graphs give the largest reported gains on hm/item-sales and amazon/item-ltv.Source: D Experimental Results · p. 12, Table 8 · p. 13, Table 9 · p. 13
  • On stack/post-votes and arxiv/author-publication, all nine variants stay within 0.14% of the filtered MLP, and mean-residual corrections raise validation MAE on both tasks.The passage attributes the mismatch between correlated history and error reduction to low-concentrated targets and MAE’s preference for conditional medians.Source: D Experimental Results · p. 13, Table 9 · p. 13

E.1 Comparisons with More Models

  • Table 10 compares the pipeline with RelGNN and RDBLearn to contextualize how much signal a simple baseline captures without learned relational aggregation.The passage notes that these models outperform the pipeline on several tasks; RelGNN reports no RelBench-v2 results.Source: §E.1 · p. 14, Table 10 · p. 14

E.2 Results without Structural Columns

  • Removing degree and age features leaves the core result intact: relational attributes improve over K = 0 baselines, and C&S can still improve performance.Table 11 compares filtered representations with and without those structural features.Source: §E.2 · p. 14, Table 11 · p. 16

E.3.1 Filtering Choices

  • K = 2 is best on most tasks, and on ten classification tasks with results at every tested depth it is best or tied on five.Deeper filters gain at most 0.21 AUROC points on amazon/user-churn but lose up to 7.1 points on trial/study-outcome, while costing more.Source: §E.3.1 · p. 14, Figure 4 · p. 15

E.3.2 Runtime

  • Filtering can become sharply more expensive as neighborhoods expand, rising on rel-avito from about 5.5 minutes at K = 2 to 60–77 minutes at K = 4.Although outputs contain only target rows, other visible nodes still contribute through intermediate propagation.Source: §E.3.2 · p. 16

E.4 Ablation Study with Labels as Lagged Features

The ablation compares MLPs with self-history and cross-entity label features, then applies C&S to the label-augmented model. History features help on some tasks, and C&S adds gains on five of eight tasks; the discussion identifies canonical label-propagation graph construction as an open challenge.

E.4 Ablation Study with Labels as Lagged Features

  • Twelve features summarize each entity’s visible history for MLP inputs; five cover availability, size, recency, and label values, five are rank-recency-weighted averages, and two use elapsed time.Source: §E.4 · pp. 16–17
  • The ablation compares a base MLP, one with self-history features, and one with both self-history and cross-entity label features, then applies C&S to the latter.Cross-entity features admit only labels observed strictly before prediction time; rows without visible history use an absence indicator and default summary values.Source: §E.4 · p. 17
  • C&S adds improvements on five of eight tasks, including +1.76 AUROC on trial/study-outcome, where target entities have no self-history by construction.Across tasks, gains vary, and avito/user-clicks shows a small degradation.Source: §E.4 · p. 17, Table 12 · p. 17

E.5 Discussion and Open Questions

  • Relational schemas do not specify which entity connections support effective label propagation, so canonical graph construction remains an open challenge.Shared records create structural links without guaranteeing correlated labels.Source: §E.5 · p. 17

Related papers

Loading 2610.20534v1…