CIKM 2026 Full Research Paper

BRIDGE: Behavior-Guided Residual Integration with Dual-Frequency Graph Evidence

Let multimodal evidence retrieve plausible items. Let behavior evidence decide their local order.

Zesheng Li Chengchang Pan^* Honggang Qi^*

University of the Chinese Academy of Sciences * Corresponding authors

Paper Code

Core finding

Global score correction can distort the catalog. Restricting behavior-guided residuals to the base top-K candidates improves ranking while preserving the backbone.

Overview

Recommendation evidence should have a scope.

Multimodal recommendation combines visual, textual, and collaborative signals, but stronger cross-view alignment does not always improve ranking. Our diagnostics show an effective range: moderate alignment helps, while stronger alignment suppresses recommendation-specific variation.

BRIDGE separates retrieval from local correction. A dual-frequency graph encoder builds the base ranking, training-only co-user overlap becomes signed behavior evidence, and the residual is applied only inside the base candidate set.

Method

Retrieve broadly. Calibrate locally.

BRIDGE keeps multimodal representation learning and behavior-guided score correction in distinct, testable roles.

BRIDGE architecture with multimodal graph encoding, dual-frequency base ranking, and behavior-guided candidate calibration — BRIDGE first constructs the base ranking with DFGE, then uses BEN evidence to authorize CRI corrections only within the selected candidate set.

DFGE

Graph-smoothed ID, image, and text channels are separated into representation-spectral bands and fused into the base ranking geometry.

BEN

Training-only co-user overlap is aggregated over each user history and normalized into signed candidate evidence.

CRI

The evidence-weighted residual changes scores only inside the detached base top-K set during both training and inference.

Results

Consistent gains across three Amazon domains.

All methods use the same processed splits, BEIT3 features, and full-sort protocol. Competitive results are averaged over five seeds.

Recall@20

0.1128

+3.6% over the best baseline

NDCG@20

0.0525

+8.5% over the best baseline

A smaller but stable gain in the densest of the three evaluated domains.

Main comparison against the strongest baseline in each domain.
Dataset	Best baseline R@20	BRIDGE R@20	Gain	Best baseline N@20	BRIDGE N@20	Gain
Baby	0.1089	0.1128	+3.6%	0.0484	0.0525	+8.5%
Sports	0.1187	0.1262	+6.3%	0.0538	0.0594	+10.4%
Electronics	0.0725	0.0778	+7.3%	0.0335	0.0385	+14.9%

Mechanism analysis

What actually drives the improvement?

The controls support a qualified conclusion: candidate restriction and behavior evidence do most of the work; spectral routing provides a smaller complementary gain.

BRIDGEFull model

0.1128

No frequency decompositionEqual-capacity routing

0.1110

Global correctionResidual reaches the catalog

0.1092

No top-K calibrationCandidate scope removed

0.1025

No behavior evidenceBEN removed

0.1016

Cross-view agreement is highest in the low band, while ranking signal is stronger in the complete and high-band representations — Low-frequency evidence is more consistent across views, while later bands retain more private ranking signal.

Behavior evidence increases the candidate coefficient, score correction, and local rank promotion — Higher BEN evidence opens CRI, increases the local correction, and promotes supported candidates.

BRIDGE lowers head-item exposure while improving recall across head, mid, tail, and cold item strata — BRIDGE improves recall across popularity strata while reducing concentration on head items.

Reproducibility

A controlled evaluation path.

Datasets: 3 Amazon domains
Main results: 5 random seeds
Reference hardware: 1 x RTX 4090, 24 GB
Behavior graph: Training split only

View training instructions

Citation

Reference BRIDGE.

Proceedings metadata: ACM DOI 10.1145/3799682.3840961.

@inproceedings{li2026bridge,
  title     = {BRIDGE: Behavior-Guided Residual Integration with
               Dual-Frequency Graph Evidence},
  author    = {Li, Zesheng and Pan, Chengchang and Qi, Honggang},
  booktitle = {Proceedings of the 35th ACM International Conference on
               Information and Knowledge Management},
  year      = {2026},
  publisher = {ACM},
  doi       = {10.1145/3799682.3840961},
  isbn      = {979-8-4007-2539-5/2026/11}
}