Skip to content

๐Ÿ•ธ๏ธ Graph Learning

๐Ÿ’ฌ EMNLP2026 ยท 10 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (44) ยท ๐ŸŽž๏ธ ECCV2026 (1) ยท ๐Ÿ“ท CVPR2026 (8) ยท ๐Ÿ”ฌ ICLR2026 (118) ยท ๐Ÿ’ฌ ACL2026 (24) ยท ๐Ÿงช ICML2026 (35)

๐Ÿ”ฅ Top topics: Reasoning ร—3 ยท Question Answering ร—2 ยท LLM ร—2

Dual-Prompt-Driven Textual Knowledge Graph Generation With Heterogeneous Information Embedding for CTR Prediction

Addressing the severe semantic loss of discrete feature encodings and the lack of fine-grained cross-modal alignment in CTR prediction, TKE4CTR designs a dual-prompt paradigm to transform tabular records into textual narratives and hallucination-mitigated local knowledge subgraphs, followed by bidirectional cross-attention gated fusion and a feature-gated MLP that achieve substantial AUC improvements across five diverse Amazon domains.

FBV-KGQA: Forward-Backward Verification for Knowledge Graph Question Answering

FBV-KGQA proposes a forward-backward verification framework for knowledge graph question answering that generates relation-level plans, quantifies the causal attribution of critical entity/relation pivots via schema-constrained counterfactual perturbations, and checks reverse question consistency to filter out superficially plausible candidate paths.

GraphLit: Learning Text-Enriched Dynamic Character Network Representations for Literary Study

GraphLit introduces Dynamic Heterogeneous Character Networks (DHCNs) to segment long novels into temporally localized heterogeneous graphs that align characters with their contextual text segments, training an 8.2M-parameter masked graph autoencoder with block ordering and global link prediction to systematically outperform text-only, graph-only, and prior hybrid baselines across 12 character-centric tasks and quotation attribution.

LLM-Powered Holistic Enhancement for Text-Attributed Graphs

Proposes EFS (Enhancing Features and Structures), a unified framework that enriches node features using LLM-extracted keywords, ranked category predictions, and explanatory contexts, while refining the graph topology with a lightweight edge refiner without requiring any per-edge LLM API calls, achieving state-of-the-art node classification across six benchmark datasets.

Logic Meets Language: Hybrid Query Answering over Incomplete Text-Labeled Graphs

Addressing the challenge of answering queries requiring both discrete logical constraints and flexible textual semantics over incomplete text-labeled graphs, this paper formalizes a hybrid query language and introduces HybridQE, an end-to-end framework that unifies query-conditioned entity representations with neural logical operators in a shared embedding space.

Ontology-Driven Structural Regularization for Document-Level Relation Extraction

Addressing pervasive and previously overlooked structural noise in distantly supervised Document-Level Relation Extraction (DocRE)โ€”such as ontology constraint violations, missing inverse relations, and asymmetric contradictionsโ€”this paper formalizes structural well-formedness under OWL axioms and demonstrates that lightweight, model-agnostic data-centric structural regularization drastically eliminates logical contradictions in model predictions while consistently improving generalization performance.

PathSoft: Hybrid Soft Prompting with Ranked Reasoning Paths for Knowledge Graph Question Answering

PathSoft introduces a hybrid soft prompting framework for knowledge graph question answering that filters candidate reasoning chains with a lightweight PathMLP and encodes the ordered entityโ€“relation sequences into continuous soft prompts via PathSoftAdapter, delivering state-of-the-art accuracy on frozen LLMs with single-call, ultra-low latency inference.

StruProKGR: A Structural and Probabilistic Framework for Sparse Knowledge Graph Reasoning

To tackle computational bottlenecks in random walk path collection and the isolated treatment of paths in sparse knowledge graph reasoning, StruProKGR introduces a training-free framework that leverages truncated distance-guided DFS pruning for up to 54.93x path collection speedup and incorporates intra-path diminishing returns and inter-path Bayesian odds updates to capture structural dependencies, consistently outperforming existing path-based and rule-based baselines across five sparse benchmarks.

TATK: Triple-Aware Top-K Learning with Knowledge-Grounded Verification for LLM-based Sequential Recommendation

Addressing the key mismatches in LLM-based sequential recommendationโ€”weak context grounding, misalignment between generation loss and ranking metrics, and unverified candidate listsโ€”TATK introduces a Triple-Aware framework that combines metadata-KG prompt grounding and position-aware Top-K rewards during training with single-pass knowledge-grounded verification at inference time, substantially boosting full-catalog Top-K ranking performance.

TKG-Thinker: Towards Dynamic Reasoning over Temporal Knowledge Graphs via Agentic Reinforcement Learning

TKG-Thinker addresses reasoning hallucinations and rigid static workflows in temporal knowledge graph question answering (TKGQA) by formulating reasoning as a sequential decision process, equipping LLMs with temporal tool primitives via a two-stage optimization pipeline combining rejection-sampled SFT cold-start and online reinforcement learning with multi-dimensional verifiable rewards.