Research map / Graph learning / Knowledge and temporal graphs
A Goal Interaction Graph Planning Framework for Conversational Recommendation
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.
Proposes GIGF, a goal interaction graph planning framework that builds a directed heterogeneous graph over goal types and entities, encodes dialogue/knowledge/profile, and uses a dual-way fusion decoder to plan next goals, then generates responses with GPT-2.
Problem
Planning goal sequences in multi-goal conversational recommender systems (MG-CRS) so the dialogue naturally leads users to accept a final recommended goal; prior methods ignore goal-to-goal correlations and goal order, or model the two goal levels separately.
Approach
Model goal types and entities jointly in one directed heterogeneous graph (built from sequential goal co-occurrence in the corpus and domain knowledge triples), propagate representations with a heterogeneous graph attention network, and predict the next goal type and entity with a dual-way cross-attention fusion decoder over dialogue context, user profile and knowledge.
Key results
- On DuRecDial, goal entity accuracy 93.47% vs 91.92% for TPNet and goal type accuracy 97.41% vs 95.46% for MGNN.
- On DuRecDial 2.0, goal entity accuracy 96.01% vs 90.56% for MGNN, and goal type F1 97.91%.
- Response generation F1 on DuRecDial rises to 47.52% vs 43.50% for MGNN; on DuRecDial 2.0 F1 38.35% vs 36.75% for MGNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the goal interaction graph instead of removing graph learning (ablation) | Goal planning framework · DuRecDial | goal entity F1: 66.75 -> 82.75 | Helped |
| use the goal interaction graph instead of removing graph learning (ablation) | Goal planning framework · DuRecDial 2.0 | goal entity F1: 81.72 -> 89.46 | Helped |
| use heterogeneous graph attention rather than GAT on the goal-entity graph (ablation) | Goal planning framework · DuRecDial | goal entity F1: 73.02 -> 82.75 | Helped |
| use heterogeneous graph attention rather than GAT on the goal-entity graph (ablation) | Goal planning framework · DuRecDial 2.0 | goal entity F1: 85.73 -> 89.46 | Helped |
| use the knowledge fusion gate (ablation) | Goal planning framework · DuRecDial | goal entity F1: 77.84 -> 82.75 | Helped |
| use the knowledge fusion gate (ablation) | Goal planning framework · DuRecDial 2.0 | goal entity F1: 82.44 -> 89.46 | Helped |
| use knowledge attention in the goal entity decoder (ablation) | Goal planning framework · DuRecDial | goal entity F1: 73.06 -> 82.75 | Helped |
| use knowledge attention in the goal entity decoder (ablation) | Goal planning framework · DuRecDial 2.0 | goal entity F1: 84.47 -> 89.46 | Helped |
| guide response generation with predicted goals rather than omit goal guidance (ablation) | GPT-2 response generation · DuRecDial | response F1: 34.93 -> 47.52 | Helped |
| guide response generation with predicted goals rather than omit goal guidance (ablation) | GPT-2 response generation · DuRecDial 2.0 | response F1: 31.58 -> 38.35 | Helped |
Surprising findings
- Removing the goal interaction graph causes the largest drop in goal entity F1 (66.75 vs 82.75 on DuRecDial), and a plain goal-entity-only GAT recovers only part of that gap.
Limitations
- Goal planning and response generation are separate modules; response generation uses plain GPT-2 without joint training.
- Evaluated only on two Chinese/English DuRecDial variants; generalization to other MG-CRS datasets not tested.
- Goal entity F1 on DuRecDial (82.75) is only modestly above TPNet and MGNN, and some methods were reported without the goal-type gain.
Experimental setup
Datasets: DuRecDial DuRecDial 2.0
Benchmarks: DuRecDial DuRecDial 2.0
Metrics: Accuracy macro F1 word-level F1 BLEU-1 BLEU-2 DIST-2 Knowledge F1 perplexity
Baselines: MGCG UniMIND TCP MGNN TPNet
Builds on: MGNN MGCG TCP TPNet UniMIND
Closest papers by meaning
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.84)
- ESPRESSO: An Effective Approach to Passage Retrieval for High-Quality Conversational Recommender Systems (AAAI 2025 · similarity 0.84)
- Goal-Conditioned Supervised Learning for Multi-Objective Recommendation (NeurIPS 2026 · similarity 0.83)
- Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations (AAAI 2024 · similarity 0.83)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.83)
- LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.82)
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.