Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Granularity Open Intent Classification via Adaptive Granular-Ball Decision Boundary
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 MOGB, which uses adaptive granular-ball clustering and a nearest sub-centroid classifier to learn multi-granularity decision boundaries for open intent classification.
Problem
Open intent classification requires separating known intents from unknown ones, but prior boundary-based methods assume a compact spherical distribution per known intent, which is often violated and leaves intra-class open space risk.
Approach
Split each known intent class into sub-parts via adaptive granular-ball clustering (no preset cluster number), train with a nearest sub-centroid cosine classifier iteratively with the clustering, and classify using per-granular-ball radii as multi-granularity decision boundaries.
Key results
- On StackOverflow at 25% known classes, MOGB reaches 91.48 Acc and 84.43 F1-All, versus 86.72 and 80.83 for ADB.
- On BANKING at 75% known classes, MOGB reaches 82.69 Acc and 86.82 F1-All, above ADB's 81.08 and 85.96.
- Ablation at 25% known on SNIPS: removing hierarchical representation learning or multi-granularity boundaries lowers Acc from 68.00 to 62.57.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MOGB instead of the strongest reported baseline | Open intent classification · StackOverflow, 25%, 50%, and 75% known classes | F1-All: 80.83 -> 84.43; 85.83 -> 87.49; 85.99 -> 87.37 | Helped |
| Use MOGB instead of the strongest reported baseline | Open intent classification · SNIPS, 25%, 50%, and 75% known classes | F1-All: 65.72 -> 72.20; 78.82 -> 78.91; 88.61 -> 89.39 | Helped |
| Use MOGB instead of the strongest reported baseline | Open intent classification · BANKING, 25%, 50%, and 75% known classes | F1-All: 71.62 -> 75.19; 81.03 -> 81.52; 86.01 -> 86.82 | Helped |
| Add hierarchical representation learning (ablation) | MOGB · StackOverflow, SNIPS, and BANKING at 25% known classes | F1-All: 79.72 -> 84.43; 68.94 -> 72.20; 71.55 -> 75.19 | Helped |
| Add multi-granularity decision boundaries (ablation) | MOGB · StackOverflow, SNIPS, and BANKING at 25% known classes | F1-All: 75.82 -> 84.43; 68.88 -> 72.20; 74.25 -> 75.19 | Helped |
| Add both hierarchical representation learning and multi-granularity decision boundaries (ablation) | MOGB · StackOverflow, SNIPS, and BANKING at 25% known classes | F1-All: 80.06 -> 84.43; 67.54 -> 72.20; 72.20 -> 75.19 | Helped |
| Set purity threshold p_t to 0.80 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 86.41 vs 87.41 | Hurt |
| Set purity threshold p_t to 0.85 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 86.50 vs 87.41 | Hurt |
| Set purity threshold p_t to 0.90 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 87.37 vs 87.41 | Hurt |
| Set purity threshold p_t to 0.93 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 87.37 -> 87.41 | Helped |
| Set purity threshold p_t to 0.95 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 87.41 -> 87.31 | Hurt |
| Set purity threshold p_t to 0.97 (ablation) | MOGB · StackOverflow, 75% known classes | F1-All: 87.41 -> 87.31 | Hurt |
Surprising findings
- Only 57% of known intents in CLINC-FULL conform to the Gaussian assumption, per cited Gaussian hypothesis testing.
- Gains over ADB are larger at low known-class ratios and small at 75% known classes.
Limitations
- Performance depends on granular-ball quality controlled by purity and sample count thresholds (p_t, n_t), which must be tuned per dataset.
- Only the final BERT layer is fine-tuned; the paper does not study full fine-tuning.
- Evaluated only on three English intent datasets.
Experimental setup
Datasets: StackOverflow SNIPS BANKING
Benchmarks: StackOverflow SNIPS BANKING
Metrics: Accuracy F1-All F1-U (unknown) F1-K (known)
Baselines: MSP DOC OpenMax DeepUnk ADB DA-ADB
Builds on: ADB granular-ball computing BERT
Closest papers by meaning
- MIntRec2.0: A Large-scale Benchmark Dataset for Multimodal Intent Recognition and Out-of-scope Detection in Conversations (ICLR 2024 · similarity 0.87)
- Towards Multi-Intent Spoken Language Understanding via Hierarchical Attention and Optimal Transport (AAAI 2024 · similarity 0.83)
- Uni-MIS: United Multiple Intent Spoken Language Understanding via Multi-View Intent-Slot Interaction (AAAI 2024 · similarity 0.82)
- Aligner²: Enhancing Joint Multiple Intent Detection and Slot Filling via Adjustive and Forced Cross-Task Alignment (AAAI 2024 · similarity 0.82)
- Evolutionary Multimodal Reasoning via Hierarchical Semantic Representation for Intent Recognition (CVPR 2026 · similarity 0.81)
- Divide-Solve-Combine: An Interpretable and Accurate Prompting Framework for Zero-shot Multi-Intent Detection (AAAI 2025 · similarity 0.81)
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.