Source-linked AI summary
Anatomy of a Decision: Uncertainty-aware Hierarchical Intent Learning via Flow Matching for Multimodal Recommendation
Yuchen Miao, Zijun Wang, Ke Liu, Siyang Xu
TL;DR
Multimodal intent modeling must account for uncertainty in the data and the changing granularity of user interests. UHIFlow uses conditional flow matching to estimate and align modality uncertainties, then uses them to build personalized intent hierarchies. It achieves the best performance across datasets and metrics, improving over its strongest baseline by 2.81%–3.04%.
Problem
Existing multimodal intent methods overlook uncertainty in user behavior and modality features and use static structures that cannot adapt to users’ varying decision certainty.
Source: §1 · p. 2Method
UHIFlow uses conditional flow matching to quantify and synergize visual and textual uncertainty, which guides personalized coarse-to-fine intent hierarchies.
Source: §1 · pp. 2–3Results
UHIFlow achieves the best performance across all datasets and metrics, improving over the strongest baseline (MENTOR, AAAI’25) by 2.81%–3.04%.
Source: §5.2 · p. 9Takeaways & Limitations
Jointly modeling multimodal uncertainty and adaptive hierarchical intent provides a stable signal for uncertain cases to use coarse intents and confident cases to refine toward item-level granularity.
Source: §5.2 · p. 9The paper leaves sample-efficient uncertainty estimation from limited-feedback signals to future work.
Source: §6 · p. 12
Abstract
from arXiv · showhide
Modeling the underlying user intent is crucial for recommendation, but existing methods struggle with the inherent uncertainty and the dynamic, hierarchical nature of user interests. Current approaches often rely on clustering or prototype learning to discover a static set of intents. However, they face two critical challenges: (1) they overlook the uncertainty inherent in multimodal features; and (2) they assume a static and flat intent structure, failing to adapt to a user's varying decision certainty. To address these limitations, we propose UHIFlow, an Uncertainty-aware Hierarchical Intent learning framework via Flow matching. First, our Cross-modal Uncertainty Synergistic Modeling (CUSM) module leverages conditional flow matching to quantify uncertainty from visual and textual modalities and synergistically align them. Subsequently, the Uncertainty-guided Hierarchical Intent Generation (UHIG) module uses this quantified uncertainty to dynamically construct a personalized intent hierarchy, generating coarse-grained intents for uncertain users and fine-grained ones for users with clear preferences. Extensive experiments on three real-world datasets demonstrate that UHIFlow significantly outperforms state-of-the-art baselines.
1 Introduction
The introduction identifies two gaps in multimodal intent modeling: uncertainty is overlooked, and intent structures are static rather than adaptive. UHIFlow uses modality uncertainty to construct personalized intent hierarchies.
1 Introduction
- Existing multimodal intent methods treat features as deterministic, despite uncertainty from ambiguous behavior and noisy or semantically ambiguous modalities.The introduction notes that this can yield suboptimal intent representations.Source: §1 · p. 2
- Fixed, flat intent structures cannot reflect users’ coarse-to-fine decisions, from broad interests under high uncertainty to finer preferences under low uncertainty.The figure illustrates this contrast with broad category and precise-item levels.Source: §1 · p. 2, Figure 1 · p. 2
- UHIFlow’s CUSM uses conditional flow matching to quantify visual and textual uncertainty and let the modalities inform and regularize one another.The framework treats item features as distributions rather than fixed points.Source: §1 · pp. 2–3
- UHIG uses quantified uncertainty to construct personalized hierarchies, with coarse intents for uncertain users and fine-grained intents for users with clear preferences.This adapts the hierarchy to each user’s decision certainty.Source: §1 · pp. 2–3
- UHIFlow outperforms state-of-the-art baselines in experiments on real-world datasets.The introduction reports the overall result without giving metric values.Source: §1 · p. 3
2 Related Work
Related work spans multimodal recommendation, intent modeling, generative models, and uncertainty modeling. The cited approaches leave open the combination of multimodal uncertainty estimation with intent hierarchies that adapt to user certainty.
2.1 Multimodal Recommendation
- Multimodal recommendation methods use visual features and cross-modal or graph-based learning, but rarely quantify modality uncertainty or adapt intent hierarchy.Examples include VBPR, MMSSL, DiffCL, LGMRec, DiffMM, and multimodal intent methods [28].Source: §2.1 · p. 3
2.2 Intent Modeling in Recommender Systems
- Intent methods include clustering, multi-interest, disentangled, and hierarchical approaches, yet most retain fixed intent counts or static structures.The passage cites MIND, HIM, HieRec [26], and other approaches.Source: §2.2 · p. 3
2.3 Generative Models for Recommendation
- Diffusion and flow-based recommendation mainly address interaction modeling, augmentation, denoising, or generation rather than uncertainty-aware hierarchical intent learning.Flow matching learns complex distributions without restrictive priors, and FlowCF applies it to collaborative filtering.Source: §2.3 · p. 3
2.4 Uncertainty Modeling in Recommendation
- Uncertainty methods address settings such as sparsity, distribution shift, and cold start, but seldom combine multimodal uncertainty with adaptive hierarchical intent generation.Examples include UCC, UA-GNN, and WaPOIR.Source: §2.4 · p. 4
3 Preliminary
The preliminary setup represents users, items, interactions, and visual and textual item features. Its objective is to learn representations for personalized ranking.
3 Preliminary
- The setup defines user and item sets and represents their interactions in a binary matrix.An entry is 1 when a user interacted with an item and 0 otherwise.Source: §3 · p. 4
- Each item has a visual feature vector and a textual feature vector.The vectors are denoted v_i and t_i, respectively.Source: §3 · p. 4
- The objective is to learn user and item representations for predicting personalized ranked lists.The recommendation score is introduced as part of this prediction setup.Source: §3 · p. 4
4 Methodology
UHIFlow builds multimodal user and item representations, estimates visual and textual uncertainty with conditional flows, and uses uncertainty to generate and aggregate personalized hierarchical intents before recommendation.
4.1 Multimodal Interest Modeling Layer
- Graph propagation over user–item interactions refines multimodal item and user embeddings, which feed the subsequent uncertainty and intent modeling.Item features are formed by projecting concatenated visual and textual inputs; embeddings aggregate representations across graph layers.Source: User–Item Interaction Graph · p. 4, Collaborative Interest Propagation · p. 4
4.2 Cross-modal Uncertainty Synergistic Modeling (CUSM)
- Conditional flows model visual and textual item features as distributions, while cross-modal guidance aligns intermediate representations and density evolution supplies an uncertainty signal.The visual and textual flows are conditioned on the collaborative item embedding; modality uncertainty is quantified from log-density change.Source: §4.2 · pp. 5–6
4.3 Uncertainty-guided Hierarchical Intent Generation (UHIG)
- Residual quantization decomposes each user embedding into increasingly fine-grained intent codes, with hierarchy depth controlled by that user's uncertainty.High uncertainty yields larger stopping thresholds and shallower, coarser hierarchies; low uncertainty yields deeper, finer ones.Source: §4.3 · p. 6
4.4 Uncertainty-aware Hierarchical Intent Aggregation
- Uncertainty-gated attention aggregates a user's hierarchy while penalizing finer intents more strongly for users with greater uncertainty, then fuses the intent and collaborative embeddings.The penalty term is exp(−γ·l·ςu), where γ is a scaling factor.Source: §4.4 · p. 7
4.5 Prediction and Optimization
- The model trains recommendation with a BPR objective over user, interacted-item, and unobserved-item triplets, alongside module-specific auxiliary losses.Two auxiliary losses are weighted by hyperparameters in the final objective.Source: §4.5 · p. 7
5 Experiments
Across three multimodal benchmarks, UHIFlow outperforms the strongest baseline, and the experiments examine its components, uncertainty strategies, hyperparameter sensitivity, and computational efficiency.
5.1 Experimental Settings
- Experiments use three 5-core-filtered multimodal benchmarks, evaluate Recall@K and NDCG@K for K={10, 20}, and average results over five random seeds.The benchmarks use pre-extracted visual and textual features.Source: Datasets · p. 8, Evaluation Metrics · p. 8, Implementation Details · p. 8
5.2 Overall Performance Comparison
- UHIFlow improves over MENTOR by 2.81%–3.04% and achieves the best performance across all datasets and metrics.The reported improvements are significant at p < 0.05.Source: Table 2 · p. 9, §5.2 · p. 9
- Advanced multimodal, intent-based, and generative methods outperform traditional collaborative filtering, while UHIFlow exceeds recent multimodal competitors on every metric.The paper attributes remaining gaps to fixed or flat intent structures and uncertainty methods focused on denoising rather than residual uncertainty quantification.Source: §5.2 · p. 9
5.3 Ablation Study
- Removing any component degrades performance; the largest drops occur when collaborative interest propagation or hierarchical intent generation is removed.CUSM supplies the per-user uncertainty UHIG uses to set hierarchy depth; removing CUSM therefore collapses the hierarchy to fixed granularity.Source: §5.3 · p. 10
5.4 Impact of Uncertainty Modeling Strategy
- Flow Matching consistently outperforms Gaussian, Wasserstein, and energy-based uncertainty strategies in the reported comparison.The paper describes Gaussian and Wasserstein approaches as relying on restrictive assumptions and EBMs as unstable to train.Source: §5.4 · p. 11
5.5 Parameter Sensitivity Analysis
- Recall@20 first rises and then falls as each of the four loss weights increases, while performance remains robust over a reasonable range such as µ1, µ2 ∈[0.1, 0.5].Figure 3 varies uncertainty loss, intent loss, cross-modal guidance, and intent diversity weights on Baby, Sports, and TikTok.Source: §5.5 · p. 11, Figure 3 · p. 11
5.6 Efficiency Analysis
- UHIFlow’s flow matching adds moderate training overhead and keeps inference speed close to the fastest baselines, unlike diffusion methods with the highest training and inference costs.The efficiency comparison uses an NVIDIA A100; the paper says expensive flow computation is performed offline during training.Source: Figure 4 · p. 12, §5.6 · p. 12
6 Conclusion
UHIFlow uses multimodal uncertainty to adapt personalized intent-structure depth and demonstrates superiority over state-of-the-art baselines. Future work will explore sample-efficient uncertainty estimation from limited-feedback signals.
6 Conclusion
- UHIFlow demonstrates superiority over state-of-the-art baselines in extensive experiments.Source: §6 · p. 12
- CUSM uses conditional flow matching to quantify and synergize multimodal uncertainties, which adaptively control UHIG’s personalized fine-to-coarse intent structures.Source: §6 · p. 12
- Future work will explore sample-efficient uncertainty estimation from limited-feedback signals.Source: §6 · p. 12