[Paper Review] DisenHCN: Disentangled Hypergraph Convolutional Networks for Spatiotemporal Activity Prediction

Yinfeng Li, Chen Gao+4 more|arXiv (Cornell University)|Aug 14, 2022

Human Mobility and Location-Based Analysis4 citations

TL;DR

This paper proposes DisenHCN, a disentangled hypergraph convolutional network that models spatiotemporal activity prediction by constructing a heterogeneous hypergraph to capture fine-grained user similarities and complex user-activity matchings. It disentangles user embeddings into location-, time-, and activity-aware subspaces and uses efficient hypergraph convolutions to learn high-order relations, achieving 14.23% to 18.10% performance gains over SOTA methods on four real-world datasets.

ABSTRACT

Spatiotemporal activity prediction, aiming to predict user activities at a specific location and time, is crucial for applications like urban planning and mobile advertising. Existing solutions based on tensor decomposition or graph embedding suffer from the following two major limitations: 1) ignoring the fine-grained similarities of user preferences; 2) user's modeling is entangled. In this work, we propose a hypergraph neural network model called DisenHCN to bridge the above gaps. In particular, we first unify the fine-grained user similarity and the complex matching between user preferences and spatiotemporal activity into a heterogeneous hypergraph. We then disentangle the user representations into different aspects (location-aware, time-aware, and activity-aware) and aggregate corresponding aspect's features on the constructed hypergraph, capturing high-order relations from different aspects and disentangles the impact of each aspect for final prediction. Extensive experiments show that our DisenHCN outperforms the state-of-the-art methods by 14.23% to 18.10% on four real-world datasets. Further studies also convincingly verify the rationality of each component in our DisenHCN.

Motivation & Objective

• Address the limitations of existing methods in modeling fine-grained user similarities and entangled user preferences in spatiotemporal activity prediction.
• Overcome the inability of tensor decomposition and graph-based models to capture multi-grained, multi-type user similarities and one-to-many user-activity interactions.
• Decouple user preferences into distinct aspects—location, time, and activity—to improve modeling accuracy and interpretability.
• Develop a hypergraph-based framework that explicitly models complex, high-order relationships among users, locations, times, and activities.
• Demonstrate superior performance and robustness on sparse data, particularly in low-activity user regimes.

Proposed method

• Construct a heterogeneous hypergraph with multiple types of hyperedges to encode fine-grained user similarities (e.g., same location, same activity) and complex one-to-many user-spatiotemporal activity matchings.
• Disentangle user embeddings into three distinct subspaces—location-aware, time-aware, and activity-aware—using a disentangled representation learning strategy.
• Design an efficient hypergraph convolution layer (Eff-HGConv) that performs both intra-type and inter-type aggregation across different hyperedge types to preserve aspect-specific features.
• Introduce an independence-modeling module with a regularization coefficient γ to enforce statistical independence among disentangled embedding components.
• Optimize the model end-to-end using a joint loss function that combines prediction loss and disentanglement regularization.
• Utilize a single-layer architecture to avoid over-smoothing, leveraging the hypergraph’s ability to capture high-order relations without deep stacking.

Experimental results

Research questions

• RQ1 Can a heterogeneous hypergraph effectively model fine-grained, multi-type user similarities in spatiotemporal activity prediction?
• RQ2 Does disentangling user embeddings into location-, time-, and activity-aware subspaces improve prediction performance and model interpretability?
• RQ3 How does the proposed hypergraph convolution mechanism compare to standard GCNs in capturing high-order dependencies?
• RQ4 To what extent does DisenHCN mitigate performance degradation under data sparsity compared to existing SOTA methods?
• RQ5 What is the optimal configuration of the disentanglement regularization coefficient γ and number of aggregation layers L for robust performance?

Key findings

• DisenHCN achieves a 14.23% to 18.10% performance gain over state-of-the-art methods on four real-world datasets, demonstrating its superiority in spatiotemporal activity prediction.
• The model maintains stable and strong performance across all datasets, with the optimal γ value being 3e-3 for Telecom, TalkingData, and 4SQ, and 1e-2 for TWEET.
• Performance degrades significantly when γ exceeds 1e-2, indicating that over-regularization harms optimization and generalization.
• The model achieves peak performance with only one aggregation layer, confirming that the hypergraph structure effectively captures high-order relations without deep stacking.
• DisenHCN significantly outperforms MHCN and HWNN on sparse user groups, with performance gaps widening as data sparsity increases, especially for users with fewer than five activity records.
• Ablation studies confirm the effectiveness of each component, including the hypergraph construction, disentanglement module, and Eff-HGConv layers.