SuperMAN: Interpretable and Expressive Networks over Temporally Sparse Heterogeneous Data
Maya Bechler-Speicher, Andrea Zerio, Maor Huri, Marie Vibeke Vestergaard, Ran Gilad-Bachrach, Tine Jess, Samir Bhatt, Aleksejs Sazonovs
TL;DR
SuperMAN is a novel interpretable neural network framework designed to learn from irregular, heterogeneous temporal data by modeling signals as implicit graphs, achieving state-of-the-art results in healthcare and social media tasks.
Contribution
It introduces SuperMAN, a new interpretable model that handles sparse, heterogeneous temporal data as implicit graphs, with flexible interpretability and high performance.
Findings
State-of-the-art accuracy in predicting Crohn's disease and hospital stays.
Effective detection of fake news using heterogeneous signals.
Interpretability aids in understanding disease progression.
Abstract
Real-world temporal data often consists of multiple signal types recorded at irregular, asynchronous intervals. For instance, in the medical domain, different types of blood tests can be measured at different times and frequencies, resulting in fragmented and unevenly scattered temporal data. Similar issues of irregular sampling occur in other domains, such as the monitoring of large systems using event log files. Effectively learning from such data requires handling sets of temporal sparse and heterogeneous signals. In this work, we propose Super Mixing Additive Networks (SuperMAN), a novel and interpretable-by-design framework for learning directly from such heterogeneous signals, by modeling them as sets of implicit graphs. SuperMAN provides diverse interpretability capabilities, including node-level, graph-level, and subset-level importance, and enables practitioners to trade…
Peer Reviews
Decision·ICLR 2026 Poster
1. It handles irregular, heterogeneous data: GMAN introduces a novel way to learn from sets of sparse, irregular time-series signals without any resampling or imputation. By converting each signal into a graph (e.g. linking events by time), the model preserves timing gaps and irregular patterns that other methods might obscure. This design avoids the information loss incurred by aligning or filling in missing data, enabling the model to exploit the full richness of the raw data (e.g. varying int
1. While GMAN’s grouping mechanism is powerful, it requires a priori decisions about how to partition features or signals. The model’s performance can depend on choosing sensible groups – a process that may need domain expertise or extensive tuning. In the experiments, the authors manually tried several grouping schemes (including no grouping vs. clinically guided groupings) and selected the best performing one. This indicates an added complexity: users must either have prior knowledge to guide
* Well-motivated representation of irregular, multi-signal data: Modeling each signal trajectory as a path graph and learning over a set of such graphs avoids time-gridding/imputation artifacts; the paper also justifies recoverability of structure under a tree-metric condition and gives a clear O($m·K·n²·d\psi$) complexity with GPU-friendly implementation notes. * Theory for interpretability–expressivity trade-off: ExtGNAN enables multivariate within-graph feature grouping; the authors prove GMA
* Novelty relative to GNAN could be made crisper in the writing: While theorems show strict expressivity gains, the narrative at times reads like an incremental extension of GNAN; emphasize the two-level mixing (within-graph feature grouping and across-graph subset mixing) and the new recoverability/complexity results as core contributions. * Scalability: ExtGNAN performs dense pairwise aggregation across nodes, giving per-graph O(n²) cost and overall O(m·K·n²·dψ); this may limit very long traj
1. Addresses an under-explored irregular-sampling problem. 2. Good design combining additive interpretability with graph-set flexibility. 3. Compelling real-world demonstrations (Crohn’s disease, LOS prediction).
1. Evaluation section briefly described; lacks detailed baselines beyond GNAN. 2. Interpretability examples are qualitative; quantitative faithfulness metrics missing. 3. Computational efficiency compared with standard GNNs unclear.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Bayesian Modeling and Causal Inference · Domain Adaptation and Few-Shot Learning