Research map / Time series / Causal and generative time series

SiBBlInGS: Similarity-driven Building-Block Inference using Graphs across States

ICML 2024 · Time series › Causal and generative time series · approach: Robust forecasting and evaluation · 1 citations (OpenAlex)

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 SiBBlInGS, a graph-regularized dictionary learning framework that extracts sparse building blocks and their per-trial temporal traces across multiple states, with cross-state similarity controlled by a state graph and channel graphs.

Problem

Identifying interpretable, sparse building blocks from multi-state, multi-trial time-series data where trials vary in duration, sampling, and missing samples, and where building-block composition may adapt subtly across states.

Approach

Extends re-weighted l1 graph filtering dictionary learning with channel-similarity graphs (per state) and a state-similarity graph P that regularizes building-block matrices across states, with per-BB weights nu allowing both state-invariant and state-specific BBs; trial temporal traces are updated independently per trial.

Key results

• On synthetic data (N=100, p=10, D=3), SiBBlInGS recovers ground-truth building blocks and traces with correlations close to one and outperforms listed baselines.
• On the monkey reaching data, all identified BBs show within/between state correlation ratio greater than 1, and multi-class logistic regression on the traces accurately predicts the reach direction state.
• On Google Trends, the identified BBs recover meaningful clusters (college, Passover, winter, COVID) with state-specific temporal patterns, and other methods with p=5 produced less meaningful BBs (qualitative).

Measured effects reported in the paper

Surprising findings

• Hopkins shows a less pronounced COVID-related search peak in Maryland, where the university and hospital are located, attributed to local familiarity.

Limitations

• Assumes Gaussian noise statistics whereas Poisson may be more suitable
• Linear mapping from building blocks to observations
• Does not handle missing channels or cross-trial channel identity changes, nor directional connectivity

Experimental setup

Datasets: synthetic data Google Trends monkey somatosensory cortex reaching task recordings EEG recordings of epileptic patient

Benchmarks: synthetic data ground truth recovery Google Trends monkey somatosensory reaching data epilepsy EEG

Metrics: Jaccard index correlation with ground truth within/between state correlation ratio logistic regression state prediction accuracy

Baselines: PARAFAC Tucker decomposition PCA (global and local) sparse PCA NONFAT NNDTN mTDR dPCA

Builds on: re-weighted l1 graph filtering (RWL1-GF) dictionary learning PARAFAC demixed PCA targeted dimensionality reduction

Closest papers by meaning

• Learning Causal Relations from Subsampled Time Series with Two Time-Slices (ICML 2024 · similarity 0.84)
• Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.83)
• CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.83)
• Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.83)
• Benchmarking Structural Inference Methods for Interacting Dynamical Systems with Synthetic Data (NeurIPS 2024 · similarity 0.83)
• Less Greedy Equivalence Search (NeurIPS 2025 · similarity 0.83)