How InterpGN Bridges Interpretability and Accuracy in Time Series Classification

Background

Time‑series classification is a core task in many domains (e.g., healthcare, neuroscience, automation). Deep neural networks (DNNs) achieve state‑of‑the‑art accuracy but provide little insight into their decisions, which is problematic for safety‑critical applications. Shapelet‑based models are intrinsically interpretable because they use discriminative subsequences (shapelets) as logical predicates, yet they usually lag behind DNNs in predictive performance.

Problem

The goal is to obtain a model that retains the transparent reasoning of shapelet‑based experts while matching the predictive strength of modern deep learning classifiers.

InterpGN Architecture

InterpGN (Interpretability‑Gated Networks) combines a shapelet‑based interpretable expert with a DNN inside a mixture‑of‑experts (MoE) framework. A confidence‑driven gating function decides, for each sample, whether to rely solely on the expert (preserving interpretability) or to blend the expert’s output with the DNN (handling complex patterns).

Shapelet‑based Interpretable Expert

The expert extracts a set of K shapelets S_k (each of length L ) from the training series. For a given time‑series X, the Euclidean distance between S_k and every subsequence of X is computed. Instead of a hard threshold, InterpGN uses a radial‑basis‑function (RBF) predicate to convert distances into probabilities: p_k(X) = exp(-\epsilon \cdot d^2(S_k, X_{sub})) where d is the Euclidean distance, \epsilon controls kernel steepness, and X_{sub} is the best‑matching subsequence. The resulting probabilities form a logical feature vector z = [p_1, …, p_K].

Shapelet Bottleneck Model (SBM)

SBM receives the predicate vector z and applies a linear classifier: ŷ = softmax(W z + b) The parameters W and

b</b> are learned jointly with the shapelets. The overall expert loss combines three terms:</p><ul><li>Cross‑entropy loss <code>L_ce = -∑_c y_c log ŷ_c

Shapelet‑diversity regularizer L_div that penalises redundant shapelets (e.g., encouraging orthogonal distance vectors) L1 regularisation L_reg = λ_reg ||W||_1 to promote sparsity and select the most informative predicates The total expert loss is L_expert = L_ce + λ_div L_div + λ_reg L_reg .

Confidence‑driven Gating

The gating function evaluates the expert’s confidence using a modified Gini index on the predicate vector z : g(z) = 1 - ∑_k p_k^2 If g(z) < τ (confidence above threshold τ ), the sample is classified by SBM alone. Otherwise, the DNN’s prediction ŷ_DNN is blended with the expert’s output: ŷ = α ŷ_expert + (1-α) ŷ_DNN The blending weight α is derived from the gating confidence.

Training Objective

The full MoE loss mixes the expert loss and the MoE cross‑entropy loss: Loss = β(t)·L_expert + (1-β(t))·L_MoE The weight β(t) can be a fixed hyper‑parameter or follow a cosine‑annealing schedule, encouraging the expert to keep learning useful shapelets throughout training. The entire system is end‑to‑end differentiable and optimized with stochastic gradient descent (e.g., Adam).

Experiments

UEA Multivariate Time‑Series Benchmark

InterpGN was evaluated on 30 datasets from the UEA archive. It achieved an average accuracy of 0.760 , an average rank of 3.5 , topped 8 datasets, and placed in the top‑3 on 16 datasets, outperforming several recent state‑of‑the‑art baselines.

MIMIC‑III In‑Hospital Mortality Prediction

Using the clinical MIMIC‑III dataset (48‑hour mortality prediction), InterpGN obtained:

Accuracy: 0.703

F1: 0.657

Recall: 0.569

Precision: 0.784

ROC‑AUC: 0.703

Local explanations highlighted shapelets that correspond to clinically relevant patterns (e.g., heart‑rate spikes, mean arterial pressure trends), demonstrating the model’s interpretability in a real‑world setting.

Ablation and Hyper‑parameter Studies

Key findings from systematic ablations:

Increasing the number of shapelets K improves expressive power but may reduce interpretability if too many predicates are active.

Cosine‑annealed β yields better focus on useful expert predictions.

RBF predicates consistently outperform linear‑threshold predicates in both accuracy and the quality of learned shapelets.

Kernel steepness ε, regularisation weights λ_reg and λ_div affect the trade‑off between sparsity and diversity.

Extended Experiments

Variants were tested with alternative DNN experts (FCN, Transformer, PatchTST, TimesNet), different shapelet distance metrics (cosine similarity, Pearson correlation), and alternative SBM classifiers (bilinear, attention‑based). Some variants gave marginal gains on specific datasets but did not surpass the baseline InterpGN across the full benchmark.

Limitations

In certain classes the most influential shapelet may not correspond to an intuitively meaningful subsequence, potentially misleading users. The current design uses a single interpretable expert and a single DNN; extending the MoE to multiple experts could increase flexibility and robustness.

Conclusion

InterpGN demonstrates that a confidence‑driven gated mixture can balance interpretability and predictive performance for time‑series classification. The framework delivers high‑quality shapelets, transparent local/global explanations, and competitive accuracy on both benchmark and clinical datasets.

Resources

Paper: Shedding Light on Time Series Classification using Interpretability Gated Networks – https://openreview.net/forum?id=n34taxF0TC Code repository: https://github.com/YunshiWen/InterpretGatedNetwork