ContraLSP: Contrastive Sparse Perturbations Transform Time‑Series Explanation

Opening

Recently, the Alibaba Cloud big‑data engineering team, together with Nanjing University, Penn State, and Tsinghua University, announced that their paper “Explaining Time Series via Contrastive and Locally Sparse Perturbations” was accepted at ICLR 2024. The paper proposes ContraLSP, a perturbation‑based explanation framework that combines a counterfactual perturbation objective with a sparsity‑inducing compressor.

Background

Providing reliable explanations for predictions in finance, gaming, healthcare, and other domains is crucial for ethical and legal compliance. While many methods generate saliency maps for images, text, or graphs, explanations for time‑series models remain under‑explored, and adapting existing explainers to time‑series data is challenging due to their inherent complexity and low interpretability.

Challenges

Existing explanation techniques rely on saliency methods that interact with arbitrary models, using gradients, attention, Shapley values, or LIME. These approaches often produce instance‑level saliency maps but suffer from feature inter‑correlation and generalization errors. Perturbation‑based methods modify data using baselines or generative models, yet the perturbed “non‑salient” regions may be out‑of‑distribution, introducing bias.

Breakthrough

ContraLSP formulates a concrete perturbation that preserves the original label y while minimizing a mask m that highlights important regions. The optimization objective consists of three terms: (1) consistency between the original and perturbed inputs through the black‑box model f, (2) sparsity of the mask, and (3) smoothness of the mask.

(1) Counterfactual perturbations via contrastive learning

ContraLSP learns counterfactual samples by contrastive learning, pulling the perturbed instance toward negative samples and pushing it away from positive ones. A triplet loss is used to enforce this behavior.

(2) Smoothness‑constrained sparse gates

To ensure the mask is both sparse and smooth, a time‑trend‑guided L0 regularizer is applied. The mask is generated through a gated sigmoid function whose temperature is learned from the time trend, producing soft masks that better adapt to temporal data.

Application

ContraLSP has been integrated into Alibaba’s Feitian Big‑Data AI governance platform for time‑series metric drilling and anomaly detection, with future work planned for root‑cause analysis on time‑series data.

Paper title: Explaining Time Series via Contrastive and Locally Sparse Perturbations

Authors: Liu Zichuan, Zhang Yingying, Wang Tianchun, Wang Zefan, Luo Dongsheng, Du Mengnan, Wu Min, Wang Yi, Chen Chunlin, Fan Lunting, Wen Qingsong

Paper link: https://openreview.net/pdf?id=qDdSRaOiyb

Slides: https://github.com/zichuan-liu/ContraLSP/blob/main/intro_contralsp_slides.pdf