How Alibaba’s New Time‑Series Models Are Redefining Forecasting and Anomaly Detection

Pathformer: Adaptive Multi‑Scale Transformer for Time Series Forecasting

The Pathformer model builds on the Pathways architecture to provide a fully‑featured multi‑scale transformer that jointly models time resolution and time distance via a dual‑attention mechanism. It introduces adaptive pathways that route and aggregate multi‑scale features layer‑by‑layer, improving prediction accuracy and generalization across diverse time‑scale patterns in cloud resource workloads.

ContraLSP: Contrastive Sparse Perturbation Framework for Time Series Explanation

ContraLSP addresses the need for reliable explanations in intelligent‑operation scenarios by generating counterfactual samples that create information‑null perturbations while preserving the data distribution. A sample‑specific sparse gating mechanism produces near‑binary masks that succinctly capture temporal trends and highlight salient features. The overall objective optimizes explanation consistency under label constraints.

MACE: Multi‑Pattern Aware Frequency‑Domain Anomaly Detection

MACE extracts normal‑mode subspaces in the frequency domain and maps each sample onto these subspaces. The reconstruction error grows with the distance from the corresponding normal‑mode subspace, enabling precise anomaly scoring. Context‑aware Fourier transform and inverse mechanisms exploit frequency sparsity for high‑throughput implementation, while Peak and Valley convolutions amplify short‑term anomalies.

LARA: Lightweight Data‑Dependent Retraining for Anomaly Detection

LARA tackles the scarcity of early‑stage observations and the high cost of retraining. A “rumination” module reuses historical data similar to new observations, while linear mapping functions M_z and M_x transform old hidden states and reconstructions to current‑distribution estimates. The derived convex loss guarantees fast convergence and avoids over‑fitting.

Paper Links

Pathformer: Multi‑Scale Transformers With Adaptive Pathways For Time Series Forecasting – PDF – Code: GitHub

ContraLSP: Explaining Time Series via Contrastive and Locally Sparse Perturbations – PDF – Code: GitHub

MACE: Learning Multi‑Pattern Normalities in the Frequency Domain for Efficient Time Series Anomaly Detection – arXiv

LARA: A Light and Anti‑overfitting Retraining Approach for Unsupervised Time Series Anomaly Detection – arXiv