Skip to content

ch05 — Gen4 Graphs and Attention (2020–2023)

Companion notebook: notebooks/ch05_gen4.ipynb

Graphs: modeling sensor relations explicitly — GDN

In multivariate series, "which sensors move together" is central information for anomaly detection. GDN (Deng & Hooi, AAAI 2021):

  1. a learnable embedding \(v_i\) per sensor
  2. a top-k graph from embedding cosine similarity
  3. graph attention aggregates neighbors to forecast the next value
  4. the robustly normalized maximum of per-sensor deviations is the score — which also tells you which sensor caused it

Attention: association discrepancy — Anomaly Transformer

The insight of Anomaly Transformer (Xu et al., ICLR 2022): anomalous points can only attend to their local neighborhood (they fail to associate with the global pattern).

  • series-association: the attention distribution learned from data
  • prior-association: a distance-based Gaussian (the baseline of local concentration)
  • the KL distance between them (association discrepancy) is small for anomalies — multiplied with reconstruction error to form the score

DCdetector (KDD 2023) implements the same intuition without reconstruction, purely by contrasting two attention views (patch-wise vs in-patch).

TimesNet: folding periods into 2D

Find the dominant period by FFT, fold the 1D series into a 2D (period × phase) tensor, and apply 2D convolutions — capturing intraperiod patterns and interperiod evolution at once.

The honest question

This generation sits at the center of the benchmark-inflation controversy. Anomaly Transformer's original numbers were PA-F1 based; re-evaluated without PA the rankings change dramatically (Kim et al., AAAI 2022; TSB-AD, NeurIPS 2024). Check the actual VUS-PR rankings on our leaderboard — the comparison against Gen1–2 baselines is the point.