Add DCdetector anomaly detection model

[Copilot]

Adds DCdetector (KDD 2023) — a dual-attention contrastive model for time series anomaly detection — as a new pypots.anomaly_detection model.

Architecture

• Multi-scale patching: splits input into two complementary views per patch size — patch-wise (inter-patch) and in-patch (intra-patch) — then embeds each with a Conv1D + positional encoding
• DAC attention (DACStructure): computes separate softmax attention for each view, upsamples both back to full window size [B, H, L, L], averages over the channel dimension
• RevIN normalization: uses existing pypots.nn.modules.revin.RevIN
• Minimax contrastive loss: prior_loss − series_loss (symmetric KL divergence between the two attention views)

Anomaly scoring

Unlike reconstruction-based detectors, DCdetector has no reconstruction head. predict() is fully overridden to compute per-time-step anomaly scores as softmax(−KL_series − KL_prior) × temperature (temperature=50 per paper), with threshold derived from the training set distribution.

New files

File	Purpose
pypots/nn/modules/dcdetector/layers.py	BackboneDCdetector + all sub-modules
pypots/anomaly_detection/dcdetector/core.py	_DCdetector (contrastive loss, ModelCore)
pypots/anomaly_detection/dcdetector/model.py	User-facing DCdetector (BaseNNDetector)
tests/anomaly_detection/dcdetector.py	Full test suite

Usage

from pypots.anomaly_detection import DCdetector

model = DCdetector(
    n_steps=100,       # must be divisible by every patch size
    n_features=55,
    anomaly_rate=0.01,
    patch_sizes=[5, 10, 20],
    d_model=256,
    n_heads=4,
    e_layers=3,
)
model.fit(train_set)
results = model.predict(test_set)  # {"anomaly_detection": np.ndarray}

n_steps must be divisible by every value in patch_sizes; the constructor asserts this upfront.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Add DCdetector</issue_title>
<issue_description>### 1. Model description

DCdetector should be included in PyPOTS as an anomaly detection model

@inproceedings{yang2023dcdetector,
title={DCdetector: Dual Attention Contrastive Representation Learning for Time Series Anomaly Detection},
author={Yiyuan Yang and Chaoli Zhang and Tian Zhou and Qingsong Wen and Liang Sun},
booktitle={Proc. 29th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD 2023)},
location = {Long Beach, CA},
pages={3033–3045},
year={2023}
}

2. Check open-source status

• The model implementation is publicly available

3. Provide useful information for the implementation

Check out the official code repository here https://github.com/DAMO-DI-ML/KDD2023-DCdetector

4. Your contribution

Will make a PR</issue_description>

<agent_instructions>DCdetector uses RevIN as a submodule. RevIN is in PyPOTS already. You can find it in pypots/nn/modules/revin/layers.py</agent_instructions>

Comments on the Issue (you are @copilot in this section)

• Fixes Add DCdetector #812

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[sonarqubecloud]

Quality Gate passed

Issues
6 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
2.0% Duplication on New Code

See analysis details on SonarQube Cloud

[WenjieDu]

@yyysjz1997 Hi Yiyuan, as the 1st author of DCdetector, could you help review this PR please

[yyysjz1997]

@yyysjz1997 Hi Yiyuan, as the 1st author of DCdetector, could you help review this PR please

Hi Wenjie, I just noticed that the hyperparameter patch_size is usually a small prime number list, e.g., [3,5,7]. And the window size hyperparameter n_steps should be divisible by every patch size (e.g., 105 for patch_size = [3,5,7]).

[WenjieDu]

@yyysjz1997 Got it. Do you have any further comments on this PR?