CONECT

This repository provides training and evaluation pipelines for CONECT.

---

1. Environment Setup

We provide two environment specifications:

• env.yaml — recommended for reproducible setup with Conda.
• env.txt — an explicit (more strictly pinned) Conda package list.

Option A (Recommended): Create from env.yaml

conda env create -f env.yaml
conda activate <ENV_NAME>

Option B: Create from env.txt (Explicit / Locked)

conda create -n <ENV_NAME> --file env.txt
conda activate <ENV_NAME>

If you already have an environment and want to update it using env.yaml:

conda env update -f env.yaml --prune

---

2. Running the Scripts

We provide three runnable shell scripts:

• CONECT_fintune.sh — Fine-tune on a dataset (training + evaluation).
• CONECT_Pretrain.sh — Pre-train primarily on a dataset.
• CONECT_Zeroshot.sh — Zero-shot evaluation using a trained checkpoint (evaluation only).

Before running: open each script and set dataset paths, output paths, and hyperparameters to match your environment.

2.1 Fine-tuning (Training + Evaluation)

Fine-tuning trains the model on the target dataset and then evaluates on the corresponding evaluation split.

bash ./scripts/CONECT_fintune.sh

2.2 Pre-training

Pre-training is primarily performed on the dataset(s) specified in the script to produce a strong initialization checkpoint.

bash ./scripts/CONECT_Pretrain.sh

2.3 Zero-shot Evaluation (Checkpoint-only Evaluation)

Zero-shot evaluation runs evaluation directly from an existing checkpoint, without additional training on the target dataset.

bash ./scripts/CONECT_Zeroshot.sh

---

3. Baselines

For major deep learning baselines, please refer to the implementations organized in the TEMPO repository:

• TEMPO (baseline implementations): https://github.com/DC-research/TEMPO/tree/main

Other baselines used in our experiments:

1. TabPFN-TimeSeries: https://github.com/PriorLabs/tabpfn-time-series
2. Chronos: https://github.com/amazon-science/chronos-forecasting
3. Moirai: https://github.com/redoules/moirai
4. TGTSF: https://github.com/VEWOXIC/TGTSF
5. ContextMatters (paper): https://arxiv.org/pdf/2410.12672
6. TimeXer: https://github.com/thuml/TimeXer

Some baselines are external repositories/papers. Please follow their official installation and usage instructions.

---

4. Notes and Sugguestions

1. Experiments use commonly adopted public benchmark datasets. Please note that for some datasets with substantial noise (e.g., Bitcoin), it is recommended to apply appropriate normalization / preprocessing before training and evaluation.

2. Due to large variations in data distribution and dataset quality, achieving stable and strong performance may require careful and repeated tuning of relevant hyperparameters (e.g., learning rate, batch size, context window length, regularization, and scheduler settings).

3. The M5 dataset is from the official Kaggle M5 Forecasting competition. We also provide a corresponding data loader for M5.

---

Acknowledgements

This codebase builds on and references multiple open-source repositories listed above. We thank the authors for their valuable contributions.

---