LogicGate: Adaptive Rule-Based Modeling of Exogenous Effects for Time Series Forecasting

Official implementation of the paper LogicGate: Adaptive Rule-Based Modeling of Exogenous Effects for Time Series Forecasting.

1. Project Overview

LogicGate is a forecasting framework for exogenous-aware time series prediction.
Compared with static rule sets and shallow fusion methods, this implementation focuses on:

• dynamic rule filtering based on current exogenous context,
• cross-modal fusion between time-series patches and rule semantics,
• dual-track self-optimization of rule quality during training.

2. Installation

pip install -r requirements.txt

Recommended environment:

• Python 3.10+
• PyTorch 2.0+

If you use optimization mode, set API key in environment variables:

# Linux/Mac
export OPENAI_API_KEY=your_key

# Windows PowerShell
$env:OPENAI_API_KEY="your_key"

3. Data and Rule Config

Data

Put datasets under dataset/ (or pass custom paths via --root_path and --data_path).

Rule Config

• Default global rule file: config/rule_patterns.json
• Dataset-specific rule files: config/rule_config/
  • rule_patterns_BE.json
  • rule_patterns_DE.json
  • rule_patterns_ETTh1.json
  • rule_patterns_FR.json
  • rule_patterns_NP.json
  • rule_patterns_wet_bulb.json

Rule files are used by:

• --rules_list (load rule text for model input)
• --rule_config_path (rule-pattern filtering/analysis path)

4. Quick Start

4.1 Baseline Training

python train.py \
  --task_name long_term_forecast \
  --is_training 1 \
  --model_id demo_baseline \
  --model_comment baseline \
  --model logicgate \
  --data ETT_exog \
  --root_path ./dataset/electricity-price \
  --data_path NP.csv \
  --features MS \
  --target OT \
  --seq_len 168 \
  --label_len 48 \
  --pred_len 24 \
  --enc_in 3 --dec_in 3 --c_out 1 \
  --train_epochs 10 \
  --batch_size 16 \
  --learning_rate 0.001 \
  --optimization_mode baseline

4.2 Monitoring Mode

python train.py ... --optimization_mode monitoring

4.3 Full Optimization Mode

python train.py ... \
  --optimization_mode optimization \
  --optimization_config config/optimization_config_stage4.json \
  --rules_list config/rule_patterns.json \
  --rule_config_path config/rule_patterns.json

Or use the stage-4 runner:

python run_optimization_stage4.py

4.4 Dataset Script Entry

python script/run_NP.py \
  --task_name long_term_forecast \
  --is_training 1 \
  --model_id np_demo \
  --model_comment test \
  --model logicgate \
  --data ETT_exog

5. Key Configurations

Important config files:

• config/rule_patterns.json
• config/optimization_config.json
• config/optimization_config_stage4.json (for stage-4 optimization flow)

Common arguments in train.py:

• --optimization_mode: baseline | monitoring | optimization
• --rules_list: rule file used for model rule text loading
• --rule_config_path: rule file path used in filtering/analysis
• --optimization_config: optimization pipeline config path
• --rule_filter_method: basic | fast
• --num_exog_vars: number of exogenous variables