Skip to main content

Dual-Input Deep Neural Network (DNN)

This DNN combines two views of the same star: the raw light curve and a set of simple, engineered numbers about that curve. Together, they make a stronger prediction.

What it does​

  • Uses the light curve to catch transit shapes
  • Uses features (like variability and frequency) to add context
  • Merges both to decide: CANDIDATE or FALSE POSITIVE

When to use it​

  • You have light curves and can compute basic features (we support 24+ common ones)
  • You want better accuracy and explainability than using the curve alone

Data needed​

  • Light curve CSV per star → normalized to 3000 points
  • Feature table per star → statistical, variability, frequency, transit features

How the model works (simple)​

  1. Time-series branch (small CNN) learns shapes from the curve
  2. Feature branch (dense layers) learns patterns from numbers
  3. Both branches are merged → final dense layers → binary output

Architecture snapshot​

  • Time series: 1D conv → pooling → dropout → embedding
  • Features: Dense → batch norm → dropout → embedding
  • Merge: Concatenate → Dense → Sigmoid

What you’ll see (example outputs)​

Feature correlation

Feature importance

Training history

Architecture/visualization

Typical results (example)​

  • Train ≈ 0.98 | Validation ≈ 0.97 | Test ≈ 0.96 (can vary by dataset)

Tips​

  • Keep preprocessing consistent between training and prediction
  • Use SHAP plots to understand which features matter for a given star

A dual-branch DNN that fuses learned features from light curve time series with engineered tabular features. The pipeline includes feature extraction, preprocessing, and comprehensive explainability via SHAP.

Highlights​

  • Dual-input: 1) time series branch (CNN stack), 2) feature branch (dense stack)
  • Robust preprocessing: missing data handling, outlier removal, robust scaling
  • Explainability: SHAP for global and local feature importances
  • End-to-end training script with configuration-driven hyperparameters

Architecture​

  • Time series branch: 1D conv + max pooling + dropout on 3000-point normalized sequences
  • Feature branch: Dense layers with batch norm + dropout on statistical/astronomical features
  • Fusion: Concatenation followed by dense layers to binary output (Candidate vs False Positive)

Data processing​

  • Light curves: normalize (median/MAD), sigma clipping, fixed length 3000
  • Feature extraction: 24+ features
    • Statistical: mean, std, median, skew, kurtosis
    • Variability: amplitude, flux excursions, range
    • Frequency: dominant frequency, spectral features
    • Transit: dip detection, periodicity
  • KOI catalog integration, imputation, scaling

Training pipeline​

  • Load + preprocess sequences and features
  • Train/validation/test splits
  • Build and train model with configured hyperparameters
  • Evaluate metrics and generate SHAP explanations

Reported metrics (snapshot)​

  • Train acc: 0.9757; Val acc: 0.9669; Best val loss: 0.0797 at epoch ~59
  • Test acc: 0.9560 on 182 samples; features used: 12 (example snapshot)

Usage examples​

  • Training: python train.py
  • Config: configs/model_config.yaml for architecture, training, preprocessing
  • Programmatic:
    • create_dual_input_dnn_model(sequence_length=3000, feature_dim=36, ...)
    • load_and_preprocess_data() returns sequences, features, labels

Notes​

  • SHAP-based explainability is integrated for model trust
  • Ensure consistent preprocessing for batch prediction
  • TensorFlow 2.x, scikit-learn, pandas, numpy, SHAP are typical dependencies