🚀 Spaceship Titanic: TensorFlow Decision Forests

A robust baseline solution using TensorFlow Decision Forests (TF-DF) for the Kaggle Spaceship Titanic competition.

📌 Overview

This project aims to predict whether a passenger on the Spaceship Titanic was transported to an alternate dimension. We utilize TensorFlow Decision Forests (TF-DF), a library that enables the training of tree-based models (such as Random Forest) using the familiar and easy-to-use Keras API.

This approach is highly effective for tabular data as it requires minimal preprocessing compared to classical Neural Networks.

📊 Dataset

The dataset is sourced from the Kaggle competition: Spaceship Titanic.

• Total Training Data: 8,693 entries with 14 features.
• Target: Transported (Boolean) - Whether the passenger was transported.
• Key Features: HomePlanet, CryoSleep, Cabin, Destination, Age, VIP, and luxury amenities expenditure (RoomService, FoodCourt, etc.).

🛠️ Methodology & Workflow

The code in this repository covers the following end-to-end steps:

1. Exploratory Data Analysis (EDA)

We analyzed the distribution of both numerical and categorical data to understand passenger characteristics and identify patterns.

2. Data Preprocessing

While TF-DF handles many data types natively, some adjustments were necessary:

• Handling Missing Values: Null values in numerical and boolean columns were imputed with 0.
• Boolean Conversion: Since TF-DF does not currently support boolean data types directly, columns like Transported, VIP, and CryoSleep were converted to integer (0 or 1).
• Dropping Columns: Removed PassengerId and Name as they are irrelevant for training.

3. Feature Engineering

The Cabin feature, which contains data in the format Deck/Num/Side, was split into three more informative features:

• Deck
• Cabin_num
• Side

4. Modeling

We utilized the standard Random Forest algorithm from TF-DF.

• Model: tfdf.keras.RandomForestModel()
• Data Split: 80% Training, 20% Validation.
• Data Format: Converted Pandas DataFrame to tf.data.Dataset for optimal performance.

📈 Evaluation & Performance

The model was evaluated using accuracy metrics on the validation set and Out-of-Bag (OOB) data.

• Training Time: ~54 seconds.
• OOB Accuracy: ~79.73%
• Validation Accuracy: 80.25%

Feature Importance

Based on the NUM_AS_ROOT metric (how often a feature appears as the root of a tree), the most influential features are:

1. CryoSleep (Highly dominant)
2. RoomService
3. Spa
4. VRDeck

💻 How to Run

1. Install Dependencies:

   pip install tensorflow tensorflow_decision_forests pandas numpy seaborn matplotlib

2. Run the Notebook: Open the notebook file (e.g., spaceship_titanic_tfdf.ipynb) in Jupyter Notebook, Google Colab, or a Kaggle Kernel.

3. Output: The script will generate a submission.csv file ready for upload to Kaggle.

🤝 Contribution

This repository is designed as a learning baseline. Feel free to fork and experiment with:

• Using GradientBoostedTreesModel instead of Random Forest.
• Conducting deeper hyperparameter tuning.
• Implementing more advanced missing data imputation techniques.

---

Created based on TensorFlow Decision Forests v1.2.0 implementation.