Titanic Survival Prediction: A Comparative Practice Between Traditional Machine Learning and Deep Learning

This project is based on the Titanic passenger dataset, comparing the performance of traditional machine learning and deep learning methods in the survival prediction task, focusing on the handling of imbalanced datasets and the application of the F1 score evaluation metric. The project covers the complete workflow including data preprocessing, feature engineering, model construction, and evaluation, aiming to explore the applicable scenarios and practical value of different algorithms.

The Titanic sinking incident is a classic case in data science. The core goal of this project to build a survival prediction system is not only accuracy but also to explore model evaluation methods under imbalanced datasets and the performance differences between the two paradigms. When data is imbalanced, accuracy tends to be artificially high, so the F1 score (harmonic mean of precision and recall) is chosen as the main evaluation metric.

The Titanic dataset includes features such as demographics (age, gender, cabin class), family relationships (number of siblings/spouses, number of parents/children), and boarding ports. Feature engineering needs to handle missing values (e.g., filling with age median), encode categorical variables (label/one-hot encoding), and can also extract implicit features from names/tickets (such as family relationships from surnames, cabin types from ticket prefixes).

Implement models such as logistic regression (baseline model, strong interpretability), random forest (ensemble trees resist overfitting), support vector machine (optimal hyperplane in high dimensions), and gradient boosting trees (serial weak learners). Tuning uses grid search + cross-validation, and imbalanced data is handled by adjusting class weights or oversampling to improve the recognition ability of minority classes.

The neural network architecture includes an input layer (number of neurons determined by feature dimension), hidden layers (depth and width affect expressive ability), and an output layer (Sigmoid activation for binary classification). Training uses backpropagation to calculate gradients, optimizers (Adam/SGD) to update weights, and prevents overfitting through Dropout, early stopping, and L2 regularization; the learning rate needs to be set appropriately.

Using the F1 score as the main evaluation metric, the confusion matrix provides a comprehensive performance view (precision, recall, etc.). Comparative results: On small-scale structured data, well-tuned traditional models perform as well as or even better than deep learning, with faster training and stronger interpretability; the advantages of deep learning are reflected in large-scale complex data.

The project demonstrates the complete life cycle of data science. Key gains: Evaluation metrics need to be combined with business and data characteristics; there is no absolutely optimal algorithm, only those suitable for specific problems; feature engineering is still important in traditional models; deep learning is not the optimal solution for all problems. The experience has reference value for competitions, business modeling, and academic research.