Reliable Online Decision-Making with Machine Learning

Overview

Download the Presentation Slides

A research project on enhancing online conversion rates with machine learning predictions, integrating extensive ML algorithms with stock simulation predictions.

Multi-Model Time Series Prediction

Employed various machine learning models—including XGBoost, Transformer, VAE, and GAN—to model financial time series (e.g., the CSI 300 Index). Successfully achieved classification predictions for future trends and regression predictions for maximum prices. The Transformer model demonstrated a 19.11% improvement over the exponential smoothing method in RMSE metric.

Trustworthy Framework Design

Participated in designing and implementing a unified framework named Trustworthy Online Decision Making. By introducing the ε-tolerant OMS algorithm, this framework mitigates the risk of blind trust in machine learning predictions during online search tasks, theoretically ensuring the algorithm's robustness and consistency.

Strategic Optimization

For two classic online decision-making scenarios — One-max Search and One-way Trading — developed decision rules equipped with a real-time trust adjustment mechanism. Experimental results demonstrate that this mechanism achieves significantly better Competitive Ratio performance compared to the benchmark algorithm when handling noisy predictive signals, while closely approximating the offline optimal solution.