CPU Latency Prediction Using a Regression Model with Ensemble Methods

Definition

Ensemble methods combine predictions from multiple models to improve the overall prediction accuracy and the main model's robustness.

Approach

• In our case, we used ensemble methods in order to enhance the coefficient of determination $\textbf R^2$ and mitigate individual model biases and variances for the prediction of CPU latency in Cloud environments.
• Initially, 4 algorithms including LGBMRegressor, XGBRegressor, RandomForestRegressor, and CatBoostRegressor were trained on the dataset. Each model captures different aspects of the data.
• Post-training, predictions from the 4 models were averaged using a simple averaging approach

Evaluation metric

The coefficient of determination $\textbf R^2$ was used to assess the models' performances.

$$ R^2 = 1 - \frac{\sum_{i}(y_i - \hat{y}i)^2}{\sum{i}(y_i - \bar{y})^2} $$

• $\sum_{i}(y_i - \hat{y}_i)^2 \quad \text{represents the Residual Sum of Squares (RSS).}$
• $\sum_{i}(y_i - \bar{y})^2 \quad \text{represents the Total Sum of Squares (TSS).}$

Results

The results are summarized in the following table

Model	Training R²	Validation R²
LGBM	0.9830	0.9741
XGBoost	0.9961	0.9723
RandomForest	0.9956	0.9747
CatBoost	0.9906	0.9756

Dataset	Average R²
Training	0.9913
Validation	0.9742

Note : Big thanks to SoAI for suggesting this fun challenge !