https://github.com/drkbluescience/autogluon_cameroon_air_quality

Finished 5th in the Cameroon Air Quality Prediction competition, later refining the model to achieve a score better than the 1st place submission using AutoGluon.
https://github.com/drkbluescience/autogluon_cameroon_air_quality

autogluon automl feature-engineering machine-learning regression-analysis tabular-data

Last synced: over 1 year ago
JSON representation

Finished 5th in the Cameroon Air Quality Prediction competition, later refining the model to achieve a score better than the 1st place submission using AutoGluon.

• Host: GitHub
• URL: https://github.com/drkbluescience/autogluon_cameroon_air_quality
• Owner: drkbluescience
• Created: 2024-11-13T12:16:29.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-11-14T10:06:17.000Z (over 1 year ago)
• Last Synced: 2025-01-14T11:33:07.255Z (over 1 year ago)
• Topics: autogluon, automl, feature-engineering, machine-learning, regression-analysis, tabular-data
• Language: Jupyter Notebook
• Homepage:
• Size: 4.16 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# **Cameroon Air Quality Prediction - AutoGluon**

## Introduction

This study focuses on predicting air quality in Cameroon, specifically the concentration of particulate matter (**PM2.5**), using various machine learning techniques. The dataset includes weather and air quality features collected from different cities across Cameroon.

## Leaderboard Achievement

Here’s a snapshot of the position on the leaderboard during the **Cameroon Air Quality Prediction** competition, showing the score in 5th place at the end of the competition. After further model improvements, the 1st-place score was surpassed.

![Leaderboard](images/leaderboard.png)

## Methodology

The analysis began with an exploration of the dataset, where **data inconsistencies** were addressed. Features with a **single value**, such as **'sunrise'**, **'sunset'**, and **'snowfall_sum'**, were removed. Redundant variables, including **city**, **longitude**, and **latitude**, were also eliminated to reduce unnecessary complexity in the models.

### Feature Engineering

Enhancing predictive power involved analyzing the distribution of **PM2.5** concentrations across different cities, leading to the creation of a new feature:

- **Distance from Bafoussam**, the city with the highest PM2.5 levels.

  

## Models

Several machine learning models were initially employed to predict **PM2.5** levels, including:

- **CatBoost**

- **LightGBM (LGBM)**

- **XGBoost (XGB)**

- **GradientBoostingRegressor**

- **ExtraTreesRegressor**

- **RandomForestRegressor**

- **AdaBoostRegressor**

- **MLPRegressor**

These models were evaluated using a **9-split RepeatedKFold cross-validation** strategy to ensure reliable results.

However, after initial testing, **AutoGluon** was introduced and ultimately provided the best performance, surpassing all other models in predictive accuracy.

## Results

Among all models tested, **CatBoost** performed well, achieving a **root mean squared error (RMSE)** of **3.11078**. However, **AutoGluon** outperformed every other model, achieving the **lowest RMSE of 2.97008**.

### Key Result Comparison

| Model                     | RMSE      |

| -------------------------- | --------- |

| **CatBoost**               | 3.11078   |

| **AutoGluon**              | **2.97008** |

This result demonstrates a **significant improvement** over the other models, indicating the superior predictive capabilities of **AutoGluon** for this particular task.

While other models, such as **CatBoost** and **ExtraTrees**, provided competitive results, **AutoGluon’s automatic model selection and hyperparameter tuning** led to the best performance, further validating its effectiveness as an **AutoML tool** for air quality prediction.

## Conclusion

This study highlights the critical role of **feature engineering** in improving model performance, as well as the superiority of **AutoGluon** over other machine learning models for the task of predicting **PM2.5** concentrations. AutoGluon’s automated approach to model selection and optimization resulted in a more accurate prediction, achieving the best RMSE score and outperforming traditional models like **CatBoost** and **XGBoost**.