Machine Learning Analysis using OpenML
This project utilized the vast amount of data on OpenML to get insights into what methods for machine learning produce the most accurate classifiers. OpenML is a platform where people use different datasets and machine learning methods for classification tasks. The results from all the user uploaded runs are stored and made available for others to view. Analysis into which sklearn and weka primitives are the most accurate across the data was conducted. Additionally, dataset features based on metadata were created and used to cluster datasets using kmeans. Analysis of how different types of datasets respond to different machine learning methods was investigated.
This code for this project is split into 4 jupyter notebooks. OpenML_Download.ipynb is where all the data for the analysis is downloaded from OpenML. dataset_features.ipynb is where the features for all datasets used in the analysis were created. These features were then used to create dataset clusters using kmeans in dataset_cluster_analysis.ipynb. Clusters were used to identify how different weka primitives varied in classification accuracy for different types of datasets. In primitive_analysis.ipynb, sklearn and weka primitives were evaluated in terms of adjusted accuracy across all the data. The best and worst primitives for classification were identified. Some parts of the download and analysis were conducted using Brown CCV to handle the large amount of data and code for that is included in the notebooks.
The data associated with this project can be downloaded [here][download-url] [download-url]: https://brownbox.brown.edu/download.php?hash=de911e03 The file is 0.5 GB compressed and 5.7 GB once uncompressed