Multi-Dataset Machine Learning Benchmark

A notebook-based benchmark that downloads user-provided CSV datasets from Google Drive, applies a shared preprocessing pipeline, trains 50 scikit-learn algorithms across four machine-learning categories, evaluates the successful runs, and produces comparative plots for the top and bottom performers.

This README documents the supplied notebook and its recorded outputs only. No unrecorded results or dataset provenance has been added.

Overview

The project is designed to compare a broad collection of supervised and unsupervised algorithms on multiple tabular datasets. For each uploaded dataset, the notebook performs data cleaning, identifies the target column, creates a reproducible train/test split, scales the features, trains the configured models, saves the result records to CSV, and visualizes category-level comparisons.

The supplied run processed three datasets:

Dataset	Rows	Columns	Target used by the run	Input mechanism
Dataset_1_ID_1oQnqc	253,680	22	Diabetes_012	Google Drive CSV
Dataset_2_ID_1YB7Oz	70,692	22	Income	Google Drive CSV; fallback to last column
Dataset_3_ID_1caYPh	253,680	22	Income	Google Drive CSV; fallback to last column

The notebook does not specify the original dataset names, feature descriptions, licensing, or a public source citation beyond the Google Drive links entered in the notebook.

Key Features

• Accepts one or more Google Drive file links and extracts their file IDs.
• Downloads and loads CSV datasets with gdown and pandas.
• Handles several missing-value tokens: empty strings, nan, null, none, and ?.
• Uses mean imputation for numeric-looking columns and mode imputation for other columns.
• Converts the feature matrix and target to numeric values.
• Uses an 80/20 train/test split with random_state=42.
• Applies StandardScaler to the training and test features.
• Limits model fitting and clustering to a reproducible sample of up to 1,000 training rows.
• Evaluates classification, regression, and clustering runs and records dimensionality-reduction successes.
• Saves results to ColabData/custom_datasets_results.csv.
• Produces per-dataset top-five/worst-five plots and overall comparison plots.

Project Workflow

Google Drive CSV links
        ↓
Download and load datasets
        ↓
Missing-value handling and numeric conversion
        ↓
Target-column selection
        ↓
80/20 train/test split
        ↓
StandardScaler fitted on training data
        ↓
Up to 1,000 sampled training rows for model fitting
        ↓
50 algorithms across four categories
        ↓
Metrics and status records
        ↓
CSV export and comparison plots

Algorithms

The notebook defines exactly 50 algorithms in four categories.

Classification — 15 algorithms

1. Logistic Regression
2. Ridge Classifier
3. SGD Classifier
4. Perceptron
5. Passive Aggressive
6. Linear SVC
7. SVC with RBF kernel
8. K-Nearest Neighbors
9. Gaussian Naive Bayes
10. Multinomial Naive Bayes
11. Bernoulli Naive Bayes
12. Decision Tree
13. Random Forest
14. Extra Trees
15. HistGradientBoosting

Regression — 15 algorithms

1. Linear Regression
2. Ridge Regression
3. Lasso Regression
4. Elastic Net
5. Huber Regressor
6. SGD Regressor
7. Support Vector Regression
8. K-Nearest Neighbors Regressor
9. Decision Tree Regressor
10. Random Forest Regressor
11. Extra Trees Regressor
12. HistGradientBoosting Regressor
13. Gradient Boosting Regressor
14. AdaBoost Regressor
15. Poisson Regressor

Clustering — 10 algorithms

1. K-Means
2. Mini-Batch K-Means
3. Bisecting K-Means
4. Agglomerative Clustering
5. Birch
6. DBSCAN
7. OPTICS
8. Mean Shift
9. Spectral Clustering
10. Gaussian Mixture

Dimensionality Reduction — 10 algorithms

1. PCA
2. Incremental PCA
3. Kernel PCA
4. Truncated SVD
5. NMF
6. FastICA
7. Factor Analysis
8. Isomap
9. Locally Linear Embedding
10. MDS

Evaluation Metrics

The executed evaluation code records the following metrics:

• Accuracy (%) and weighted F1 score for classification.
• MAE, RMSE, and R² for regression.
• Silhouette score and number of detected clusters for clustering.
• Number of transformed components and execution time for dimensionality reduction.

calinski_harabasz_score is imported in the notebook but is not used in the executed result records or plots.

Recorded Results

The following values are taken directly from the notebook’s stored output. “Best” and “worst” follow the notebook’s own ordering: highest accuracy for classification, lowest MAE for regression, and highest/lowest silhouette score for clustering.

Dataset 1 — Dataset_1_ID_1oQnqc

Category	Best recorded result	Worst recorded result
Classification	Ridge Classifier — 84.27% accuracy	Gaussian Naive Bayes — 8.76% accuracy
Regression	Huber Regressor — MAE 0.3027	AdaBoost Regressor — MAE 0.5628
Clustering	Spectral Clustering — silhouette 0.2759	OPTICS — silhouette -0.1553

Other recorded classification leaders were SVC RBF (84.24%), Linear SVC (84.20%), Random Forest (84.02%), and Logistic Regression (83.86%). Other recorded regression leaders were Support Vector Regression (0.3664 MAE), K-Nearest Neighbors Regressor (0.4079), Gradient Boosting Regressor (0.4247), and Extra Trees Regressor (0.4317).

Dataset 2 — Dataset_2_ID_1YB7Oz

Category	Best recorded result	Worst recorded result
Classification	Ridge Classifier — 32.58% accuracy	Gaussian Naive Bayes — 12.87% accuracy
Regression	Support Vector Regression — MAE 1.3749	Decision Tree Regressor — MAE 1.8615
Clustering	Spectral Clustering — silhouette 0.2370	OPTICS — silhouette -0.1936

Other recorded classification leaders were Linear SVC (32.48%), Logistic Regression (32.12%), SVC RBF (32.01%), and Multinomial Naive Bayes (30.79%). Other recorded regression leaders were Huber Regressor (1.3876 MAE), Gradient Boosting Regressor (1.3896), Linear Regression (1.4010), and Ridge Regression (1.4010).

Dataset 3 — Dataset_3_ID_1caYPh

Category	Best recorded result	Worst recorded result
Classification	Ridge Classifier — 37.41% accuracy	Passive Aggressive — 18.64% accuracy
Regression	Support Vector Regression — MAE 1.3077	Decision Tree Regressor — MAE 1.7177
Clustering	Spectral Clustering — silhouette 0.2734	OPTICS — silhouette -0.1840

Other recorded classification leaders were Linear SVC (37.28%), Logistic Regression (36.97%), SVC RBF (36.79%), and Multinomial Naive Bayes (35.68%). Other recorded regression leaders were Huber Regressor (1.3162 MAE), Ridge Regression (1.3306), Linear Regression (1.3306), and SGD Regressor (1.3333).

Visualizations

All images below were extracted from the notebook’s actual embedded PNG outputs and renamed for clarity. The plots show the recorded run; they are not newly generated benchmark results.

Dataset 1

Dataset 2

Dataset 3

Overall comparisons

No dimensionality-reduction plot is recorded in the supplied notebook output. The dimensionality-reduction loop records execution status, runtime, and component count, but does not display a chart.

Installation

The project does not include a requirements.txt file. Install the packages imported by the notebook in a Python 3 environment:

pip install numpy pandas matplotlib seaborn scikit-learn gdown ipython jupyter

Usage

Google Colab

1. Open project_for_ml_dinal.ipynb in Google Colab.
2. Install the dependencies listed above if they are not already available.
3. Replace MY_DATASET_LINKS with one or more accessible Google Drive file links, separated by spaces or new lines.
4. Run the notebook cells in order.
5. Review the printed performance report and plots.
6. Find the exported results at ColabData/custom_datasets_results.csv.

The notebook uses Colab-style cell annotations such as #@title and #@markdown, and its recorded run uses a Python 3 kernel.

Local Jupyter or VS Code

The notebook is standard .ipynb Python code and can also be opened in Jupyter or VS Code with the Jupyter extension. The supplied project does not contain a separate VS Code configuration or a recorded VS Code run. Ensure that the Google Drive links are accessible and that the working directory permits creation of the ColabData/ folder.

Project Structure

project_for_ml_dinal/
├── project_for_ml_dinal.ipynb
├── README.md
└── assets/
    ├── dataset1_classification_results.png
    ├── dataset1_regression_results.png
    ├── dataset1_clustering_results.png
    ├── dataset2_classification_results.png
    ├── dataset2_regression_results.png
    ├── dataset2_clustering_results.png
    ├── dataset3_classification_results.png
    ├── dataset3_regression_results.png
    ├── dataset3_clustering_results.png
    ├── overall_classification_comparison.png
    ├── overall_regression_comparison.png
    └── overall_clustering_comparison.png

ColabData/ is created at runtime by the notebook and is not part of the supplied upload. The recorded execution also writes custom_datasets_results.csv into that directory.

Limitations and Observations

• The notebook has no requirements.txt, license file, author information, or separate source modules.
• Dataset provenance is not specified beyond the Google Drive links embedded in the notebook.
• The default target is Diabetes_012; if that column is absent, the code silently uses the last column. In the recorded run, this resulted in Income for Datasets 2 and 3.
• The same target is used for classification, regression, and preprocessing. Whether each target is semantically appropriate for all three task types is not documented in the project.
• Models are fitted on at most 1,000 randomly sampled training rows, while evaluation uses the full test split. Therefore, the displayed results are sample-fit benchmark results rather than full-training results.
• The notebook uses one train/test split and does not perform cross-validation or hyperparameter tuning.
• Several algorithms can fail on particular datasets; failures are captured in results_df, but the supplied printed report shows successful results only.
• Dimensionality-reduction outputs are recorded as component counts and status records, but are not visualized in the notebook output.
• The imported Calinski–Harabasz metric is not calculated in the execution loop.
• The notebook does not save trained models, preprocessing scalers, or a machine-readable experiment configuration beyond the results CSV.

Future Improvements

The following are proposed improvements, not features documented as already implemented:

• Add a pinned requirements.txt and a license.
• Document each dataset’s original name, schema, feature meanings, and license.
• Validate target semantics separately for classification and regression.
• Replace silent target fallback with an explicit configuration or validation error.
• Add cross-validation, hyperparameter tuning, and confidence intervals.
• Record class distributions and baseline scores.
• Add dimensionality-reduction visualizations and use the imported Calinski–Harabasz metric where appropriate.
• Save fitted models, scalers, and a complete experiment manifest.
• Add automated tests for data loading, preprocessing, and result-schema consistency.

Technologies

• Python 3
• Jupyter Notebook / Google Colab-compatible notebook
• NumPy
• Pandas
• Matplotlib
• Seaborn
• scikit-learn
• gdown
• IPython

Author

Author information is not specified in the supplied notebook.