EasyResearch

Machine learning without programming: analysis cores for researchers who do not code. They follow the rules of EasyClassifier:

• every step that learns from data uses training data only;
• models use their authors' default settings with fixed random seeds (no tuning), so results are reproducible;
• when several models are compared, the selected one gets a separate, honest final score;
• every run writes a plain-language report with a Methods paragraph, figures, tables, predictions, the trained model, citations and a log.

The cores have no user interaction (settings in, results out), so a terminal wizard, the desktop application or a web server can all drive them.

This repository holds:

Install and run

Windows, without Python: download EasyResearchDesktop-0.1.0-setup.exe from the Releases page and run it. It installs EasyResearch Desktop with its own Python and every package it needs; no administrator rights are required. Windows may show "Windows protected your PC" because the program is not code-signed: choose More info, then Run anyway.

Python (any system):

pip install easyresearch    # installs EasyClassifier, PyTorch and statsmodels
easyresearch forecast --demo
easyresearch forecast sales.csv --value Sales --horizon 12
easyresearch regress houses.xlsx --target Price
easyresearch signals ECG200_TRAIN.tsv
easyresearch models forecast            # list models; * = default fast set

Add --models all, or a list such as --models ets,lstm,nbeats, to choose models. Results go to a new folder in ./Results (or --out). Deep models use a GPU automatically when PyTorch finds one.

Figures: --figures residuals,comparison draws only the named figures (the default is the task's usual set), --figure-theme chooses colour-blind safe (default), greyscale or high-contrast colours, and --figure-format png+pdf or png+svg adds vector files for journals.

Forecasting

• Data: a CSV/Excel table with a date column (detected automatically) and the value to forecast. Hourly, daily, business-daily, weekly, monthly, quarterly and yearly series are recognised; the season length follows (24, 7, 5, 52, 12, 4, none). Repeated time points are averaged; missing points are filled with the last known value (never a later one). Without a date column, rows are taken in order with no season.
• Models: reference rules (naive, seasonal naive) are always included. Statistical: exponential smoothing (damped trend, seasonal) and Theta. Machine learning on the last L values: Ridge, Random Forest, Gradient Boosting, KNN with the Hassanat distance. Deep (PyTorch): MLP, LSTM, GRU, TCN, N-BEATS, Transformer. Default set: ETS, Theta, Ridge, Random Forest (deep models are optional because each takes minutes).
• Honest evaluation: never shuffled. Models are compared at up to five rolling origins before a final period (the last ~20%); the selected model is then scored in the final period, which played no part in choosing it, retrained before each origin. The automatic horizon is the usual one for the time unit, shortened (with a note) when the series is too short to evaluate it honestly; a series too short for any horizon is refused with the number of values it needs.
• Measures: MASE (used to choose, by default), MAE, RMSE, sMAPE and MHSP (mean Hassanat similarity percentage) – any of MASE, RMSE, MAE or MHSP can choose the model (--metric mhsp).
• Outputs: the future forecast with an 80% band (its coverage in the final period is checked and reported), final-period forecasts, comparison table, five figures, report, trained model, settings.json (what was asked and what was used) and versions.txt (the software used).

MHSP

For each prediction, the Hassanat similarity between actual a and forecast f is (1 + min(a, f)) / (1 + max(a, f)), with the signed form (1 + min + |min|) / (1 + max + |min|) when the smaller value is negative. MHSP is 100 × the mean of these similarities (100% = perfect), computed on the values as they are. It is also reported by the regression core. Reference: Hassanat, A. B., et al. (2024). A novel outlier-robust accuracy measure for machine learning regression using a non-convex distance metric. Mathematics, 12(22), 3623. https://doi.org/10.3390/math12223623

Signal classification (ECG, EEG, sensors)

• Data: one recording per row – a CSV/Excel table with a label column and the signal values in the other numeric columns, or UCR/UEA archive files (.tsv/.txt, label first). Gaps inside a recording are filled from its own neighbouring values.
• Models: KNN with the Hassanat distance (each recording scaled to 0–1), ROCKET (10,000 random convolution kernels), signal features + Random Forest, and deep FCN, ResNet, InceptionTime and LSTM.
• Evaluation: EasyClassifier's own – stratified cross-validation to compare; nested cross-validation for the final score (up to 2,000 recordings, without deep networks) or an untouched stratified 20% test set (larger data, or when deep networks are compared, because they take minutes each on a normal computer).
• The demo (--demo) uses synthetic heartbeats for trying the tool; it is not patient data.
• Not yet: several channels per recording (multichannel EEG) and long recordings to be cut into windows.

Regression

See the EasyResearch Desktop README for details: linear and ridge regression, decision tree, random forest, gradient boosting, SVR, KNN with the Hassanat distance and a neural network; R², RMSE, MAE, median absolute error, MAPE and MHSP against the reference of always predicting the average. XGBoost and LightGBM are added when installed (pip install -e ".[full]").

Development

pip install -e ".[test]"
python -m pytest tests

From the repository root, python tools/run_all_tests.py runs every suite (these cores and the desktop adapters); see docs/DEVELOPMENT.md. benchmarks/easyresearch_benchmarks.py measures accuracy, run time, memory and agreement with reference implementations, and benchmarks/easyresearch_honesty.py measures whether the reported scores are honest, against data that is never touched (results in benchmarks/results/easyresearch_honesty).

The tests include a check that no forecasting model – statistical, machine learning or deep – changes its forecast when every value after the forecast origin is changed.

Citing

If you use EasyResearch in published work, please cite it (GitHub's "Cite this repository" button uses CITATION.cff); every run also writes the exact references for the methods it used to citations.txt.

Licence

MIT. Ahmad Hassanat.