A framework of PERformance METRICS (PerMetrics) for artificial intelligence models

"Knowledge is power, sharing it is the premise of progress in life. It seems like a burden to someone, but it is the only way to achieve immortality." --- Thieu Nguyen

Quick notification

• The version 1.2.0 has serious problem with calculate multiple metrics (OOP style), please update to version 1.2.1 as soon as possible for your sake.
• Add more metrics to version 1.2.2

Introduction

• PerMetrics is a python library for performance metrics of machine learning models.

• The goals of this framework are:

  • Combine all metrics for regression, classification and clustering models
  • Helping users in all field access to metrics as fast as possible
  • Perform Qualitative Analysis of models.
  • Perform Quantitative Analysis of models.

Dependencies

• Python (>= 3.6)
• Numpy (>= 1.15.1)

User installation

Install the current PyPI release:

pip install permetrics==1.2.2

Or install the development version from GitHub:

pip install git+https://github.com/thieu1995/permetrics

Example

• Permetrics version >= 1.2.0

from numpy import array
from permetrics.regression import RegressionMetric

## For 1-D array
y_true = array([3, -0.5, 2, 7])
y_pred = array([2.5, 0.0, 2, 8])

evaluator = RegressionMetric(y_true, y_pred, decimal=5)
print(evaluator.RMSE())
print(evaluator.MSE())

## For > 1-D array
y_true = array([[0.5, 1], [-1, 1], [7, -6]])
y_pred = array([[0, 2], [-1, 2], [8, -5]])

evaluator = RegressionMetric(y_true, y_pred, decimal=5)
print(evaluator.RMSE(multi_output="raw_values", decimal=5))
print(evaluator.MAE(multi_output="raw_values", decimal=5))


## All metrics

EVS = evs = explained_variance_score
ME = me = max_error
MBE = mbe = mean_bias_error
MAE = mae = mean_absolute_error
MSE = mse = mean_squared_error
RMSE = rmse = root_mean_squared_error
MSLE = msle = mean_squared_log_error
MedAE = medae = median_absolute_error
MRE = mre = MRB = mrb = mean_relative_bias = mean_relative_error
MPE = mpe = mean_percentage_error
MAPE = mape = mean_absolute_percentage_error
SMAPE = smape = symmetric_mean_absolute_percentage_error
MAAPE = maape = mean_arctangent_absolute_percentage_error
MASE = mase = mean_absolute_scaled_error
NSE = nse = nash_sutcliffe_efficiency
NNSE = nnse = normalized_nash_sutcliffe_efficiency
WI = wi = willmott_index
R = r = PCC = pcc = pearson_correlation_coefficient
AR = ar = APCC = apcc = absolute_pearson_correlation_coefficient
R2s = r2s = pearson_correlation_coefficient_square
CI = ci = confidence_index
COD = cod = R2 = r2 = coefficient_of_determination
ACOD = acod = AR2 = ar2 = adjusted_coefficient_of_determination
DRV = drv = deviation_of_runoff_volume
KGE = kge = kling_gupta_efficiency
GINI = gini = gini_coefficient
GINI_WIKI = gini_wiki = gini_coefficient_wiki
PCD = pcd = prediction_of_change_in_direction
CE = ce = cross_entropy
KLD = kld = kullback_leibler_divergence
JSD = jsd = jensen_shannon_divergence
VAF = vaf = variance_accounted_for
RAE = rae = relative_absolute_error
A10 = a10 = a10_index
A20 = a20 = a20_index
A30 = a30 = a30_index
NRMSE = nrmse = normalized_root_mean_square_error
RSE = rse = residual_standard_error

RE = re = RB = rb = single_relative_bias = single_relative_error
AE = ae = single_absolute_error
SE = se = single_squared_error
SLE = sle = single_squared_log_error

• Permetrics version <= 1.1.3

##  All you need to do is: (Make sure your y_true and y_pred is a numpy array)
## For example with RMSE:

from numpy import array
from permetrics.regression import Metrics

## 1-D array
y_true = array([3, -0.5, 2, 7])
y_pred = array([2.5, 0.0, 2, 8])

y_true2 = array([3, -0.5, 2, 7])
y_pred2 = array([2.5, 0.0, 2, 9])

### C1. Using OOP style - very powerful when calculating multiple metrics
obj1 = Metrics(y_true, y_pred)  # Pass the data here
result = obj1.root_mean_squared_error(clean=True, decimal=5)
print(f"1-D array, OOP style: {result}")

### C2. Using functional style
obj2 = Metrics()
result = obj2.root_mean_squared_error(clean=True, decimal=5, y_true=y_true2, y_pred=y_pred2)  
# Pass the data here, remember the keywords (y_true, y_pred)
print(f"1-D array, Functional style: {result}")

## > 1-D array - Multi-dimensional Array
y_true = array([[0.5, 1], [-1, 1], [7, -6]])
y_pred = array([[0, 2], [-1, 2], [8, -5]])

multi_outputs = [None, "raw_values", [0.3, 1.2], array([0.5, 0.2]), (0.1, 0.9)]
obj3 = Metrics(y_true, y_pred)
for multi_output in multi_outputs:
    result = obj3.root_mean_squared_error(clean=False, multi_output=multi_output, decimal=5)
    print(f"n-D array, OOP style: {result}")

# Or run the simple:
python examples/RMSE.py

# The more complicated tests in the folder: examples

The documentation includes more detailed installation instructions and explanations.

Changelog

• See the ChangeLog.md for a history of notable changes to permetrics.

Important links

• Official source code repo: https://github.com/thieu1995/permetrics

• Official documentation: https://permetrics.readthedocs.io/

• Download releases: https://pypi.org/project/permetrics/

• Issue tracker: https://github.com/thieu1995/permetrics/issues

• This project also related to my another projects which are "meta-heuristics" and "neural-network", check it here

  • https://github.com/thieu1995/mealpy
  • https://github.com/thieu1995/opfunu
  • https://github.com/thieu1995/metaheuristics
  • https://github.com/chasebk

Metrics

Contributions

Citation

• If you use permetrics in your project, please cite my works:

@software{thieu_nguyen_2020_3951205,
  author       = {Thieu Nguyen},
  title        = {A framework of PERformance METRICS (PerMetrics) for artificial intelligence models},
  month        = jul,
  year         = 2020,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.3951205},
  url          = {https://doi.org/10.5281/zenodo.3951205}
}

@article{nguyen2019efficient,
  title={Efficient Time-Series Forecasting Using Neural Network and Opposition-Based Coral Reefs Optimization},
  author={Nguyen, Thieu and Nguyen, Tu and Nguyen, Binh Minh and Nguyen, Giang},
  journal={International Journal of Computational Intelligence Systems},
  volume={12},
  number={2},
  pages={1144--1161},
  year={2019},
  publisher={Atlantis Press}
}

Future works

Classification

• F1 score
• Multiclass log loss
• Lift
• Average Precision for binary classification
• precision / recall break-even point
• cross-entropy
• True Pos / False Pos / True Neg / False Neg rates
• precision / recall / sensitivity / specificity
• mutual information

HIGHER LEVEL TRANSFORMATIONS TO HANDLE

• GroupBy / Reduce
• Weight individual samples or groups

PROPERTIES METRICS CAN HAVE

• Min or Max (optimize through minimization or maximization)
• Binary Classification
  • Scores predicted class labels
  • Scores predicted ranking (most likely to least likely for being in one class)
  • Scores predicted probabilities
• Multiclass Classification
  • Scores predicted class labels
  • Scores predicted probabilities
• Discrete Rater Comparison (confusion matrix)