Is my AI ethic ?
TransparentAI is a toolbox in Python to answer the question "Is my AI ethic ?" based on the European Commission requierements.
The research of ethic in the Artificial Intelligence field is a hot topic. More than 70 papers between 2016 and 2019 (Algorithm Watch, 2019). But many papers just present the question of "What should be an ethic AI" not the "How to do it". In consequence, many developers become fustrated and still don't really know to do it in practive (Peters, May 2019).
TransparentAI is an answer to this question. The philosophy is that, in coherence with the Ethics Guidelines for Trustworthy AI by the European Commission, you can find out easily (in Python) if "your AI is ethic" !
New tool : This is a new tool so if you found any bugs or other kind of problems please do not hesitate to report them on the issues GitHub page from the library here : https://github.com/Nathanlauga/transparentai/issues. I hope you will enjoy this tool and help me to improve it!
Documentation is available here : API Documentation.
- Installation
- Compatible model and data type
- Getting started
- UE Commision requirements
- Contributing
- Credits and ressources
- Author
- License
You can install it with PyPI :
pip install transparentai
Or by cloning GitHub repository
git clone https://github.com/Nathanlauga/transparentai.git
cd transparentai
python setup.py install
Version 0.2 :
| Objects | What the tool can handle |
|---|---|
| Data | can only handle tabular dataset (numpy array, pandas DataFrame) |
| Model | can only handle classification and regression model |
Coming for version 0.3 :
- Data : Explore Image dataset, Text dataset
- Model : Clustering model
Take a look on the Getting started page of the documentation or you can search specific use cases in the examples/ directory.
In this section I created a binary classifier based on Adult dataset. The following variables will be used :
| variable | description |
|---|---|
data |
Adult dataset as DataFrame |
clf |
Classifier model |
y_true |
True labels for train set |
y_true_valid |
True labels for valid set |
y_pred |
Predictions labels for train set |
y_pred_valid |
Predictions labels for valid set |
df_valid |
Dataframe for valid set |
X_train |
Features for train set |
X_valid |
Features for valid set |
privileged_group = {
# For gender attribute Male peoples are considered to be privileged
'gender':['Male'],
# For marital-status attribute Married peoples are considered to be privileged
'marital-status': lambda x: 'Married' in x,
# For race attribute White peoples are considered to be privileged
'race':['White']
}
from transparentai import fairness
fairness.model_bias(y_true_valid, y_pred_valid, df_valid, privileged_group)Output:
{
"gender": {
"statistical_parity_difference": -0.07283528047741014,
"disparate_impact": 0.4032473042703101,
"equal_opportunity_difference": -0.04900038770381182,
"average_odds_difference": -0.026173142849183567
},
"marital-status": {
"statistical_parity_difference": -0.11667610209029305,
"disparate_impact": 0.27371312304160633,
"equal_opportunity_difference": 0.08345535064884008,
"average_odds_difference": 0.03867329810319946
},
"race": {
"statistical_parity_difference": -0.0420778376239787,
"disparate_impact": 0.5964166117990216,
"equal_opportunity_difference": -0.0004408949904296522,
"average_odds_difference": -0.002870373184105955
}
}
This metrics can be not easy to understand so you can use the returns_text=True so that you can get ths insight :
fairness_txt = fairness.model_bias(y_true_valid, y_pred_valid, df_valid, privileged_group, returns_text=True)
print(fairness_txt['gender'])Output:
The privileged group is predicted with the positive output 7.28% more often than the unprivileged group. This is considered to be fair.
The privileged group is predicted with the positive output 2.48 times more often than the unprivileged group. This is considered to be not fair.
For a person in the privileged group, the model predict a correct positive output 4.90% more often than a person in the unprivileged group. This is considered to be fair.
For a person in the privileged group, the model predict a correct positive output or a correct negative output 2.62% more often than a person in the unprivileged group. This is considered to be fair.
The model has 3 fair metrics over 4 (75%).
And if you like to get visual help use the plot_bias function :
privileged_group = {'gender': ['Male']}
from transparentai import fairness
fairness.plot_bias(y_true_valid, y_pred_valid, df_valid, privileged_group, with_text=True)
from transparentai.models import explainers
explainer = explainers.ModelExplainer(clf, X_train, model_type='tree')
explainer.explain_global_influence(X_train, nsamples=1000)Output:
{
'age': 0.08075649984055841,
'fnlwgt': 0.05476459574744569,
'education-num': 0.08048316800088552,
'capital-gain': 0.06879137962639843,
'capital-loss': 0.018367250661071737,
'hours-per-week': 0.06009733425389803
}
explainer.plot_global_explain()
explainer.plot_local_explain(X_valid.iloc[0])
from transparentai.models import classification
# You can use custom function with lambda
metrics = ['accuracy', 'roc_auc', 'f1', 'recall', 'precision', lambda y_true, y_pred: sum(y_true-y_pred)]
classification.compute_metrics(y_true_valid, y_pred_valid, metrics)Output:
{
'accuracy': 0.812011415808413,
'roc_auc': 0.8272860034692258,
'f1': 0.5682530635508691,
'recall': 0.5244608100999474,
'precision': 0.6200248756218906,
'custom_1': 586
}
classification.plot_performance(y_true, y_pred, y_true_valid, y_pred_valid)
from transparentai.datasets import variable
variable.plot_variable(data['age'])
variable.plot_variable(data['capital-loss'], legend=data['income'], ylog=True)
variable.plot_variable(data['workclass'])
The birthdate column was generated based on the age column.
variable.plot_variable(data['birthdate'], legend=data['income'])
timestamp variable was generated randomly, it represents the time of the prediction.
from transparentai import monitoring
monitoring.plot_monitoring(y_true, y_pred, timestamp, interval='month', classification=True)
Estimate your training CO2 consumption.
from transparentai import sustainable
sustainable.estimate_co2(hours=24, location='France', watts=250)Output:
3.18437946896484
Evaluate your training kWh consumption.
from transparentai import sustainable
kWh, clf = sustainable.evaluate_kWh(clf.fit, X, Y, verbose=True)Output:
Location: France
Baseline wattage: 4.79 watts
Process wattage: 18.45 watts
--------------------------------------------------------------------------------
------------------------------- Final Readings -------------------------------
--------------------------------------------------------------------------------
Average baseline wattage: 3.53 watts
Average total wattage: 16.04 watts
Average process wattage: 12.51 watts
Process duration: 0:00:07
--------------------------------------------------------------------------------
------------------------------- Energy Data -------------------------------
--------------------------------------------------------------------------------
Energy mix in France
Coal: 3.12%
Petroleum: 16.06%
Natural Gas: 33.56%
Low Carbon: 47.26%
--------------------------------------------------------------------------------
------------------------------- Emissions -------------------------------
--------------------------------------------------------------------------------
Effective emission: 1.32e-05 kg CO2
Equivalent miles driven: 5.39e-12 miles
Equivalent minutes of 32-inch LCD TV watched: 8.14e-03 minutes
Percentage of CO2 used in a US household/day: 4.33e-12%
--------------------------------------------------------------------------------
------------------------- Assumed Carbon Equivalencies -------------------------
--------------------------------------------------------------------------------
Coal: 995.725971 kg CO2/MWh
Petroleum: 816.6885263 kg CO2/MWh
Natural gas: 743.8415916 kg CO2/MWh
Low carbon: 0 kg CO2/MWh
--------------------------------------------------------------------------------
------------------------- Emissions Comparison -------------------------
--------------------------------------------------------------------------------
Quantities below expressed in kg CO2
US Europe Global minus US/Europe
Max: Wyoming 2.85e-05 Kosovo 2.93e-05 Mongolia 2.86e-05
Median: Tennessee 1.40e-05 Ukraine 2.04e-05 Korea, South 2.34e-05
Min: Vermont 8.00e-07 Iceland 5.26e-06 Bhutan 3.26e-06
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
Process used: 3.10e-05 kWh
import transparentai.utils as utils
utils.check_packages_security(full_report=True)Output:
+==============================================================================+
| |
| /$$$$$$ /$$ |
| /$$__ $$ | $$ |
| /$$$$$$$ /$$$$$$ | $$ \__//$$$$$$ /$$$$$$ /$$ /$$ |
| /$$_____/ |____ $$| $$$$ /$$__ $$|_ $$_/ | $$ | $$ |
| | $$$$$$ /$$$$$$$| $$_/ | $$$$$$$$ | $$ | $$ | $$ |
| \____ $$ /$$__ $$| $$ | $$_____/ | $$ /$$| $$ | $$ |
| /$$$$$$$/| $$$$$$$| $$ | $$$$$$$ | $$$$/| $$$$$$$ |
| |_______/ \_______/|__/ \_______/ \___/ \____ $$ |
| /$$ | $$ |
| | $$$$$$/ |
| by pyup.io \______/ |
| |
+==============================================================================+
| REPORT |
| checked 77 packages, using default DB |
+==============================================================================+
| No known security vulnerabilities found. |
+==============================================================================+
The European Commission defined seven requirements that allow to make a trustworthy AI.
These requirements are applicable to different stakeholders partaking in AI systems’ life cycle: developers, deployers and end-users, as well as the broader society. By developers, we refer to those who research, design and/or develop AI systems. By deployers, we refer to public or private organisations that use AI systems within their business processes and to offer products and services to others. End-users are those engaging with the AI system, directly or indirectly. Finally, the broader society encompasses all others that are directly or indirectly affected by AI systems. Different groups of stakeholders have different roles to play in ensuring that the requirements are met:
- Developers should implement and apply the requirements to design and development processes.
- Deployers should ensure that the systems they use and the products and services they offer meet the requirements.
- End-users and the broader society should be informed about these requirements and able to request that they are upheld.
The below list of requirements is non-exhaustive. 35 It includes systemic, individual and societal aspects:
- Human agency and oversight: Including fundamental rights, human agency and human oversight
- Technical robustness and safety: Including resilience to attack and security, fall back plan and general safety, accuracy, reliability and reproducibility
- Privacy and data governance: Including respect for privacy, quality and integrity of data, and access to data
- Transparency: Including traceability, explainability and communication
- Diversity, non-discrimination and fairness: Including the avoidance of unfair bias, accessibility and universal design, and stakeholder participation
- Societal and environmental wellbeing: Including sustainability and environmental friendliness, social impact, society and democracy
- Accountability: Including auditability, minimisation and reporting of negative impact, trade-offs and redress.
This table allows you to in details for each requirements and if it's possible how to control if the aspect is ethic with TransparentAI. Some aspects do not have technical implementation in this tool because it requires legal or other knowledge. If you want to understand the differents aspect and requirements you can read details in the Ethics Guidelines for Trustworthy AI paper.
| UE requirements | Aspect | TransparentAI implementation |
|---|---|---|
| 1. Human agency and oversight | Fundamental rights | No technical implementation. |
| Human agency | No technical implementation. | |
| Human oversight | Control AI performance over time with monitoring.monitor_model or monitoring.plot_monitoring |
|
| 2. Technical robustness and safety | Resilience to attack and security | Try different input scenario in the model to see how it handles it with models.explainers.ModelExplainer |
| Fallback plan and general safety | Check if your Python's package are secure with utils.check_packages_security |
|
| Accuracy | Validate your AI performance with models.classification.plot_performance or models.regression.plot_performance |
|
| Reliability and Reproducibility | No technical implementation. | |
| 3. Privacy and data governance | Privacy and data protection | No technical implementation. |
| Quality and integrity of data | Check if the variable is coherent in its distribution with datasets.variable.plot_variable |
|
| Access to data | No technical implementation. | |
| 4. Transparency | Traceability | Generate a performance validation report with utils.reports.generate_validation_report |
| Explainability | Explain the local or global behavior of your model with models.explainers.ModelExplainer |
|
| Communication | No technical implementation. | |
| 5. Diversity, non-discrimination and fairness | Avoidance of unfair bias | Check if your AI is biased on protected attributes with fairness.model_bias or fairness.plot_bias |
| Accessibility and universal design | No technical implementation. | |
| Stakeholder Participation | No technical implementation. | |
| 6. Societal and environmental well-being | Sustainable and environmentally friendly AI | Get the kWh value of the AI training with utils.evaluate_kWh |
| Social impact | Check if your AI is biased on protected attributes with fairness.model_bias or fairness.plot_bias |
|
| 7. Accountability | Auditability | Generate a performance validation report with utils.reports.generate_validation_report |
| Minimisation and reporting of negative impacts | No technical implementation. | |
| Trade-offs | No technical implementation. | |
| Redress | No technical implementation. |
See the contributing file.
PRs accepted.
For this submodule, I have to say I was mainly inspired by one tool so all the credit has to be attributed to AIF360 by IBM.
I used some of the metrics proposed in the tools (Statistical Parity Difference, Equal Opportunity Difference, Average Odds Difference, Disparact Impact and Theil Index).
I used some metrics function of the scikit-learn Python package.
I choose to used the Shap library because this tool was tested and aproved by a lot of people in the community, and even if I found some papers showing some problems (e.g. "Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods" (Slack and al., November 2019)), I decided to use it because if you want to biased Shap, you have to do it intentionally at the AI creation.
I was inspired by some graphics on Kaggle. But mainly I use some code on matplotlib website and the Python graph gallery.
I used different packages that implement great features such as :
energyusage: A Python package that measures the environmental impact of computation.safety: Safety checks your installed dependencies for known security vulnerabilities.
Again thanks to researchers and developers that contributed in this really important field, without them I don't think I'll be able to create this tool.
This work is led by Nathan Lauga, french Data Scientist.
This project use a MIT License.
Why ?
I believe that the code should be re-used for community projects and also inside private projects. AI transparency needs to be available for everyone even it's a private AI.