A framework to measure how ethic, equitable and integrated is an AI algorithm
The solution presented is a framework that allows to measure and explain the predictive behaviour of Artificial Intelligence (AI) algorithms. This model provides transparency of the actions carried out by the algorithms that are analysed, it tries to confirm that the decision is equitable and integrated and that it does not present bias due to any characteristic of the input datasets, so the algorithm is Ethic. The model has previously been tested with two proofs of concept (PoC). In the first, a book recommendation algorithm was analysed and in the second, a incidents, inquiries and complaints algorithm was analysed. We are currently conducting a new PoC that analyses an algorithm for detecting cancerous melanomas on the face.
Problem or opportunity
AI presents a series of problems that are inherent in its construction, behaviour, and results. Another factor that adds to this problem is the use and application of AI, which is increasingly broad, covering all fields of knowledge, business sectors, etc. With the increasingly widespread use of AI, it is important to resolve the Ethics, Transparency and Fairness of algorithms and their predictions. The problem we want to solve is when a decision is made by a machine, two requirements must be fulfilled:
- There must be a balance between the dataset used and the programming of the algorithm, which avoids discrimination and bias, and provides as much fairness as possible. If the algorithm is trained with data that has bias, it will cause the results obtained to be discriminatory.
- The algorithm used must meet transparency conditions, i.e., the result obtained must be explainable to any user in a relatively simple way. When an algorithm decides about a subject, a person or an entity, this subject has the right to know what process has been followed to make that decision.
The challenge facing AI is to create explanations that are both complete and interpretable, and it is difficult to achieve interpretability and completeness at the same time.
The application of this model will allow to solve the explainability and integrity of any decision taken by an AI algorithm, no longer black boxes and the decision will come with an argumentation and justification of the decision that will include the absence of discrimination and bias.
This model can help lay the foundations for ethical certification of AI algorithms that make decisions.
The solution addresses a complex problem. A set of techniques have been used that resolve partial parts that together form a whole system that responds to the problems and needs explained in the previous points. The technologies used are Artificial Intelligence (AI), NLP and a set of visualization techniques for more agnostic models, techniques for transparency, techniques for discrimination and bias. PCA (which is a dimension reduction technique focused on retaining those dimensions/components that explain as much of the variance in the data as possible) or T-SNE (which is another dimension reduction technique more focused on retaining close point distances and is used as an alternative to PCA as it can capture more complex patterns). To address transparency and fairness, we started with dimension reduction techniques applying the TruncatedSVD, PCA or T-SNE method as the case it may be. From here, the predictive models were reviewed to evaluate and ensure the behaviour, specifically the prediction, contrasting the results according to the prepared input datasets. To analyse the discrimination tests were applied: z-test, Wilcoxon, Sum Rank and Signed Rank. To analyse the bias, mapping techniques between vocabulary and real number vectors Word2vec, doc2vec (from Google) and FasText (from Facebook) were applied. In many cases, there is discrimination linked to bias and it is necessary to apply the so-called tests jointly and/or sequentially.