What if it were possible to read the outcomes of a Machine Learning (ML) model simply and within the context of each problem?

“White boxes” and “black boxes” are terms used to describe ML algorithms. “Black boxes” are highly complicated models that are challenging to understand, even for specialists in the field. On the other hand, “white boxes” represent simpler models that data scientists can understand. However, neither of these models is easily understandable to those outside the field, making it difficult for project stakeholders to understand and accept the models’ decisions.

To close this gap, the field of Explainable AI has been developed. It intends to enhance transparency and understanding of model decisions, ultimately leading to greater confidence and security when applying ML models in critical domains. (IBM, n.d.)

To aid in the interpretation of models, more and more Python libraries and techniques are becoming available. Currently, SHAP, LIME, and ELI5 are the most widely used. These Python libraries help us comprehend the significance of each variable for the predictions made. Interpreting the meaning of each contribution (positive vs. negative value, high vs. low values) and what they mean in the context of each problem/business remains challenging even with the existence of these libraries.

But what if these contributions could be translated into an appropriate language for each challenge, making them easier for people who do not know ML?

The approach entails merging a Large Language Model (LLM) with the LIME library to comprehend the model’s decisions better and explain each decision in the context of the challenge.

A LLM is a model that uses advanced natural language processing (NLP) algorithms to understand and generate text. These models are trained on large volumes of textual data, such as books and web articles, to learn how language is structured and how words and phrases relate to each other. A famous example of an LLM is the GPT (Generative Pre-trained Transformer). (Souza, 2023)

And how is this possible?

By combining the problem context with the results from the LIME library, we can input this information into an LLM prompt. This enables us to translate complex numerical contributions into simpler explanations that fit the specific context of each problem or business.

Let’s analyse an example to help you grasp this proposition a little better: a classification problem with an ML model that uses clinical factors such as blood pressure, glucose, insulin, BMI, diabetes pedigree function (the likelihood of acquiring diabetes based on family history), age, number of pregnancies, and skin thickness to predict whether a patient would develop diabetes.

Let’s see this individual:

Why did the ML model predict this person will not get diabetes?

This is how the LIME library tackles it. It is clear to us that the model indicates this person has a 63% chance of not getting diabetes. However, what does the 0.19 contribution for glucose levels below 96 mean? It becomes challenging to evaluate the results with just these values.

We obtained a report on the interpretation of the ML model using an LLM from OpenAI, the GPT-3.5 model and provided the theoretical context of the problem. In this case, the LLM was instructed that this was a classification problem to predict predisposition to diabetes and that the goal was to provide an interpretation of the results obtained. The obtained report was as follows:

“The model’s interpretation of these factors is as follows:

– Glucose <= 96.00: This has the highest negative contribution of -0.193, indicating that glucose levels below 96 mg/dL are associated with a lower risk of developing diabetes.

– 111.25 < Insulin <= 125.00: Insulin levels between 111.25 and 125.00 have a positive contribution of 0.073, suggesting that higher insulin levels increase the risk of developing diabetes.

– BMI > 37.52: A BMI greater than 37.52 has a positive contribution of 0.067, indicating that obesity is a risk factor for diabetes.

– DiabetesPedigreeFunction <= 0.25: A lower Diabetes Pedigree Function, with a contribution of -0.061, suggests a lower genetic predisposition to diabetes.

– Age between 23.00 and 31.00: This age range has a negative contribution of -0.041, indicating that younger age is associated with a lower risk of developing diabetes.

– Pregnancies between 3.00 and 6.75: This range has a negative contribution of -0.014, suggesting that having had between 3 and 6.75 pregnancies is associated with a lower risk of developing diabetes.

– 72.41 < BloodPressure <= 80.00: Blood pressure levels between 72.41 and 80.00 have a small positive contribution of 0.004, indicating that higher blood pressure may be a risk factor for diabetes.

– SkinThickness between 29.00 and 29.15: This range has a very small negative contribution of -0.001, suggesting that skin thickness in this range is not a significant risk factor for diabetes.

Overall, the model suggests that this patient has a moderate risk of developing diabetes, with the most significant risk factors being higher glucose levels, higher insulin levels, and obesity. However, the patient’s younger age, lower genetic predisposition, and lower number of pregnancies may be protective factors. It is important to note that these are statistical correlations and should be validated with medical advice. Nevertheless, the model’s contributions provide valuable insights into the factors that may be relevant for predicting diabetes risk in this patient.”

Through the LLM’s response, stakeholders can determine whether this prediction is clinically reasonable and comprehend what led the model to anticipate that the patient will not have diabetes.

This simplification in expressing the results eliminates the concept of “black boxes,” resulting in transparency and accountability in model usage. This allows all stakeholders to comprehend and use the results without concern.

The range of problems to which this solution can be used is increased by the fact that the LIME library can deal with image classifiers, regression, and classification problems, among other problem types.

In a world where Artificial Intelligence and Machine Learning are increasingly prominent solutions, the ability to interpret “black boxes” facilitates the development of ethical and responsible behaviour and the useful application of these technologies for the good of humanity.

BI4ALL aims to be at the forefront of this path. If you are interested in enhancing your business’s Data Science and Artificial Intelligence capabilities, we are available to discuss your project.

Bibliography

IBM. (n.d.). Retrieved from https://www.ibm.com/topics/explainable-ai

Souza, A. (2023, julho). Medium. Retrieved from https://medium.com/blog-do-zouza/tudo-o-que-voc%C3%AA-precisa-saber-sobre-llm-large-language-model-a36be85bbf8f

state, A. (n.d.). Retrieved from https://www.activestate.com/blog/white-box-vs-black-box-algorithms-in-machine-learning/