LITERATURE REVIEW
There is a great need to understand how things occur. Hence the need for
comprehensibility or interpretability. Interpretation can be said to be
the action of explaining something to know how that aspect works, why
the part behaves in a particular manner and any impact that the element
in question can have on attributes depending on it. In machine learning,
interpretability is highly important nowadays as these models run vital
critical processes worldwide in financial, academic, and political
scenarios.
Why is interpretability important? From the literature reviewed,
interpretability is essential based on TWO reasons. One aspect of being
a human being is that we need to put an explanation to every event that
occurs in our lives. This need thus has driven the urge to be able to
explain certain causal circumstances and why they informed a particular
decision. This need is equally essential in machine learning as we need
to explain why a particular model chooses a specific prediction instead
of another. Thus the interpretability of machine learning models can aid
human beings in learning new aspects and also find meaning in their
decisions. (C Molnar 2019)
The other is handling uncertainty and biases in decisions. Much research
has developed new models to handle uncertainty and bias in machine
learning models. Certainly, interpretability is an important aspect that
can be used to address discrimination and uncertainty. Machine learning
models pick preferences from training data caused by missing values in
some datasets, which leads to uncertainty in these models when it comes
to predicting the outcomes based on the training data they have. This
makes the models unstable and gives outrageous predictions that cannot
be explained. Thus interpretability can be used to handle this issue and
offer a reprieve for models and their developers. (C Molnar 2019)
Review of literature by Christoph Molnar in his book ”Interpretable
machine learning,” we see that interpretable machine learning models can
be classified according to FOUR fronts. The figure below illustrates the
classification of interpretable machine learning models.