STAGE Junior Data Scientist

Machine learning algorithms deployed in real-world situations often have lower
performances than the ones demonstrated during R&D. This can be due to a
combination of factors: the training dataset is not representative of all the possible
cases, overfitting of the evaluation dataset, the model is under-specified by the available
data, the real-world data has a distribution slightly different In medical applications, a
drop of performances can have dire consequences and should be mitigated as much as
possible. This internship goal is to tackle part of this problem: how to detect if a new
example is an anomaly and if a prediction should be made on it?
Anomaly (or outliers) detection is the identification of examples that do not fit the
distribution of the majority of the data. In our case, this would allow us to identify
samples that have been created using a different protocol and where the diagnosis
algorithm would not be meaningful. A large literature exists on the subject going from
unsupervised methods to fully-supervised ones.
Your tasks will be to: Make a complete bibliography of the subject Choose and implement state-of-the-art methods on toy data Test the chosen methods on VitaDX real problem