Projects
Recommendation Systems
As part of my job as a DataScientist at MercadoLibre I’m currently working on developing a Deep-Learning-based recommendation system to give real-time customized recommendations to our users based on their recent navigation history. We have framed the problem as a sequence model where the task is to predict the user’s next purchase. We use an attention-based RNN to interpret the user’s navigation intent and, based on this, rank a pool of heuristically generated candidates that will serve as a recommendation.
Natural Language Processing
I have worked in different projects involving Natural Language Processing. Most recently, we have developed a questions-classifier for MercadoLibre buyers’ questions. We took advantage of a state of the art Language Model that then we trained using custom data to capture the domain-specific knowledge. We used Transfer Learning to fine tune the model weights for the classification tasks, needing a relatively small amount of labeled data to achieve a satisfactory performance. The classifier was able to capture some pretty complex examples in the data.
Healthcare applications of Machine Learning
In the context of my Master’s thesis, I approached a critical problem in the healthcare sector from a Machine Learning perspective. Together with the Operations Management team of one of the biggest hospitals in Buenos Aires, we developed a tool to forecast discharges in the following 24 hours in order to provide management with crucial information to optimize the allocation of the free beds. The problem was modelled as classification task where the objective was predicting whether a patient was going to be discharged in the following 24 hours or not. The model incorporated the full patient’s clinic records and was able to predict the odds of a discharge during the day with a satisfactory performance (AUC Roc of 0.85). The most challenging aspect of the project was gathering the data coming from the different highly unstandardized databases of the hospital, making sense of it and processing it in order to get a useful dataset for the classification task at hand.
You can access the full thesis document here
Customer Segmentation and Life Time Value
As part of my job as a DataScientist at Despegar.com I worked intensively on analyzing, assesing and iterating the company’s customer segmentation model. Eventually, this model evolved into a LTV model, whose objective was to predict the travel expenses in which users were going to incur during the course of a year. We developed a classifier that would allocate users into one out of ten buckets according to their expected travel expenses. These predictions were then used to build marketing campaing audiences that allow to bid and communicate appropriately based on the prospective LTV of the customers being targeted.