Peter Papai

Fraud detection systems often work in a human-in-the-loop fashion, where an ML model suggests suspicious transactions for review by a human expert. My goal was to minimize the number of reviews without increasing the number of uncaught fraudulent transactions. I built a model to predict which transaction would be likely to get rejected by a human expert. I was responsible for creating features and training the model using Scala and Spark ML in a batch job that ran daily. I deployed the model online using the MLeap framework. The next iteration of the model was implemented in Python and exported in PMML format for online deployment. I was responsible for A/B testing the model as well. Because it entailed comparing costs, variance reduction techniques had to be used to measure the impact of the new algorithm. The new algorithm reduced the workload of the human experts by one-third.

Came up with the idea to apply counterfactual techniques to find the optimal room prices. By extending the decision tree of Spark ML, I implemented a greedy algorithm that maximizes margin. I designed and evaluated A/B tests to validate the approach. Later, I implemented iterative improvements of the original idea in Python. The company attributed a 20% increase in profit to the success of this project.

Had to work around a legacy system when designing an algorithm for predicting the click-through rate for banner ads. Using an off-the-shelf algorithm was out of the question. We also faced a cold-start problem because the company was fairly new, and the data was very sparse. I implemented a tree augmented naïve Bayesian model with hierarchical CTR smoothing. It was the 1st model we deployed, utilizing our system architecture and data collection practices to the fullest.

Implemented a factorization machine-based model that used context features to estimate the conversion rate for user-hotel pairs. I extended Scala and Spark ML to train the model. Cross-validation happened on several metrics from information retrieval, e.g., NDCG or binary hit rate. This model was part of a model stack that generated a 1 – 2% increase in margin for the company.

For matching hotels from suppliers' inventory to hotels in our own inventory, I used techniques from NLP (TF-IDF, character CNN) to create an embedding for hotel records. The hotels were compared in this embedding space. The model was deployed as a daily batch job. It increased the percentage of matching hotels from around 80% to around 90%. Removing duplicates from our search page significantly improved the user experience based on our internal UX research.

Led the work on the ML component of an internal tool to monitor the health of our IT systems. The modeling involved forecasting quantiles for hundreds of time series in a fully autonomous way. The models were written in Python using scikit-learn, NumPy, SciPy, and PyTorch. The deployment happened via Docker containers. The loss due to technical errors decreased by 90% in the quarter following the deployment.