Industrial applications of Data Science

The objective of this thesis is twofold. The first objective is to detect duplicate addresses within a dataset containing addresses from 142 different countries, written in various languages. The primary challenge in this task is the lack of training data, which is addressed through two approaches: Active Learning, designed for efficient learning with minimal training data, and Un-supervised Learning techniques, such as Clustering and Dimensionality Reduction, applied to vector embeddings generated by Large Language Models. The second objective focuses on forecasting Alarm and Production Times of machines for a 14-day horizon, leveraging estimated daily production volumes. Since both target variables are interdependent on each other, various experiments are conducted using Statistical methods, Machine Learning, and Deep Learning models to explore the best possible way to forecast the target variables. The Temporal Fusion Transformer model is introduced as a key component of this research, offering promising capabilities for multi-horizon, multiple-time series forecasting and placing a strong emphasis on interpretability. Additionally, XGBoost, a powerful gradient-boosting algorithm, is used for its ability to excel in multi-time series forecasting tasks, along with Croston and TSB models that work well with intermittent data. Notably, the transformer model leverages additional features during training that may not be available during the forecasting phase.