Post by Yichuan (Michael) Cai

Thanks Paarth Chothani as the podcast host and Uber ML platform team's support as well as Manoj Panikkar Bob Zheng Dhruv Ghulati from Uber AI and other teammates from Uber to build all the relevant products This video explains how Uber developed forecasting models to improve the airport experience for both drivers and riders and market efficiency . Key points generated from Youtube by #Gemini include: - The Airport Challenges : Unlike typical Uber operations, airports involve unique dynamics such as designated staging lots and contracts, requiring drivers to make a conscious decision to go to the airport and wait and solve airport supply deficit challenges . - Estimated Time to Request (ETR) : To address driver concerns about wait times, Uber developed a product displaying the ETR which provides drivers with information on how long they might have to wait for a trip at the airport, empowering them to make informed decisions and building trust. - Complexity of ML in Physical Spaces : Highlight that machine learning in the physical world is significantly more complex than in the digital world. Factors like curb space, varied airport layouts, and the high cost of error (e.g., drivers wasting time) necessitate robust and accurate models. - Technical Approach and Evolution: The team started with simple heuristics and gradually evolved their models. They began with XGBoost for ETR predictions and later incorporated more advanced models like Gaussian Mixture Models to predict both mean and variance. Eventually, they adopted a Transformer architecture for its ability to handle parallel training, capture long-range dependencies, dynamic weighting, and multivariate data. - Data as a First-Class Citizen: Emphasizing data quality, the speakers stressed that accurate incoming data is crucial for model performance. They found that improving data quality yielded better ROI than solely focusing on model complexity. - Feature Engineering: The models incorporate various real-time and external data sources, including flight information (delays, passenger) from providers like OAG, and even weather data, to accurately predict driver demand and rider conversion rates. They also discussed the importance of feature importance analysis and making trade-offs on features based on their impact and the cost of maintaining their data pipelines. - Model Design Decisions and Data Quality: A significant improvement in precision was achieved by transitioning from a stacked model (which amplified errors) to a single, end-to-end model for predicting wait times. They also emphasized the importance of cross-functional collaboration to ensure data quality and alignment on metric definitions across different teams. - Global vs. Partition Models : The choice between building a global model or partition models depends on data availability. Smaller airports with sparse data often necessitate a global model initially, with potential for heuristic customizations. #ml #AI #transformer