How Uber’s Predictive Machine Learning Is Changing User Experience?

Uber has continually invested in AI over the years, and the technology is now widely used as a critical component throughout the company. “Uber has been an AI powered company for years and years and years,” said CEO Dara Khosrowshahi . Uber has made substantial use of artificial intelligence (AI) and machine learning (ML) to maintain its leadership in the transportation business. Uber operates in over 10,000 cities across more than 70 countries, handling 25 million trips daily with 137 million monthly active users. Machine learning (ML) is integral to almost every interaction within Uber’s apps, enhancing various aspects of the user experience and operational efficiency. Phase 1 (2016-2019): Predictive Machine Learning During this initial phase, Uber primarily focused on predictive machine learning for tabular data use cases. The company employed algorithms like XGBoost for critical tasks such as ETA predictions, risk assessment, and pricing. Uber also ventured into deep learning for 3D mapping and perception in self-driving cars. In 2016, Uber acquired Geometric Intelligence to create Uber AI Labs, demonstrating its commitment to advancing AI research. Gary Marcus , founder of Geometric Intelligence, stated: “We live in an era where data is central to almost every major company and for almost any of them, even incremental improvements in artificial intelligence and machine learning can have immediate, significant payoffs,” Uber began its machine learning (ML) journey in 2015, with applied scientists developing models in Jupyter Notebooks™ and engineers designing customised pipelines for production deployment. Initially, there was no standardised approach for developing reliable and reproducible pipelines for large-scale training and prediction operations, nor was there a simple mechanism to store and compare training experiment findings. Putting a model into production necessitated the creation of a specialised serving container. Uber developed Michelangelo in early 2016 to standardise machine learning procedures, allowing developers to construct and deploy ML models on a large scale. Michelangelo addressed the issues of scalable model training and deployment, and later launched Palette, a feature repository that allows teams to organise and share feature pipelines in batch or near-real time. Palette now includes over 20,000 characteristics that Uber teams may use to develop effective ML models. Key Michelangelo components include: Gallery : A model and ML metadata registry with a comprehensive search API. Manifold : A model-agnostic visual debugging tool for ML. PyML : A framework to expedite prototyping, validating, and productionizing Python ML models. Horovod : A tool for distributed training. Phase 2 (2019-2023): Deep Learning and Collaborative Model Development Between 2019 and 2023, Uber dramatically improved its AI and ML capabilities by adopting deep learning (DL) and encouraging collaborative model creation for high-impact ML initiatives. During this revolutionary period, Uber shifted from considering model iteration as code within a centralised ML monorepo to supporting DL as a first-class citizen in Michelangelo, the company's centralised ML platform. Michelangelo received significant improvements, including broad support for deep learning models, infrastructure for training, deploying, and monitoring DL models, and a unified user interface (UI) and code-driven experience. This flexibility benefited both traditional ML and deep learning developers, allowing for more efficient model construction. Uber also established Project Canvas, which applies software engineering approaches to machine learning development. This guaranteed that all source code, configurations, and dependencies were kept in the ML monorepo, which allowed for change tracking, code reviews, and unit testing. The project provided both local and remote training, allowing developers to debug and test programmes locally before submitting them to remote clusters for complete training. By the end of this phase, more than 60% of Uber's tier-1 models had included deep learning in production, considerably improving model performance across a range of applications. These included ETA calculation, rider-driver matching, fraud detection, and personalised recommendations, all of which saw significant accuracy and efficiency gains using deep learning models. Additionally, Uber's Customer Obsession Ticket Assistant (COTA) used machine learning and natural language processing (NLP) to improve customer service. COTA, which was built on Michelangelo, helped address more than 90% of inbound support queries in a timely and effective manner. When a new ticket was created in the customer service platform, the appropriate features were gathered and delivered to Michelangelo's ML model. The model predicted scores for each feasible solution, which were later kept and accessed by the backend service. Agents were presented with the top three alternatives

How Uber’s Predictive Machine Learning Is Changing User Experience?

Phase 1 (2016-2019): Predictive Machine Learning

Keep reading. It's free.

Related stories