We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Real-Time Forecasting Faceoff: Time Series vs. DNNs // Josh Xi // #305

Real-Time Forecasting Faceoff: Time Series vs. DNNs // Josh Xi // #305

2025/4/11
logo of podcast MLOps.community

MLOps.community

AI Deep Dive AI Chapters Transcript
People
J
Josh Xi
Topics
Josh Xi: 我在Lyft的市场团队工作，负责构建预测模型来平衡供需。我们使用实时预测模型，对每个地理哈希单元（约1平方英里）进行未来5分钟到1小时的预测。这需要处理数百万个数据点，并整合外部数据源，如事件和天气数据，但这些数据整合起来具有挑战性。我们发现，传统的基于时间序列的预测方法（如自回归模型）在实际应用中表现优于深度神经网络 (DNN)。时间序列模型更易于解释，更容易进行人工干预调整，并且在处理实时、高频数据时更有效率。我们每分钟对模型进行再训练，以保持高精度，尽管这会带来高昂的计算成本。我们也尝试了时间序列基础模型，在某些用例中效果良好，但在实时、高粒度预测方面不如传统的自回归模型。在模型训练和服务方面，自回归模型的成本远低于DNN，并且更容易进行在线再训练，从而更好地适应市场动态。我们通过在线再训练、集成多个模型以及使用启发式规则进行偏差校正来提高预测精度。我们计划将区域级离线预测模型与实时预测模型结合，以提高精度并更好地处理特殊事件。处理时空数据的模型需要考虑空间相关性，这增加了模型的复杂性和训练难度。由于人类行为的复杂性和变化性，准确预测出行需求具有挑战性。我们主要关注区域层面的异常值分析，并通过自动化的模型调整和偏差校正来处理预测误差。 Demetrios: (主要以提问和引导讨论为主，没有形成具体的核心论点)

Deep Dive

Chapters
This chapter introduces Lyft's real-time forecasting system, focusing on the challenges of balancing supply and demand in a dynamic market. The discussion highlights the use of geohashes for granular forecasting and the importance of incorporating external data sources.
  • Lyft uses real-time forecasting to manage supply and demand.
  • Forecasting is done at the geohash level (e.g., GeoHash 6, approximately one square mile).
  • External data sources (events, weather) are incorporated but pose challenges.

Shownotes Transcript

Real-Time Forecasting Faceoff: Time Series vs. DNNs // MLOps Podcast #305 with Josh Xi, Data Scientist at Lyft.

Join the Community: https://go.mlops.community/YTJoinIn Get the newsletter: https://go.mlops.community/YTNewsletter

// AbstractIn real-time forecasting (e.g. geohash level demand and supply forecast for an entire region), time series-based forecasting methods are widely adopted due to their simplicity and ease of training. This discussion explores how Lyft uses time series forecasting to respond to real-time market dynamics, covering practical tips and tricks for implementing these methods, an in-depth look at their adaptability for online re-training, and discussions on their interpretability and user intervention capabilities. By examining these topics, listeners will understand how time series forecasting can outperform DNNs, and how to effectively use time series forecasting for dynamic market conditions and decision-making applications.

// BioJosh is a data scientist from the Marketplace team at Lyft, working on forecasting and modeling of marketplace signals that power products like pricing and driver incentives. Josh got his PHD in Operations Research in 2013, with minors in Statistics and Economics. Prior to joining Lyft, he worked as a research scientist in the Operations Research Lab at General Motors, focusing on optimization, simulation and forecasting modeling related to vehicle manufacturing, supply chain and car sharing systems.

// Related LinksWebsite: https://www.lyft.com/