The Value of Big Data in Machine Learning: Detailed Illustration and Insights

The lecture introduces the concept of big data's value for machine learning, emphasizing that large, well‑segmented datasets allow more precise characterizations of user groups and avoid misleadingly broad conclusions.

Through a case study of 3,000 customers' shoe preferences, it shows how overly aggregated statistics (e.g., 50% of Chinese women like high heels) mask diverse sub‑populations, while overly granular slices (e.g., a single 5‑10‑year‑old girl) suffer from insufficient sample size, leading to untrustworthy results.

Two conflicting requirements arise: the desire for detailed segmentation to increase accuracy, and the need for enough samples in each segment to maintain statistical confidence. The article argues that big data resolves this tension by providing massive sample volumes that keep every fine‑grained cell sufficiently populated.

It then presents the second value of big data—smarter learning—by tracing four stages of machine learning evolution: from domain‑knowledge‑driven hypothesis testing, through statistical language parsing, to deep learning that relies primarily on massive data with minimal handcrafted features.

The discussion highlights how overfitting and underfitting relate to data volume and feature engineering, and how abundant data enables models to learn complex feature‑to‑target relationships without heavy reliance on prior domain knowledge.

In conclusion, big data empowers machine learning by delivering both finer data description and more intelligent learning, ultimately allowing systems to automate prediction tasks once sufficient data and clear objectives are defined.