How Big Data Powers Personalized Recommendations in Mother‑Baby E‑Commerce

Characteristics of Mother‑Baby E‑Commerce

Mother‑baby e‑commerce sites have short product cycles (typically 5‑7 days), rapidly changing user needs as customers transition from pregnancy to caring for infants, and a mobile‑first traffic pattern where about 90% of transactions occur on smartphones.

Big Data Platform Architecture

The platform consists of a data collection layer, a computing layer (offline batch processing and real‑time streaming), an algorithm layer (collaborative filtering, classification for product features, and ranking models), and a business layer that incorporates operational rules for personalized adjustments and marketing.

The BI layer serves both operations and merchants, handling user browsing logs from PC and mobile, supporting both real‑time and offline data consumption, with daily batch statistics and real‑time reporting via a subscription model.

Distributed scheduling is the core of the platform, managing task assignment, monitoring, and logging. A master‑worker architecture ensures that failure of a master node does not affect overall operation.

Recommendation Types

The system provides three main recommendation forms:

Personalized ranking : Orders online brand‑sale items (e.g., 500 daily specials) based on individual shopping intent.

Item‑to‑item recommendation : Suggests similar or complementary products.

Personalized push : Displays items of interest on the user’s profile page and sends targeted SMS or push notifications.

User Profiling

User profiles consist of static attributes (gender, age, baby’s age and gender) and dynamic attributes (brand preferences, purchase time, channel). In this domain, gender and age are especially predictive of shopping preferences, enabling predictions such as baby‑age based product recommendations.

Modeling and Evaluation

Features include attribute, statistical, and preference features derived from user behavior. Various algorithms—classification, ranking, and non‑linear models—are evaluated; ranking models often outperform classification for conversion uplift. Offline and online learning pipelines update models continuously, allowing real‑time scoring of users and items.

Combining multiple algorithms (e.g., collaborative filtering on browsing and purchase data plus user features) yields a hybrid re‑ranking system that has shown significant lift in key metrics, with some resources improving by up to 500% after deployment.