Can Machine Learning Predict Baby Life Stages to Boost E‑Commerce Recommendations?

Background

Life‑stage influences consumer behavior and has been studied for decades in marketing and sociology. In e‑commerce, most recommendation systems focus on historical behavior, while the relationship between a consumer's life stage and purchasing decisions remains under‑explored. This work proposes a dynamic fusion algorithm to infer life stages in the mother‑and‑baby vertical.

Main Contributions

We provide an industrial‑grade solution for life‑stage inference.

We develop a dynamic fusion method that continuously improves prediction accuracy while saving computational resources and supporting multiple children.

We validate the solution with real‑world data, showing its effectiveness.

Mother‑and‑Baby Life‑Stage Division

Based on domain knowledge, a child's life stage is divided by age: pre‑birth (pregnancy), 0‑6 months (newborn), 6‑12 months, 2‑3 years (nursery), 3‑7 years (kindergarten). Different stages correspond to distinct product interests, enabling targeted recommendations.

Dynamic Fusion Method

The algorithm predicts a probability distribution over life stages each month using a multivariate logistic regression model. New distributions are shifted in time (Δ) and fused with previous ones. The shift is defined by Equation (1):

If the highest‑probability stages of two distributions match, they are merged using Equation (2):

The merged distribution is normalized by a factor (see Equation (3)) to ensure a valid probability distribution.

Feature Engineering

Features are derived from five consumer actions: search, click, collect, add‑to‑cart, and purchase. They include:

Category features : first‑level and leaf categories to avoid sparsity.

Category attribute features : brand, size, etc., combined with category.

Product attribute features : intrinsic product properties.

Search term features : keywords that directly indicate age (e.g., “3‑year‑old baby clothes”).

Title features : extracted age‑related keywords from product titles.

Temporal features : recent month behavior and same‑month behavior from the previous year.

Experiments

We conducted offline and online experiments using six months of Taobao data (Sept 2016 – Feb 2017). Figures show the distribution of life stages in training and test sets.

Dynamic fusion consistently outperforms the memory‑less baseline, especially in months with sparse consumer behavior, achieving up to 15 % higher accuracy and up to 21.76 % improvement in recall.

Conclusion

The proposed dynamic fusion method effectively infers infant life stages from parental e‑commerce behavior, offering a practical tool for recommendation systems. Future work may explore alternative machine‑learning models or hierarchical approaches to further boost performance.