How Dynamic Fusion Improves Life‑Stage Prediction for Baby Products Using AI

Background

Consumer life‑stage has long been studied in marketing and sociology, but most e‑commerce research focuses on recommending items based on past behavior rather than inferring the consumer's life stage. We propose a dynamic fusion algorithm to infer life stages for mother‑and‑baby shoppers.

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 handling multiple children.

We validate the solution with real data, showing significant gains in inference accuracy.

Mother‑and‑Baby Life‑Stage Segmentation

We divide the infant’s life stage into several age‑based periods: pre‑birth (pregnancy), 0‑6 months (newborn), 6‑12 months, 2‑3 years (nursery), and 3‑7 years (kindergarten). Different stages correspond to distinct purchasing patterns, such as diapers for newborns and clothing for toddlers.

Dynamic Fusion Method

For each month we predict a probability distribution over life stages using a multinomial logistic regression model. When a new distribution is generated, we translate the previous distribution forward by a time shift Δ (see Equation (1)) and then compare the top‑probability stages. If they match, we fuse the distributions using the normalization in Equation (2); otherwise we keep both distributions and record how many times each has been fused. The stage with the highest fusion count is the final prediction.

Feature Engineering

Features are derived from five consumer actions: search, click, favorite, add‑to‑cart, and purchase. We use:

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

Category‑attribute features : combinations such as brand and size.

Product attributes : intrinsic item properties.

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

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

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

Experiments

Using six months of Taobao data (Sept 2016 – Feb 2017), we compare the dynamic fusion method with a memory‑less baseline that only uses the latest month’s features. Results show the fusion method improves accuracy by up to 15 % and recall by up to 22 % in sparse months, while matching baseline performance in dense months.

The dynamic fusion method maintains or exceeds baseline performance across all months, especially when consumer behavior data are sparse. It also gradually corrects earlier mis‑predictions in subsequent months.

Conclusion

We introduced a dynamic fusion algorithm for inferring infant life stages from parental e‑commerce behavior, detailed the feature engineering process, and demonstrated its superiority over a memory‑less approach through extensive offline experiments. Future work may explore richer models and multi‑level architectures for further improvements.