Robust Spatio-Temporal Purchase Prediction via Deep Meta Learning

Background Sales forecasting for online and offline retail is critical for inventory planning, especially during large shopping festivals, but historical data are often scarce and spatial‑temporal patterns are heterogeneous.

The authors present a AAAI 2021 paper titled Robust Spatio‑Temporal Purchase Prediction via Deep Meta Learning , which tackles this challenge using deep meta‑learning to model city‑wide purchase behavior across regions and time slots.

1. Spatio‑Temporal Representation Generation An amortization network is designed to approximate the posterior distribution of region‑ and date‑type representations, incorporating both sales features and spatial attributes such as POI distribution and demographics. Linear transformation modules capture coarse sales baselines, while a specialized jump‑LSTM models nonlinear temporal patterns; a feature‑fusion module combines temporal and static features.

2. Spatio‑Temporal Purchase Prediction Using the generated representations, a generative model with two inputs—recent sales features and sampled spatio‑temporal embeddings—produces final sales forecasts via a fully‑connected network.

3. Alternating Spatio‑Temporal Training Inspired by multi‑view learning, two views (spatial and temporal) are constructed; training alternates between them to capture complementary information, improving representation quality. The complete training procedure is illustrated in Algorithm 1.

Experiments The model is evaluated on a large JD.com dataset (2015‑2019) covering Beijing’s 18 districts, including purchase orders, shopping‑cart data, and static region features. Results (Table 1, Figure 4‑5) show that the proposed STMP method outperforms baselines by at least 30 % in MSE across regular days and shopping‑festival peaks, and remains robust when only a single week of festival data is available.

Overall, the deep meta‑learning approach effectively captures heterogeneous spatio‑temporal purchase patterns, enabling accurate sales forecasting even with limited historical observations.