Deep Learning for Expired POI Detection at Amap: Feature Engineering, RNN, Wide&Deep, and Attention‑TCN

Amap maintains tens of millions of POIs (Points of Interest) such as schools, hotels, and supermarkets, but many become obsolete over time, harming map freshness and user experience. Detecting expired POIs is framed as an asymmetric binary classification problem, solved with multi‑source trend features and high‑dimensional sparse attributes.

Feature Engineering

Temporal features are built by aggregating POI‑related metrics into monthly statistics, requiring at least two years of history. Auxiliary features include a one‑hot vector marking the last month of manual verification, industry‑type one‑hot encodings, and a special "gate" sequence to handle newly created POIs with many zero values.

Variable‑length RNN that truncates sequences after POI creation.

Fixed‑length sequence with a binary "gate" feature (0 before creation, 1 after).

RNN Stage

RNN 1.0 : A multi‑layer LSTM consumes aligned temporal and auxiliary features; the final hidden state passes through a SoftMax layer to output expiration probability, enabling automated or manual handling based on confidence.

RNN 2.0 : Improves detection of high‑traffic POIs by training a dedicated head‑segment model and applying different missing‑value strategies (zero‑fill for hot POIs, interpolation for sparse ones), resulting in better recall for popular locations.

Wide&Deep Stage

Wide & LSTM : Encodes static attributes, status flags, and industry types as high‑dimensional sparse vectors (with presence flags) and concatenates them with the LSTM hidden state before the SoftMax layer.

Wide & Dual‑LSTM : Handles long‑term sequences without new features and short‑term sequences with new features in separate RNN modules; their hidden states are merged with the Wide part for final prediction.

Wide & Attention‑TCN : Replaces the Deep RNN with a Temporal Convolutional Network (TCN) that preserves causality, uses residual and dilated convolutions, and adds an attention mechanism over all time steps. The aggregated TCN vector is coupled with the Wide part, yielding higher accuracy and faster inference.

Conclusion & Outlook

The progressive evolution—from simple LSTM models to Wide‑Deep and Attention‑TCN architectures—demonstrates that rich, well‑engineered features combined with appropriate deep‑learning structures significantly boost expired POI detection. Future work will integrate fine‑grained contextual factors (geography, business health, competition) to achieve more personalized, precise mining.