Expired POI Detection in Amap Using Deep Learning: Feature Engineering, RNN, Wide&Deep, and TCN Models

Gaode (Amap) maintains tens of millions of POIs (Points of Interest) such as schools, hotels, gas stations, and supermarkets. Over time many POIs become obsolete due to closure, demolition, relocation, or renaming, which degrades map freshness and user experience. Detecting expired POIs efficiently is therefore critical.

Because on‑site data collection is costly and slow, algorithmic mining—especially time‑series models built on large‑scale trend data—plays a key role in improving POI quality.

Expired POI detection can be abstracted as an asymmetric binary classification problem. The project builds a hybrid model that combines multi‑source trend features with high‑dimensional sparse attribute and status features.

2. Feature Engineering

Expired mining relies on sensing changes that accompany POI expiration. Temporal features are constructed by linking POIs with various auxiliary signals and aggregating them into monthly statistics. A sequence length of at least two years is used to capture periodic patterns.

Auxiliary (non‑temporal) features include:

One‑hot vectors indicating the month of the last manual verification; a verification flag suggests a lower expiration probability.

One‑hot encoding of industry type, since categories like catering have higher churn than landmarks.

Special handling of newly created POIs that have many zero values in the early part of the sequence, which can cause false trends. Two mitigation strategies are proposed: (1) variable‑length RNN that only consumes data after POI creation, and (2) a “gate” sequence (0 before creation, 1 after) added as an extra feature.

3. RNN Stage

The project first adopts a multi‑layer LSTM (RNN 1.0) that ingests the temporal and auxiliary features, outputs a hidden state, and feeds it to a SoftMax layer to compute expiration probability. The model’s confidence intervals drive automated or manual remediation.

RNN 2.0 focuses on high‑traffic (“hot”) POIs, whose data distribution differs markedly from tail POIs. For hot POIs, missing dimensions are zero‑filled to avoid noise, while for sparse tail POIs an interpolation scheme normalizes other dimensions and applies weighted averaging.

4. Wide&Deep Stage

To overcome the limited feature richness of pure RNNs, the team integrates static and state features using a Wide&Deep architecture. The Wide part encodes non‑temporal attributes (property, status, industry type) as high‑dimensional sparse vectors with presence flags and One‑Hot encodings. The Deep part remains the LSTM sequence model.

Feature encoding is illustrated below:

The Wide and Deep hidden layers are concatenated before the final SoftMax layer, enabling joint training and balanced weighting.

A “Dual‑LSTM” variant splits the sequence into a long‑RNN (without new features) and a short‑RNN (with new short‑term features), then fuses both hidden states with the Wide part.

Finally, the team replaces the Deep RNN with a Temporal Convolutional Network (TCN) and adds an attention mechanism (Wide&Attention‑TCN). TCN offers causal convolutions, parallel processing of the whole sequence, and dilated residual blocks for long‑range dependencies.

Attention aggregates hidden vectors from all time steps, producing a weighted summary that is then combined with the Wide part.

The Wide&Attention‑TCN model achieves higher accuracy and better computational efficiency compared with the Dual‑LSTM baseline.

5. Summary and Outlook

The deployment of deep learning for expired POI mining has progressed through iterative cycles of feature expansion, model redesign, and performance validation. Key take‑aways are the importance of rich, reliable features, appropriate representation, and model structures that fit the business scenario.

Future work will combine macro‑level trend signals with micro‑level contextual factors (geography, business health, competition) to enable finer‑grained, personalized expiration predictions.

6. References

1. Sepp Hochreiter and Jürgen Schmidhuber, “Long Short‑Term Memory,” Neural Computation, 1997.

2. B. Hidasi et al., “Session‑based recommendations with recurrent neural networks,” CoRR, 2015.

3. H‑T. Cheng et al., “Wide&Deep learning for recommender systems,” 1st Workshop on Deep Learning for Recommender Systems, 2016.

4. P. Covington, J. Adams, and E. Sargin, “Deep neural networks for YouTube recommendations,” ACM RecSys, 2016.

5. S. Bai, J. Z. Kolter, V. Koltun, “An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling,” 2018.

6. N. Pappas and A. Popescu‑Belis, “Multilingual Hierarchical Attention Networks for Document Classification,” 2017.