Machine Learning Solutions for User Feedback Intelligence at Amap (Gaode Maps)

Background : Amap, a leading Chinese navigation and location data provider, receives massive amounts of user feedback (text, images, videos) that are valuable for improving map quality and services. The company seeks to automate the processing of this feedback using machine learning.

Problem : Daily user feedback reaches hundreds of thousands. The current rule‑based pipeline requires manual verification of each report, leading to low accuracy, high labor cost, and slow response times.

Proposed Solution : Decompose the feedback handling workflow into three stages—Intelligence Recognition, Intelligence Localization, and Intelligence Verification—and replace the rule‑based classification with data‑driven machine‑learning models. The workflow is further split into six hierarchical levels, where only the last three levels (recognition, localization, verification) need limited human intervention.

Business & Model Alignment : User feedback contains both structured choice fields (source, major type, sub‑type) and free‑text descriptions. Descriptions are categorized as valid, invalid, or empty. Valid descriptions are further classified into product, data, and forwarding categories, with data further split into road‑related and topic‑related sub‑categories.

Model Selection : Text descriptions are vectorized using word2vec embeddings to capture semantic relationships, avoiding the sparsity of one‑hot/TF‑IDF representations. Deep learning models—specifically LSTM networks—are chosen over traditional statistical methods because they handle sequential data and mitigate gradient issues with the LSTM cell.

Model Architecture : Each feedback’s word‑embedding sequence is fed into an LSTM. The final LSTM hidden state is concatenated with the structured choice fields, passed through fully connected layers, and classified with a softmax output for multi‑class prediction. (See accompanying architecture diagram.)

Practical Experience :

Fine‑tuning pre‑trained models on limited labeled data yields ~3% accuracy gains.

Hyper‑parameter tuning experiments (initialization with SVD, dropout before LSTM, Adam optimizer, batch sizes 64‑128, data shuffling) improved model stability.

Ensembling the top‑5 models via voting increased overall accuracy by 1.5%.

Confidence‑based post‑processing: a simple per‑class thresholding strategy outperformed separate confidence models, and low‑confidence predictions are offered as top‑N suggestions to operators.

Results :

Intelligence Classification: product class accuracy >96%, data class recall >99%.

Intelligence Recognition: valid description accuracy >96%.

Automation rate increased dramatically, reducing manual workload by ~80% and cutting per‑task cost by 4/5.

Conclusion & Outlook : The project demonstrates a repeatable methodology for integrating NLP and deep‑learning techniques into complex business processes, delivering substantial efficiency gains and setting a foundation for further automation and model enhancements.