POI Category Tagging: Multi‑Label Classification, Feature Engineering and Model Design

POI (Point of Interest) refers to entities such as buildings, shops, bus stops, lakes, roads, etc., that appear on a map. In map search, POIs are the retrieval objects, analogous to webpages in web search, and selecting a POI displays a floating balloon.

Category tags summarize POI attributes along a category dimension (e.g., a Watson’s store is tagged as "cosmetics", the mall it belongs to is tagged as "shopping mall"). Tags enrich front‑end information display and support search‑side category recall, which cannot be satisfied by simple text matching or synonym expansion.

The tag system is built on three principles: (1) real user query expressions, (2) objective real‑world category distribution and product‑manager insights, and (3) hierarchical relationships among tags. A multi‑branch tree is constructed for each top‑level category (e.g., shopping).

Tagging POIs is essentially a multi‑label classification problem, not a single‑label one. Challenges include:

Multi‑label nature (a POI may belong to several tags simultaneously).

Textual issues: short POI names, low‑frequency expressions, ambiguous names.

Comprehensive issues: domain‑specific nuances (e.g., distinguishing nightclubs from bars, high‑frequency vs. low‑frequency brands).

High precision and coverage are required because low‑quality tags lead to wrong recall or missed results. Over 20 top‑level categories and thousands of sub‑tags demand efficient, high‑recall methods.

Technical solution overview : The pipeline consists of feature engineering, sample engineering, classification models, and multi‑path fusion.

Feature engineering : Use generic textual features (POI name, typecode, source category, brand) and high‑frequency proprietary features. Textual features address the main tagging problem, while non‑textual features (typecode, source category) are incorporated via a wide‑&‑deep style.

Sample engineering : Because manual labeling is infeasible for millions of tags, click‑log data and external resources are leveraged. Click logs provide massive, user‑intent‑rich samples but are noisy and biased toward high‑frequency queries. External resources supplement low‑frequency coverage. A two‑stage active‑learning workflow cleans and iteratively refines samples.

Click‑sample mining : Transform the required tag → POI samples into tag → query → POI by defining seed queries for each tag and retrieving corresponding clicks. Query generalization expands high‑frequency seed queries to low‑frequency synonyms using methods such as word2vec embeddings, synonym dictionaries, session context, and recommendation techniques (SimRank, DeepWalk). 点击数据：query -> POI 需要样本：tag -> POI 解决方案：tag -> query -> POI

Model design :

Initial hierarchical multi‑class classifiers built per non‑leaf node.

Later, per‑tag binary classifiers (one‑vs‑rest) to handle multi‑label scenarios and reduce conflict.

Unified deep model based on textCNN, modified to (a) concatenate non‑textual features, (b) replace softmax with multiple sigmoid outputs, (c) compress multi‑label targets into a vector, and (d) adopt a weighted loss that accounts for class imbalance (weight = (h+ck)/ck).

Model improvements (textCNN with feature fusion) yielded >5% accuracy gain, 10× efficiency increase over rule‑based methods, and substantial long‑tail performance.

Evaluation : Tag quality is measured by precision and coverage; online A/B tests show that the new tag system raises category‑search relevance to 94% usage, delivering noticeable search quality improvements.

Summary & future work : The current system solves the majority of POI tagging using generic features and deep models. Remaining challenges include leveraging review/description attention mechanisms, image‑based category cues, external knowledge graphs, and crowdsourced labeling to further improve low‑frequency and ambiguous cases.