How AI‑Driven Data Mining Revives POI Freshness: A Deep Dive into Expired POI Detection

1. Introduction

Point‑of‑Interest (POI) data on digital maps includes restaurants, hotels, stations, parking lots and many other venue types. Enhancing timeliness means detecting POIs that have closed, moved, renamed or been demolished and marking them offline as quickly as possible. Even a modest stale‑rate of 0.1 % per year can generate hundreds of millions of negative user interactions.

2. Problem Overview

Expired POIs degrade user experience. The solution consists of two phases: discovery and handling . Discovery is driven mainly by large‑scale data mining (the “mining line”) and is complemented by manual collection and sentiment analysis. Handling uses three mechanisms—manual verification, automated labeling and automated offline removal—each with distinct cost and risk profiles.

3. Evolution of Data‑Mining Techniques

3.1 Stage 1 – Attribute‑Based POI Expiration Mining

Early work relied on POI‑internal attributes because external signals were scarce. POI attributes are grouped into:

Basic information : name, coordinates, address, category, phone, timestamps, source.

Deep/Dynamic information : images, ratings, reviews, deals, pricing.

Relational information : parent‑child links, citations, co‑occurrence.

Two strategy families were explored:

Single‑POI strategies – e.g., comment semantic mining : a TextCNN classifier processes user reviews to detect explicit mentions of closure.

Multi‑POI strategies – e.g., "original" relationship mining : extract old‑new name pairs from titles or aliases, aggregate them, and flag the older name as expired.

Additional heuristics include:

Same‑address strategy : build a graph of POIs sharing the same street number, decompose it into maximal connected components, and treat the older component as potentially expired.

Same‑phone strategy : group POIs by identical phone numbers, filter out chain stores or duplicates, then use name similarity and spatial distance to infer renames or relocations.

3.2 Stage 2 – Behavior‑Based POI Expiration Mining

3.2.1 Temporal Anomaly Detection

When usage‑behavior signals (e.g., order volume, navigation requests) become available, a declining activity trend is taken as a proxy for POI expiration. The modeling pipeline evolved as follows:

Random Forest (RF) on monthly activity features – high accuracy, low over‑fit, easy to deploy.

Recurrent Neural Network (RNN) – captures sequential dependencies missed by RF.

Model‑fusion pipeline (RF → RNN → fusion) – combines the strengths of both.

Wide & Deep architecture – encodes numerous categorical and continuous features in a wide component, while the deep component (RNN) learns temporal patterns; the two are merged via a fully‑connected layer.

This pipeline achieved stable production performance and markedly reduced the manual verification workload.

3.2.2 Event Anomaly Detection

Temporal models are ineffective for POI types lacking clear usage signals (e.g., gas stations, ATMs). An event‑based model extracts abnormal sessions from navigation logs: when a user reaches an expired POI and encounters an error or immediately initiates a second navigation, the session is flagged as anomalous.

Key challenges:

Massive log volume – requires efficient pre‑filtering to avoid wasteful computation.

High behavioral randomness – need to isolate patterns that are truly correlated with expiration.

Solution framework:

Define concrete abnormal scenarios (e.g., error after arrival, rapid re‑navigation).

Extract session‑level features per POI ID within a sliding time window.

Combine these features with static POI attributes (type, city) and feed a Logistic Regression (LR) model that outputs an expiration score.

Despite its simplicity, the LR model effectively suppresses noise and complements the temporal detectors.

3.3 Stage 3 – Human‑Place Relationship Mining

When usage trends are sparse, the system builds “human‑place” relationships to identify internal user groups that depend on a POI. The pipeline consists of three layers:

Data layer : ingest heterogeneous data sources (e.g., transaction logs, check‑in records) and link them to POIs.

Behavior layer : represent group behavior as a multi‑channel matrix (days × behaviors). Each behavior channel forms a separate matrix slice, preserving periodicity and allowing easy addition of new channels.

Model layer : apply a deep convolutional neural network to the multi‑channel matrix for binary classification (expired vs. active). A subsequent recall‑oriented expansion step increases coverage of internal relationships.

This architecture mitigates feature sparsity, learns temporal patterns, and improves recall for POIs with limited direct signals.

4. Summary and Future Directions

Behavior‑driven POI expiration mining has become the primary engine for enhancing POI timeliness. Ongoing research aims to:

Model spatio‑temporal shifts of internal user groups to further reduce reliance on dense attribute data.

Develop targeted mining capabilities for specific expiration phenomena (e.g., renames, relocations).

Construct comprehensive POI life‑cycle portraits.

Shift from purely reactive (offline) handling toward proactive (online) enrichment of the POI ecosystem.