Detecting Offline Merchant Service Issues Using Machine Learning and Big Data at Nuomi

Nuomi, an O2O lifestyle service platform, faces critical offline merchant problems that directly affect user retention and brand reputation, such as stores being closed, under renovation, or refusing service, which are reflected in refund reasons like “cannot use”, “closed”, and “no service”.

By analyzing a month’s worth of refund requests and complaint data, the team identified that merchant‑related issues dominate the top complaints, prompting a need to mine these cases using transaction, product, and user‑generated content features.

The solution distinguishes between single‑store and multi‑store orders: for single‑store orders, user intent is clear, allowing direct extraction of order‑level features; for multi‑store orders, location data (LBS) and other big‑data signals are used to infer store‑level attributes such as verification rates and foot traffic.

A machine‑learning pipeline is built using samples from refunds, complaints, transactions, and UGC, with manually labeled “served” vs “not served” cases. After denoising, a decision‑tree classifier is trained, incorporating information gain for feature selection and a regional matrix to account for city‑level performance differences.

For multi‑store scenarios lacking direct order data, the model leverages crowd density and other big‑data indicators to construct a funnel model that identifies anomalous merchants.

The resulting insights are fed into an online platform for customer‑service and quality‑control teams to intervene promptly, and confirmed cases are looped back into the training set, improving model precision and recall while enabling a closed‑loop operational process.