How Alibaba’s Traffic Quality Team Detects and Mitigates Advertiser Arbitrage

Background

The Alibaba Mama traffic quality team focuses on cleaning invalid traffic, which includes low‑quality clicks (e.g., duplicate click billing, frequency control) and outright cheating (e.g., clicks with zero conversion probability).

Cheating traffic always has a conversion probability of zero, such as clicks generated by crawlers, but a zero‑frequency flow does not necessarily indicate cheating; it may simply be a new product with many clicks but no conversions yet.

Advertiser Arbitrage Overview

Advertiser arbitrage is defined as advertisers using cheating tactics (fake clicks, fake conversions) to deceive the platform and obtain more ad resources at lower cost. During the recent Double‑11 promotion, two main arbitrage risks were targeted: quality‑score cheating and malicious over‑investment.

Quality‑Score Cheating

Some advertisers hire black‑gray market workers to artificially inflate click and conversion rates, thereby improving their quality scores and securing higher ad placements at low cost.

Malicious Over‑Investment

Due to billing delays, certain advertisers purchase large volumes of search terms with high unit prices but limited budgets, resulting in a surge of clicks that far exceed their budget. The risk team detects abnormal backend operations and imposes strict penalties.

Arbitrage Impact

Arbitrage reduces the platform's revenue per mille (RPM) because high‑quality slots are occupied by low‑conversion traffic, lowering overall clicks and conversions. Although the platform’s CTR/CVR models self‑heal over time, there is a lag during which the model is repeatedly deceived and corrected.

Challenges

Cheating techniques evolve rapidly.

Ground‑truth labels are hard to define.

Supervised training struggles to find a universal solution.

Crowdsourced Traffic Identification

To detect crowdsourced artificial traffic, the team built a multi‑model framework consisting of a black‑phrase model, statistical baseline, and behavior‑sequence model, each contributing to recall and precision.

Black‑Phrase Model

By aggregating historical search logs, a standardized text library of black‑phrases was created. These phrases exhibit chaotic word order, mixed languages, emojis, and rapid iteration, making pure text‑based detection difficult.

Two‑stage semi‑supervised recall is used: first, unsupervised user‑dimension penalty features generate candidate clusters; second, transformed keywords train a CART regression tree with dynamic thresholds to build an offline sample pool. The resulting black‑phrase sample library reaches tens of millions with >99% manual verification accuracy.

A BERT‑based binary classifier, trained on black‑phrase samples and trusted white samples, scores all text to compensate for coverage gaps.

Statistical Baseline

Statistical anomaly features are engineered from user behavior, focusing on deviations from typical distributions rather than absolute counts. These features provide robust signals for arbitrage detection but lack granularity to pinpoint specific crowdsourced actions.

Behavior‑Sequence Modeling

Fine‑grained raw behavior sequences (search and click) are encoded as tokens like clk_true_3_2, representing click type, ad click flag, product ID, and time‑gap bucket. Sessions start with a search and are bounded by time gaps.

A Transformer with a cut‑paste auxiliary task pre‑trains on massive unlabeled sequences, producing N‑dimensional embeddings. Anomaly scores are derived by combining COPOD‑based distribution deviations and reconstruction errors from a fine‑tuned Seq2Seq model.

Graph Association Mining

Statistical baseline nodes are refined with behavior‑sequence results, then a semi‑supervised Graph Attention Network (GAT) expands recall to sparsely connected crowdsourced traffic that shares high overlap with known malicious nodes.

Arbitrage Perception

The perception layer estimates robust RPM for each advertiser and computes the difference between observed and expected RPM. Advertisers with significantly lower RPM are flagged as successful arbitrage cases, forming a superset for downstream analysis.

Insight Platform

An internal insight platform visualizes traffic distribution changes and backend behavior patterns, providing analysts with interactive curves and dimensionality‑reduced graphs to quickly assess arbitrage risk.

Downstream Handling

Based on the arbitrage list, advertisers are categorized as active arbitrage or accidental victims. Tailored penalty strategies are applied, and educational campaigns are launched to cleanse the ecosystem.

Outlook

While substantial progress has been made, many technical challenges remain, offering valuable research directions for the broader risk‑detection community.