TL;DR
This paper introduces a novel methodology to estimate the real-time monetary value of individual users in the programmatic advertising ecosystem, even when bid prices are encrypted. By developing a "Price Modeling Engine" and a browser extension called "YourAdValue," they enable users to understand how much advertisers pay to reach them, offering a significant step towards transparency in digital advertising.
What Problem Does This Paper Address?
The digital advertising landscape has become increasingly data-driven, with Real-Time Bidding (RTB) protocols facilitating instantaneous auctions for ad space. While the value of user data for advertisers is widely acknowledged, a critical gap exists in understanding precisely how much advertisers are willing to pay for access to individual users. This lack of transparency is exacerbated by the growing trend of encrypting bid prices, making it difficult to ascertain the true cost of reaching a specific user. Existing methods often rely on averages (ARPU) or focus solely on cleartext prices, failing to capture the full picture, especially in the evolving RTB ecosystem.
Methods
The authors propose a two-pronged approach:
- Price Modeling Engine (PME): This is a backend system responsible for estimating encrypted bid prices. It uses a machine learning model (specifically, a Random Forest classifier) trained on ground truth data. To obtain this ground truth, the researchers conducted their own "probe" ad-campaigns. The PME focuses on a carefully selected subset of features (e.g., user location, time of day, device type, ad format, ADX/DSP) to keep campaigns cost-effective and the model efficient.
- YourAdValue: This is a client-side browser extension that monitors a user's network traffic. It identifies RTB notifications, extracts cleartext prices directly, and uses the PME's model to estimate encrypted prices. YourAdValue then aggregates these costs to provide the user with an estimated real-time "advertising worth."
A key technical contribution is the method for tallying winning bids even when encrypted, achieved by training a model on ground truth prices obtained from their own ad campaigns.
Datasets
- Offline Weblogs: A year-long dataset of weblogs from 1594 volunteering mobile users was used to bootstrap the PME and analyze user behavior and ad traffic patterns.
- Probing Ad-Campaigns: Two phases of real-world ad-campaigns were conducted. These campaigns were designed to collect ground truth prices for both cleartext and encrypted bids, specifically targeting ADX-DSP pairs that use encryption. This dataset differs from existing ones by specifically focusing on obtaining ground truth for encrypted prices, which are harder to capture.
Key Insights and Ideas
- Encryption Matters: Contrary to previous assumptions, encrypted prices are not simply a way to prevent tampering; they can be significantly higher (around 1.7x) than cleartext prices, suggesting that encryption might also be used to hide more aggressive bidding strategies or higher user values.
- User Value Varies Wildly: While the average user might be "cheap" to reach (median cost ~25 CPM), a small but significant portion of users (~2%) are vastly more expensive, costing 10-100x more than the average.
- Beyond Averages: The paper moves beyond aggregate metrics like ARPU to provide per-user, real-time cost estimations, offering a much more granular understanding of advertising economics.
- Feature Importance: Not all data points are equally valuable for predicting bid prices. The research identifies a core set of features that are crucial for effective price modeling, enabling cost-efficient ad campaigns.
Key Results
- Accuracy: The methodology can estimate a user's advertising worth with over 82% accuracy.
- Encrypted vs. Cleartext Prices: Encrypted prices were found to be approximately 1.7x higher than cleartext prices.
- Average User Cost: The median user cost for advertisers over a year was around 25 CPM, with 73% of users costing less than 100 CPM.
- Outlier Users: Approximately 2% of users were significantly more expensive, costing 10-100x more than the average.
- Cost Increase from Encryption: About 60% of users experienced an increased cumulative cost of ~55% due to estimated encrypted prices.
Limitations and Future Work
- Cost Model Assumption: The current methodology assumes a Cost-Per-Impression model for all charge prices. The authors acknowledge that other cost models (e.g., Cost-Per-Click) exist and are not currently accounted for.
- Mobile Focus: While the methodology is presented, the primary implementation and dataset analysis focused on mobile users.
- Generalizability of Probing Campaigns: The ad-campaigns used for ground truth were limited in scope and duration. While efforts were made to select representative features, the authors note that the publishers used in their campaigns are a subset of those found in real weblogs, and over-fitting is a potential concern.
- Future Work: The authors plan to make their "YourAdValue" prototype available for community testing and exploration. They also suggest future work on investigating the effects of retargeting and exploring the use of more advanced features from professional ad-campaign planners to enhance price modeling.
Summary updated