TL;DR
This paper introduces a novel, automated methodology for measuring and characterizing Online Behavioural Advertising (OBA). Their findings reveal that OBA is prevalent, often targets sensitive topics despite regulations, and is not significantly impacted by Do-Not-Track (DNT) signals, suggesting a need for better enforcement.
What Problem Does This Paper Address?
The rise of Online Behavioural Advertising (OBA) has brought significant benefits to advertisers and publishers, but it also raises substantial privacy concerns. OBA relies on tracking users' online activities to associate them with specific interests and then serving targeted ads. However, understanding the true extent and nature of OBA has been challenging due to the complexity of the advertising ecosystem and the difficulty in distinguishing OBA from other ad types like contextual or demographic targeting. Furthermore, there's a lack of robust, scalable methods to quantify OBA and investigate its application to sensitive topics or the effectiveness of privacy controls like DNT. This paper aims to fill this gap by providing a systematic approach to measure and characterize OBA.
Methods
The authors developed a novel, automated methodology to detect and quantify OBA. Key aspects include:
- Artificial Personas: They created "artificial online personas" representing specific behavioral traits (e.g., 'cooking', 'motor sports') to mimic user browsing behavior. This allows for controlled experiments.
- Keyword-based Tagging: They leveraged multiple online tagging services (Google AdWords, Cyren, McAfee) to assign keywords to training and landing pages. This multi-source approach enhances robustness and reduces dependency on a single platform.
- Semantic Overlap Filtering: To ensure personas are not contaminated by unrelated interests, they used semantic similarity (Leacock-Chodorow similarity) to filter keywords associated with training pages.
- Ad Filtering: A series of filters were designed to distinguish OBA from other ad types:
- Retargeting Ads Filter (Fr): Removes ads pointing to previously visited sites.
- Static and Contextual Ads Filter (Fs&c): Identifies and removes ads that are not behaviorally targeted by comparing ads shown to a "clean profile" (no history) with those shown to the persona.
- Demographic and Geographical Targeted Ads Filter (Fd&g): Filters out ads likely based on location or demographics by analyzing similarities between personas receiving the same ad.
- Metrics: Two key metrics were defined:
- Targeted Training Keywords (TTK): Measures the fraction of training keywords that appear in the landing pages of ads. A high TTK indicates that the persona's defined interests are being targeted.
- Behavioural Advertising in Landing Pages (BAiLP): Captures the fraction of ads whose landing pages are tagged with keywords from the persona's training set, indicating the proportion of ads likely driven by OBA.
- Automated System: The methodology was implemented in a fully automated, scalable system using a headless browser (PhantomJS) to conduct experiments without manual intervention.
Datasets
The paper does not explicitly list pre-existing datasets for training. Instead, it describes the creation of:
- Artificial Personas: 51 regular and 21 sensitive personas were systematically selected and manually curated.
- Training Webpages: For each persona, a set of related websites was identified using Google Ad Words' hierarchical category system.
- Control Webpages: Five popular weather-related websites were used as control pages to collect ads.
- Landing Webpages: These are the pages users are directed to from the ads shown on control pages.
The authors also collected over 3.5 million ads across more than 2.9K experiments.
Key Insights and Ideas
- OBA is Pervasive: The study strongly suggests that OBA is a common practice, with a high percentage of personas being targeted based on their defined interests.
- Sensitive Topics are Targeted: Despite regulations in some regions (like Europe) prohibiting tracking of sensitive topics (health, religion, politics), the research provides evidence that the online advertising market does indeed target these areas.
- Economic Value Drives Targeting: There's a significant correlation between the level of OBA received by a persona and its estimated economic value in the advertising market (indicated by Cost Per Click bids).
- DNT is Largely Ineffective: The experiments show that enabling Do-Not-Track (DNT) signals has a negligible impact on the amount of OBA received, suggesting that support for DNT is not widely enforced by ad networks and websites.
- Geographic Bias is Minimal: The study found no significant geographical bias in OBA application between the US and Spain.
- Methodology Validation: A key contribution is the rigorous evaluation of their OBA detection system's performance using standard metrics like accuracy and false positive rates, which is often missing in prior work.
Key Results
- OBA Prevalence: 88% of analyzed personas received ads associated with all their defining keywords, and half received 26-62% of their ads via OBA.
- Sensitive Topic Targeting: For sensitive personas, 10-40% of ads received were associated with OBA, despite legal restrictions.
- Economic Correlation: A Spearman correlation of 0.44 and Pearson correlation of 0.40 were found between BAiLP and suggested CPC bids, validating the economic value hypothesis.
- DNT Ineffectiveness: The median difference in BAiLP when DNT was activated versus deactivated was close to zero, with a small interquartile range, indicating minimal impact.
- System Performance: The automated system achieved high accuracy (over 94%) and recall, with a false negative rate below 4.5% for most personas.
Limitations and Future Work
- Persona Contamination: While efforts were made to minimize contamination, the authors acknowledge it's a persistent challenge in defining precise behavioral traits.
- Tagging Service Dependency: Although multiple tagging services were used, the methodology's reliance on these services means its effectiveness is tied to their accuracy and coverage.
- "Tragedy of the Commons": The paper frames privacy concerns within this economic model, suggesting that over-harvesting of consumer privacy could lead to its destruction.
- Future Work: The authors aim to improve the transparency of the advertising ecosystem and protect consumer privacy. They suggest their methodology can pave the way for more robust tools and further research into the intrinsic aspects of online advertising. They also highlight the need for better enforcement of privacy regulations and signals like DNT.
Summary updated