TL;DR
This paper introduces "Audience Expansion," a system designed to help advertisers on LinkedIn reach more relevant users by intelligently expanding their target audiences. They propose two novel methods, campaign-agnostic and campaign-aware expansion, and a hybrid approach that significantly boosts ad inventory utilization and advertiser value without negatively impacting user experience or ad relevance.
What Problem Does This Paper Address?
Online advertising platforms like LinkedIn allow advertisers to precisely target users based on a wide array of attributes (e.g., job title, skills, location). However, manually defining these granular targeting criteria can be a complex and time-consuming task for advertisers. Even with sophisticated targeting, advertisers may struggle to reach all relevant potential customers, leading to missed opportunities and underutilization of ad inventory. This paper addresses the challenge of efficiently expanding an advertiser's initial target audience to include "lookalike" users who are likely to be interested in their ads, thereby increasing reach and improving the overall advertising ecosystem.
Methods
The paper presents two primary methods for Audience Expansion:
- Campaign-Agnostic Expansion: This method focuses on enriching individual user profiles. It leverages LinkedIn's "Similar-X" recommender system to identify attributes similar to those already present in a user's profile. For instance, if a user has the skill "Data Mining," this method might add "Big Data" and "Machine Learning" to their profile attributes, making them eligible for ads targeting those skills. This approach is immediately available for new campaigns.
- Campaign-Aware Expansion: This method takes a campaign-centric approach. It views targeting as a classification problem where the goal is to identify users who are similar to the campaign's exact target audience. It uses a "Similar-Profiles" algorithm (a specialized version of Similar-X) to find users who are highly similar to the members of the original target audience. This method is more precise as it considers the entire user profile in relation to the campaign's target.
- Hybrid Expansion: This approach combines the strengths of both campaign-agnostic and campaign-aware methods. It leverages the immediate availability of attribute expansion with the precision of lookalike modeling, offering a more comprehensive solution.
The core of these methods relies on sophisticated modeling techniques, including logistic regression for learning entity similarities and a locality-sensitive hashing (LSH) technique called Arcos for efficient nearest neighbor search in large-scale user data.
Datasets
The paper describes experiments conducted on LinkedIn's live traffic data. While specific datasets for training models are not explicitly detailed beyond "historical user-entity interaction logs" and "member profile attributes," the evaluation is based on a large-scale online A/B test involving a significant portion of LinkedIn's user base and active ad campaigns.
Key Insights and Ideas
- The Power of "Lookalikes": The core idea is that users with similar attributes or profiles to an existing target audience are highly likely to be interested in the same ads. This significantly simplifies targeting for advertisers.
- Complementary Expansion Strategies: Campaign-agnostic and campaign-aware expansion address different aspects of the problem (user profile enrichment vs. campaign audience expansion) and their combination yields superior results.
- Balancing Precision and Availability: Campaign-agnostic expansion is quick to deploy but less precise, while campaign-aware expansion is more precise but requires an offline processing time. The hybrid approach effectively bridges this gap.
- Similarity as a Predictor of Relevance: The paper demonstrates a strong correlation between the similarity scores generated by their models and actual user click-through rates, validating the effectiveness of their similarity measures.
- Ecosystem Benefits: Audience Expansion not only benefits advertisers by increasing reach and efficiency but also benefits LinkedIn by increasing ad inventory utilization and overall marketplace participation.
Key Results
The paper reports significant improvements across various metrics from large-scale online experiments:
- Impressions, Reaches, Matches, Revenue, and Value: The hybrid approach led to substantial increases, with impressions and reaches nearly doubling, and revenue and value increasing by over 100% for expansion-enabled campaigns compared to the control group.
- CTR and Dwell Time: These metrics showed minimal, statistically insignificant changes, indicating that audience expansion does not significantly degrade ad relevance or user engagement.
- CPC: While CPC increased, the paper argues this is due to more competitive auctions rather than advertisers targeting more expensive users, suggesting a healthy market dynamic.
- Coverage: The system's rollout led to a significant increase in sponsored updates coverage on desktop, demonstrating improved ad inventory utilization.
Limitations and Future Work
- CTR Improvement: While the drop in CTR for expansion-enabled campaigns was not drastic, the authors acknowledge it as an area for future improvement. They suggest that using similarity score thresholds could help target more promising users and mitigate CTR drops.
- Behavioral Signals: Future work includes incorporating more behavior-based signals (e.g., ad clicks, social actions) into the campaign-aware expansion model.
- Advanced Quality Control: The authors plan to explore more advanced methods for quality control of the expanded audience.
- Seed Audiences: They aim to allow advertisers to optionally provide seed audiences (like known converters) to learn campaign-specific performance models.
Summary updated