TLDR: This research from Pinterest introduces a novel approach to improve ad performance by integrating users’ onsite ad interactions with their offsite conversion data into a large-scale heterogeneous graph. They developed a new Knowledge Graph Embedding (KGE) model called TransRA, which efficiently learns entity representations from this complex data by designating a ‘user space’ as an anchor. The key innovation lies in an attention-based fine-tuning method that successfully integrates these graph embeddings into ad ranking models, leading to significant improvements in Click-Through Rate (CTR) and Conversion Rate (CVR) predictions, and ultimately, better online ad performance for Pinterest.
In the world of online advertising, understanding what users are interested in buying is key to showing them the right ads. For platforms like Pinterest, this means not just looking at what users do on their site (like clicking on ads or saving pins), but also what they do outside of it, such as making purchases on other websites. This external shopping data, known as offsite conversions, is incredibly valuable for improving ad effectiveness.
However, combining these two types of data – onsite interactions and offsite conversions – comes with its own set of challenges. Offsite data often lacks detailed descriptions, and its format can vary widely between different advertisers. There can also be significant differences in how onsite and offsite data are structured, making it difficult to integrate them seamlessly.
To tackle these issues, researchers at Pinterest developed a groundbreaking approach. They built a massive, interconnected network, or ‘graph,’ that brings together both users’ onsite ad interactions and their opt-in offsite conversion activities. This ‘onsite-offsite heterogeneous graph’ provides a much more complete picture of user behavior, especially their shopping interests.
Introducing TransRA: A Smarter Way to Understand Connections
To make sense of this complex graph, the team introduced a new type of Knowledge Graph Embedding (KGE) model called TransRA (TransR with Anchors). KGE models are designed to learn numerical representations, or ’embeddings,’ for entities (like users, items, or advertisers) and their relationships within the graph. Unlike some other models, TransRA doesn’t rely on pre-existing features for each entity, making it highly flexible for real-world data where such information might be scarce or inconsistent.
The core idea behind TransRA is to designate one central ‘anchor space’ – in this case, the user space. All other entity types are then connected and transformed into this user space. This design makes it much easier to use the learned embeddings in downstream applications, such as ad ranking models, because all information is aligned to a common reference point. While previous models like TransE struggled with the complexity of this combined data, and TransR created separate spaces for different entity types (making integration difficult), TransRA successfully extracts valuable information while ensuring smooth integration.
Integrating Knowledge into Ad Ranking
The next big hurdle was getting these powerful graph embeddings to work effectively within Pinterest’s existing ad ranking models. Initially, simply using the pre-trained embeddings or directly fine-tuning them within the ranking models yielded only minor or neutral improvements. This suggested that the way the ad ranking models processed tabular data (like spreadsheets) wasn’t fully leveraging the rich, interconnected information from the graph.
The breakthrough came with an innovative ‘attention-based fine-tuning’ method. This approach introduces a special layer within the ad ranking models that mimics the interactions found in the graph. Essentially, it allows the ranking model to ‘pay attention’ to the relationships between different entities, similar to how the KGE model learns from the graph itself. This method proved highly successful, leading to significant performance gains.
Also Read:
- New GPU-Based Tool Speeds Up Knowledge Graph Embeddings
- Improving Graph Learning Across Domains with Noisy Labels
Real-World Impact
The results of this research are impressive. In offline experiments, the attention-based fine-tuning method led to substantial improvements in both Click-Through Rate (CTR) and Conversion Rate (CVR) prediction models. More importantly, this framework has been put into action within Pinterest’s Ads Engagement Model. Online experiments showed a significant 2.69% increase in Click-Through Rate and a 1.34% reduction in Cost-Per-Click, a crucial metric for advertisers. This demonstrates the real-world value of understanding user behavior across both onsite and offsite activities.
The techniques presented in this paper, detailed further in the original research paper, offer a promising path for other large-scale industrial models looking to leverage diverse data sources for improved performance.
