This just in! Engineering at Meta has announced an overhaul of their recommendations processes that could lead to performance changes within ad accounts.

Personalised advertising has become an essential part of the online experience. It allows businesses to target their ideal customers with relevant ads, leading to increased conversions and sales. However, traditional recommendation systems have limitations. This is where Meta steps in with its next-generation ads recommendation engine that leverages sequence learning.

Challenges of traditional DLRM-based ads recommendations

Traditional Deep Learning Relevance Models (DLRMs) have been the workhorse of ads recommendations for a long time. However, they come with some inherent limitations:

1. Loss of sequential information: DLRMs struggle to capture the sequential nature of user behavior. A user’s browsing history, for instance, tells a story – what they looked at earlier can influence what they might be interested in later. Traditional DLRMs often fail to consider this sequence, leading to less-than-optimal recommendations.

2. Limited ability to capture granular information: DLRMs rely on manually engineered features, which can be a cumbersome and time-consuming process. These features might not always capture the nuances of user behaviour, leading to inaccurate recommendations.

3. Reliance on human intuition: DLRMs depend on human intuition for feature engineering. This can be subjective and limit the effectiveness of the recommendations.

Meta’s sequence-learning solution

Meta’s next-generation recommendation engine overcomes these limitations by employing sequence learning at its core. This approach allows the system to capture the sequential nature of user behavior and extract more granular information, resulting in more relevant and personalised ad recommendations.

Event-based features (EBFs) (The building blocks)

The foundation of Meta’s new system lies in event-based features (EBFs). EBFs capture a user’s interactions with different elements on the platform, such as clicking on a like button or visiting a product page. This granular data provides a richer picture of user behaviour compared to traditional features.

Event sequence model (the engine)

The event sequence model is the heart of Meta’s recommendation engine. It’s a person-level event summarisation model that consumes sequential event embeddings. These embeddings are generated by feeding EBFs through an embedding layer, transforming them into a dense vector representation. The model then processes these embeddings sequentially, capturing the order and relationships between user interactions.

The benefits of sequence learning

Meta’s adoption of sequence learning in its ads recommendation system has yielded significant benefits:

• Enhanced ad relevance: By capturing the sequential nature of user behaviour, the system can recommend ads that are more relevant to a user’s current interests and needs.
• Improved performance: Sequence learning models can extract more granular information from user interactions, leading to better ad performance and higher click-through rates.
• Infrastructure efficiency: The new system is more efficient in terms of infrastructure usage, allowing Meta to scale its recommendation engine more effectively.
• Accelerated research velocity: The use of sequence learning opens doors for further research and innovation in the field of ads recommendations.

The future of ads recommendations

Meta’s sequence-learning based approach represents a significant leap forward in personalised advertising. By capturing the nuances of user behaviour and leveraging the power of sequential information, Meta’s system can deliver highly relevant ads that resonate with users and drive better results for advertisers. As Meta continues to refine its approach and explore new possibilities in sequence learning, we can expect even more innovative and effective ads recommendations in the future. To tie things off, here’s a promising note from Meta themselves about the performance of their overhaul so far:

These innovations have enabled our ads system to develop a deeper understanding of people’s behavior before and after converting on an ad, enabling us to infer the next set of relevant ads. Since launch, the new ads recommendation system has improved ads prediction accuracy – leading to higher value for advertisers and 2-4% more conversions on select segments.