ユーザーシーケンスからスケーリング則まで：Metaの広告ランキングを支える多段階アーキテクチャFrom User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta’s Ads Ranking

Meta has published a technical account of how its advertising recommendation stack has moved from sequence-based models of user behavior toward a multi-stage architecture whose design is guided by scaling laws. The topic matters because ad ranking is central to Meta's business and to the experience of billions of users, and the company's engineering choices tend to influence how the wider recommendation-systems field approaches similar problems.

At the core of the work is the observation that Meta's recommendation platforms process billions of user interactions every day. Those interactions produce rich temporal signals that reflect an individual's preferences and intent across products, ads, and organic content. Rather than treating a user as a static bundle of manually engineered sparse features, Meta's approach models the order and timing of actions a person takes over time. This builds on a 2024 post in which the company described how sequence learning could capture behavioral patterns that hand-crafted features miss, framing the user's history as an ordered sequence that a model can learn from directly.

The newer piece extends that idea into a layered system. In modern recommendation pipelines, a multi-stage architecture typically separates the problem into stages such as candidate retrieval and finer-grained ranking, because scoring every possible ad against every user with a large model is computationally impractical. A common pattern is to use lighter models to narrow a vast inventory down to a manageable set, then apply progressively heavier models to rank the survivors. Meta's account appears to follow this logic, distributing sequence-based user understanding across stages so that accuracy and serving efficiency can be balanced rather than traded off bluntly.

The reference to scaling laws is notable. Scaling laws describe empirical relationships between model performance and factors such as model size, dataset size, and compute budget, and they became widely discussed through research on large language models, where predictable gains from added scale helped justify ever-larger training runs. Applying that framing to ads ranking suggests Meta is trying to understand how its recommendation models improve as it invests more parameters, data, and compute, and to use those relationships to decide where additional scale is worthwhile. In a production advertising system, that discipline is important because latency and cost constraints are strict, and indiscriminate scaling is likely to be uneconomical without a clear picture of the returns.

This work sits within a broader arc of Meta's recommendation research. The company has previously open-sourced its Deep Learning Recommendation Model, or DLRM, which became a widely cited reference design for combining dense and sparse features in large-scale recommenders. More recently, Meta researchers have explored generative approaches that reframe recommendation as a sequence prediction task, drawing on architectural ideas from the transformer models that power language systems. The current article's emphasis on user sequences and scaling laws is consistent with that trajectory, in which techniques associated with large language models are increasingly adapted to the distinct demands of retrieval and ranking.

The industry context is similarly relevant. Recommendation and ranking teams across large platforms have been moving away from heavily hand-engineered features toward learned representations of user behavior, and several companies have reported benefits from treating engagement history as a sequence. Sequence modeling and attention mechanisms allow systems to weigh recent and past actions differently and to capture longer-range patterns, though they also raise engineering challenges around storing long histories, keeping features fresh, and serving predictions within tight time budgets. A multi-stage design is one way to manage those tensions, which helps explain why it recurs across large-scale systems.

For readers less familiar with the terminology, a few prerequisites help. Sparse features are the many discrete signals, such as identifiers and categories, that recommendation models have traditionally relied on. Dense embeddings are learned numerical vectors that represent such signals in a form models can generalize from. Retrieval and ranking are the successive filtering steps that turn a huge candidate pool into a short, ordered list.

Meta presents this as an evolution of its existing system rather than a wholly new product, and the post is an engineering explanation rather than a benchmark disclosure. As with any first-party account, the specific gains and trade-offs would need independent evaluation, but the described direction aligns with well-established trends in large-scale recommendation.