Meta Lattice: Model Space Redesign for Cost-Effective Industry-Scale Ads Recommendations

TL;DR

This paper introduces Lattice, a holistic framework for cost-effective industry-scale ads recommendations built on model-space redesign. By consolidating portfolios, unifying heterogeneous data, and deploying Lattice Networks with efficiency-driven components (Zipper, Filter, KTAP, and Sketch), it enables cross-domain knowledge sharing while meeting latency and cost constraints. Empirical results on public and internal datasets show Lattice outperforms strong baselines and delivers tangible production benefits, including 10% revenue-topline gains, 11.5% ads-quality uplift, 6% conversion increase, and 20% capacity savings at Meta. Overall, Lattice demonstrates how synchronized data, model, and systems optimization can translate cutting-edge research into scalable, real-world ranking and recommendation improvements.

Abstract

The rapidly evolving landscape of products, surfaces, policies, and regulations poses significant challenges for deploying state-of-the-art recommendation models at industry scale, primarily due to data fragmentation across domains and escalating infrastructure costs that hinder sustained quality improvements. To address this challenge, we propose Lattice, a recommendation framework centered around model space redesign that extends Multi-Domain, Multi-Objective (MDMO) learning beyond models and learning objectives. Lattice addresses these challenges through a comprehensive model space redesign that combines cross-domain knowledge sharing, data consolidation, model unification, distillation, and system optimizations to achieve significant improvements in both quality and cost-efficiency. Our deployment of Lattice at Meta has resulted in 10% revenue-driving top-line metrics gain, 11.5% user satisfaction improvement, 6% boost in conversion rate, with 20% capacity saving.

Figures (6)

• Figure 1: Lattice overview: ① Even with MDMO frameworks, current model space remains scattered, with heterogeneous datasets, assorted attribution windows and siloed models and objectives. ②-③ Lattice redesigns the model space by consolidating datasets of heterogeneous formats, attribution windows and selects features on the pareto-front to form a small set of consolidated portfolios, improving knowledge sharing across domains and cutting down serving cost. ④ Foundational Lattice Network employs a MDMO-based, unified model architecture to consume multiple data formats and produces predictions for all consolidated objectives. ⑤ These models are further distilled to lean user-facing models via novel knowledge transfer and efficiency optimizations. ⑥ Deployed Lattice at Meta results in significant improvements in all key metrics and cost reduction.
• Figure 2: Lattice Networks interleave sequence and non-sequence learning to handle diverse input types in merged domains.
• Figure 3: Illustration: Lattice Filter with $T=9$ and $|\mathcal{F}|=10$ via 3 iterations.
• Figure 4: Relative Loss Improvement (%) over AFN on an industry-scale dataset.
• Figure 5: Relative Loss Improvement (%) of Lattice Zipper across various dates in a month, compared to an upperbound.