M3C: a Multi-Domain Multi-Objective, Mixed-Modality Framework for Cost-Effective, Industry Scale Recommendation
Keywords: Recommendation, efficiency, data consolidation
TL;DR: M3C is a collection of efforts to improve recommendation quality and cost-efficiency in an industry-scale deployment.
Abstract: The ever-expanding landscape of products, surfaces, policies, and regulations poses
significant challenges for recommendation systems, leading to data fragmentation
and prohibitive hikes in infrastructure costs. To address these challenges, we
propose M3C, a holistic co-design of model, data and efficiency strategies. M3C
(1) partitions the recommendation space to allow better representation learning
and encourage knowledge sharing within a subspace; (2) covers each partition
using a hierarchy of foundational and vertical networks tailored to handle multi-
domain, multi-objective tasks with mixed-modal inputs; (3) forms a unified data
representation that utilizes heterogeneous signals across domains, objectives and
optimization goals to alleviate data fragmentation, label sparsity, and to enhance
knowledge sharing; (4) improves execution efficiency and lowers costs with a suite
of stability and throughput optimizations. We show that across a diverse set of tasks
on public and industry datasets, M3C delivers up to 1% lower LogLoss compared
to 10 state-of-the-art baselines, while improving system efficiency by up to 20%.
Furthermore, in a large-scale industry setting our deployment of M3C has resulted
in 7% top-line metrics improvement in online tests with 10% capacity savings.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1181
Loading