Notification recommendation systems are critical to driving user engagement on professional platforms like LinkedIn. Designing such systems involves integrating heterogeneous signals across domains, capturing temporal dynamics, and optimizing for multiple, often competing, objectives. Graph Neural Networks (GNNs) provide a powerful framework for modeling complex interactions in such environments.

We present LinkSAGE, an innovative framework that integrates Graph Neural Networks (GNNs) into large-scale personalized job matching systems, designed to address the complex dynamics of LinkedIn's extensive professional network. Our approach capitalizes on a novel job marketplace graph, the largest and most intricate of its kind in industry, with billions of nodes and edges. This graph is not merely extensive but also richly detailed, encompassing member and job nodes along with key attributes, thus creating an expansive and interwoven network. A key innovation in LinkSAGE is its training and serving methodology, which effectively combines inductive graph learning on a heterogeneous, evolving graph with an encoder-decoder GNN model. This methodology decouples the training of the GNN model from that of existing Deep Neural Network (DNN) models, eliminating the need for frequent GNN retraining while maintaining up-to-date graph signals in near real-time, allowing for the effective integration of GNN insights through transfer learning. The subsequent nearline inference system serves the GNN encoder within a real-world setting, significantly reducing online latency and obviating the need for costly real-time GNN infrastructure. Validated across multiple online A/B tests in diverse product scenarios, LinkSAGE demonstrates marked improvements in member engagement, relevance matching, and member retention, confirming its generalizability and practical impact.

This paper introduces LiNR, LinkedIn's large-scale, GPU-based retrieval system. LiNR supports a billion-sized index on GPU models. We discuss our experiences and challenges in creating scalable, differentiable search indexes using TensorFlow and PyTorch at production scale. In LiNR, both items and model weights are integrated into the model binary. Viewing index construction as a form of model training, we describe scaling our system for large indexes, incorporating full scans and efficient filtering. A key focus is on enabling attribute-based pre-filtering for exhaustive GPU searches, addressing the common challenge of post-filtering in KNN searches that often reduces system quality. We further provide multi-embedding retrieval algorithms and strategies for tackling cold start issues in retrieval. Our advancements in supporting larger indexes through quantization are also discussed. We believe LiNR represents one of the industry's first Live-updated model-based retrieval indexes. Applied to out-of-network post recommendations on LinkedIn Feed, LiNR has contributed to a 3% relative increase in professional daily active users. We envisage LiNR as a step towards integrating retrieval and ranking into a single GPU model, simplifying complex infrastructures and enabling end-to-end optimization of the entire differentiable infrastructure through gradient descent.

In this paper, we present LiGNN, a deployed large-scale Graph Neural Networks (GNNs) Framework. We share our insight on developing and deployment of GNNs at large scale at LinkedIn. We present a set of algorithmic improvements to the quality of GNN representation learning including temporal graph architectures with long term losses, effective cold start solutions via graph densification, ID embeddings and multi-hop neighbor sampling. We explain how we built and sped up by 7x our large-scale training on LinkedIn graphs with adaptive sampling of neighbors, grouping and slicing of training data batches, specialized shared-memory queue and local gradient optimization. We summarize our deployment lessons and learnings gathered from A/B test experiments. The techniques presented in this work have contributed to an approximate relative improvements of 1% of Job application hearing back rate, 2% Ads CTR lift, 0.5% of Feed engaged daily active users, 0.2% session lift and 0.1% weekly active user lift from people recommendation. We believe that this work can provide practical solutions and insights for engineers who are interested in applying Graph neural networks at large scale.

We present LiRank, a large-scale ranking framework at LinkedIn that brings to production state-of-the-art modeling architectures and optimization methods. We unveil several modeling improvements, including Residual DCN, which adds attention and residual connections to the famous DCNv2 architecture. We share insights into combining and tuning SOTA architectures to create a unified model, including Dense Gating, Transformers and Residual DCN. We also propose novel techniques for calibration and describe how we productionalized deep learning based explore/exploit methods.

In the realm of recommender systems, the ubiquitous adoption of deep neural networks has emerged as a dominant paradigm for modeling diverse business objectives. As user bases continue to expand, the necessity of personalization and frequent model updates have assumed paramount significance to ensure the delivery of relevant and refreshed experiences to a diverse array of members. In this work, we introduce an innovative meta-learning solution tailored to the personalization of models for individual members and other entities, coupled with the frequent updates based on the latest user interaction signals. Specifically, we leverage the Model-Agnostic Meta Learning (MAML) algorithm to adapt per-task sub-networks using recent user interaction data. Given the near infeasibility of productionizing original MAML-based models in online recommendation systems, we propose an efficient strategy to operationalize meta-learned sub-networks in production, which involves transforming them into fixed-sized vectors, termed meta embeddings, thereby enabling the seamless deployment of models with hundreds of billions of parameters for online serving. Through extensive experimentation on production data drawn from various applications at LinkedIn, we demonstrate that the proposed solution consistently outperforms the best performing baseline models of those applications, including strong baselines such as using wide-and-deep ID based personalization approach. Our approach has enabled the deployment of a range of highly personalized AI models across diverse LinkedIn applications, leading to substantial improvements in business metrics as well as refreshed experience for our members.

In this paper, we present NxtPost, a deployed user-to-post content based sequential recommender system for Facebook Groups. Inspired by recent advances in NLP, we have adapted a Transformer based model to the domain of sequential recommendation. We explore causal masked multi-head attention that optimizes both short and long-term user interests. From a user's past activities validated by defined safety process, NxtPost seeks to learn a representation for the user's dynamic content preference and to predict the next post user may be interested in. In contrast to previous Transformer based methods, we do not assume that the recommendable posts have a fixed corpus. Accordingly, we use an external item/token embedding to extend a sequence-based approach to a large vocabulary. We achieve 49% abs. improvement in offline evaluation. As a result of NxtPost deployment, 0.6% more users are meeting new people, engaging with the community, sharing knowledge and getting support. The paper shares our experience in developing a personalized sequential recommender system, lessons deploying the model for cold start users, how to deal with freshness, and tuning strategies to reach higher efficiency in online A/B experiments.

In this paper, we present Que2Search, a deployed query and product understanding system for search. Que2Search leverages multi-task and multi-modal learning approaches to train query and product representations. We achieve over 5% absolute offline relevance improvement and over 4% online engagement gain over state-of-the-art Facebook product understanding system by combining the latest multilingual natural language understanding architectures like XLM and XLM-R with multi-modal fusion techniques. In this paper, we describe how we deploy XLM-based search query understanding model that runs <1.5ms @P99 on CPU at Facebook scale, which has been a significant challenge in the industry. We also describe what model optimizations worked (and what did not) based on numerous offline and online A/B experiments. We deploy Que2Search to Facebook Marketplace Search and share our deployment experience to production and tuning tricks to achieve higher efficiency in online A/B experiments. Que2Search has demonstrated gains in production applications and operates at Facebook scale.

In this paper, we present VisRel, a deployed large-scale media search system that leverages text understanding, media understanding, and multimodal technologies to deliver a modern multimedia search experience. We share our insight on developing image and video understanding models for content retrieval, training efficient and effective media-to-query relevance models, and refining online and offline metrics to measure the success of one of the largest media search databases in the industry. We summarize our learnings gathered from hundreds of A/B test experiments and describe the most effective technical approaches. The techniques presented in this work have contributed 34% (abs.) improvement to media-to-query relevance and 10% improvement to user engagement. We believe that this work can provide practical solutions and insights for engineers who are interested in applying media understanding technologies to empower multimedia search systems that operate at Facebook scale.

In this paper, we present GrokNet, a deployed image recognition system for commerce applications. GrokNet leverages a multi-task learning approach to train a single computer vision trunk. We achieve a 2.1x improvement in exact product match accuracy when compared to the previous state-of-the-art Facebook product recognition system. We achieve this by training on 7 datasets across several commerce verticals, using 80 categorical loss functions and 3 embedding losses. We share our experience of combining diverse sources with wide-ranging label semantics and image statistics, including learning from human annotations, user-generated tags, and noisy search engine interaction data. GrokNet has demonstrated gains in production applications and operates at Facebook scale.

In this paper we present a deployed image recognition system used in a large scale commerce search engine, which we call MSURU. It is designed to process product images uploaded daily to Facebook Marketplace. Social commerce is a growing area within Facebook and understanding visual representations of product content is important for search and recommendation applications on Marketplace. In this paper, we present techniques we used to develop efficient large-scale image classifiers using weakly supervised search log data. We perform extensive evaluation of presented techniques, explain practical experience of developing large-scale classification systems and discuss challenges we faced. Our system, MSURU out-performed current state of the art system developed at Facebook [23] by 16% in e-commerce domain. MSURU is deployed to production with significant improvements in search success rate and active interactions on Facebook Marketplace.

In this paper we present a deployed, scalable optical character recognition (OCR) system, which we call Rosetta , designed to process images uploaded daily at Facebook scale. Sharing of image content has become one of the primary ways to communicate information among internet users within social networks such as Facebook, and the understanding of such media, including its textual information, is of paramount importance to facilitate search and recommendation applications. We present modeling techniques for efficient detection and recognition of text in images and describe Rosetta 's system architecture. We perform extensive evaluation of presented technologies, explain useful practical approaches to build an OCR system at scale, and provide insightful intuitions as to why and how certain components work based on the lessons learnt during the development and deployment of the system.

Online professional social networks such as LinkedIn serve as a marketplace, wherein job seekers can find right career opportunities and job providers can reach out to potential candidates. LinkedIn's job recommendations product is a key vehicle for efficient matching between potential candidates and job postings. However, we have observed in practice that a subset of job postings receive too many applications (due to several reasons such as the popularity of the company, nature of the job, etc.), while some other job postings receive too few applications. Both cases can result in job poster dissatisfaction and may lead to discontinuation of the associated job posting contracts. At the same time, if too many job seekers compete for the same job posting, each job seeker's chance of getting this job will be reduced. In the long term, this reduces the chance of users finding jobs that they really like on the site. Therefore, it becomes beneficial for the job recommendation system to consider values provided to both job seekers as well as job posters in the marketplace.

User experience at social media and web platforms such as LinkedIn is heavily dependent on the performance and scalability of its products. Applications such as personalized search and recommendations require real-time scoring of millions of structured candidate documents associated with each query, with strict latency constraints. In such applications, the query incorporates the context of the user (in addition to search keywords if present), and hence can become very large, comprising of thousands of Boolean clauses over hundreds of document attributes. Consequently, candidate selection techniques need to be applied since it is infeasible to retrieve and score all matching documents from the underlying inverted index. We propose CaSMoS, a machine learned candidate selection framework that makes use of Weighted AND (WAND) query. Our framework is designed to prune irrelevant documents and retrieve documents that are likely to be part of the top-k results for the query. We apply a constrained feature selection algorithm to learn positive weights for feature combinations that are used as part of the weighted candidate selection query. We have implemented and deployed this system to be executed in real time using LinkedIn's Galene search platform. We perform extensive evaluation with different training data approaches and parameter settings, and investigate the scalability of the proposed candidate selection model. Our deployment of this system as part of LinkedIn's job recommendation engine has resulted in significant reduction in latency (up to 25%) without sacrificing the quality of the retrieved results, thereby paving the way for more sophisticated scoring models.

User behavior provides many cues to improve the relevance of search results through personalization. One aspect of user behavior that provides especially strong signals for delivering better relevance is an individual's history of queries and clicked documents. Previous studies have explored how short-term behavior or long-term behavior can be predictive of relevance. Ours is the first study to assess how short-term (session) behavior and long-term (historic) behavior interact, and how each may be used in isolation or in combination to optimally contribute to gains in relevance through search personalization. Our key findings include: historic behavior provides substantial benefits at the start of a search session; short-term session behavior contributes the majority of gains in an extended search session; and the combination of session and historic behavior out-performs using either alone. We also characterize how the relative contribution of each model changes throughout the duration of a session. Our findings have implications for the design of search systems that leverage user behavior to personalize the search experience.