In large-scale recommendation systems like LinkedIn’s, the retrieval stage is critical for narrowing billions of potential candidates to a manageable subset for ranking. LinkedIn's feed now serves suggested content based on the topical interests of members, where 2000 candidates are retrieved from several million candidates with a latency budget of a few milliseconds and inbound QPS of several thousand per second. This paper presents a novel retrieval approach that fine tunes a large causal language model (Meta’s LLaMA 3) as a dual encoder to generate high quality embeddings for both users (members) and content (items), using only textual input. We describe the end to end pipeline, including prompt design for embedding generation, techniques for fine tuning at LinkedIn scale, and infrastructure for low latency, cost effective online serving. We share our findings on how quantizing numerical features in the prompt enables the information getting encoded in the embedding facilitating greater alignment between the retrieval and ranking layer. The system was evaluated using offline metrics and an online A/B test, which showed substantial improvements in member engagement. We observed significant gains among newer members, who often lack strong network connections, indicating that high-quality suggested content aids retention. This work demonstrates how generative language models can be effectively adapted for real time, high throughput retrieval in industrial applications.

A post embedding (representation of text in embedding space that effectively captures semantic meaning) is a foundational component of LinkedIn that is consumed by product surfaces in retrieval and ranking (e.g., ranking posts in the feed or video tab). This paper presents the post embeddings used at LinkedIn, where a pre-trained transformer-based large language model (LLM) is taken as input and fine-tuned using multi-task learning across a diverse set of semantic labeling tasks. We observe positive transfer, leading to improved performance across all tasks, compared to training them independently. The generated post embeddings outperform baseline models in zero-shot learning, demonstrating its potential for broader applicability. Furthermore, the generated post embeddings' performance surpasses that of OpenAI's ADA-001 and ADA-002 embeddings on LinkedIn specific datasets and tasks. We also describe the offline evaluation methodology and the deployment to our near-line infrastructure, which makes the post embedding available for use within minutes of post creation for any downstream application. We present how the embeddings were applied in the Feed product surface, in both ranking and retrieval stages, and showcase the real world online impact to demonstrate the superior performance of these embeddings. Finally, we also share the results of applying the embeddings to the retrieval system of our video ranking product surface in LinkedIn. These embeddings have been battle-tested in production at LinkedIn for over two years, consistently powering multiple products.

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.

Closed-loop decoder adaptation (CLDA) is an emerging paradigm for both improving and maintaining online performance in brain-machine interfaces (BMIs). The time required for initial decoder training and any subsequent decoder recalibrations could be potentially reduced by performing continuous adaptation, in which decoder parameters are updated at every time step during these procedures, rather than waiting to update the decoder at periodic intervals in a more batch-based process. Here, we present recursive maximum likelihood (RML), a CLDA algorithm that performs continuous adaptation of a Kalman filter decoder's parameters. We demonstrate that RML possesses a variety of useful properties and practical algorithmic advantages. First, we show how RML leverages the accuracy of updates based on a batch of data while still adapting parameters on every time step. Second, we illustrate how the RML algorithm is parameterized by a single, intuitive half-life parameter that can be used to adjust the rate of adaptation in real time. Third, we show how even when the number of neural features is very large, RML's memory-efficient recursive update rules can be reformulated to also be computationally fast so that continuous adaptation is still feasible. To test the algorithm in closed-loop experiments, we trained three macaque monkeys to perform a center-out reaching task by using either spiking activity or local field potentials to control a 2D computer cursor. RML achieved higher levels of performance more rapidly in comparison to a previous CLDA algorithm that adapts parameters on a more intermediate timescale. Overall, our results indicate that RML is an effective CLDA algorithm for achieving rapid performance acquisition using continuous adaptation.

Closed-loop decoder adaptation (CLDA) is an emerging paradigm for achieving rapid performance improvements in online brain-machine interface (BMI) operation. Designing an effective CLDA algorithm requires making multiple important decisions, including choosing the timescale of adaptation, selecting which decoder parameters to adapt, crafting the corresponding update rules, and designing CLDA parameters. These design choices, combined with the specific settings of CLDA parameters, will directly affect the algorithm's ability to make decoder parameters converge to values that optimize performance. In this article, we present a general framework for the design and analysis of CLDA algorithms and support our results with experimental data of two monkeys performing a BMI task. First, we analyze and compare existing CLDA algorithms to highlight the importance of four critical design elements: the adaptation timescale, selective parameter adaptation, smooth decoder updates, and intuitive CLDA parameters. Second, we introduce mathematical convergence analysis using measures such as mean-squared error and KL divergence as a useful paradigm for evaluating the convergence properties of a prototype CLDA algorithm before experimental testing. By applying these measures to an existing CLDA algorithm, we demonstrate that our convergence analysis is an effective analytical tool that can ultimately inform and improve the design of CLDA algorithms.