PaperlayerView exact revision on arXiv

Source-linked AI summary

PAIR: Preference-Aligned Intent Retrieval for Premium People Recommendation at LinkedIn

Zhongkai Shangguan, Yanbin Kang, Yihan Cao, Ruijie Jiang, Qiannan Yin, Chang Yue, Zijian Zhou, Girish Kathalagiri, Changshuai Wei, Zhengming Xing

arXiv:2610.30941v1cs.IRcs.LG

TL;DR

Professional recommendation can misrank people when profile similarity misses complementary, goal-driven intent. PAIR trains embeddings from exposure-conditioned click preferences and aggregates click history for dynamic intent. On LinkedIn Premium retrieval, a 0.6B model achieves 0.450 MAR@100, and logged preferences remain conditioned on historical exposure.

  • Problem

    Profile-similarity retrieval can miss complementary professional matches, and downstream ranking cannot recover candidates excluded during retrieval.

    Source: §3.2 · pp. 3–4

  • Method

    PAIR trains embeddings from exposure-conditioned click preferences and combines profile representations with positional click-history embeddings for efficient intent updates.

    Source: §1 · p. 2, §4.3 · p. 5

  • Results

    0.450 MAR@100: a 0.6B-parameter PAIR model improves 56% relative to Holistic-171 and 2.3× over the off-the-shelf base encoder.

    Source: §6 · p. 8

  • Takeaways & Limitations

    PAIR supports dynamic intent updates through O(Kd) vector operations and multiple embedding dimensions without online LLM re-inference.

    Source: §6 · p. 9

    Because preferences are observed only within exposed candidate sets, PAIR inherits selection bias from the historical retrieval policy.

    Source: §4.1 · p. 5

Abstract

from arXiv · show

General-purpose embeddings trained on profile similarity often misrank candidates in professional networks, where relevance is inherently complementary rather than symmetric. For example, job seekers usually engage with recruiters rather than peers with similar credentials. Current user recommendation systems are often built on profile similarity or social signals such as shared connections, which are effective in friend-oriented social networks where users tend to connect with similar people or within existing social circles. However, this does not hold in professional platforms like LinkedIn, where interactions are often goal-driven and cross-role interactions, requiring matching based on complementary intents rather than similarity or social proximity. In this paper, we propose Preference-Aligned Intent Retrieval, a retrieval framework that is designed to model user intent preference instead of profile similarity. Our approach is inspired by the preference-learning perspective of Direct Preference Optimization, but instantiates preference learning as an exposure-conditioned multi-negative contrastive objective over embeddings. This formulation simplifies training and scales efficiently to large candidate sets without requiring an explicit reference model. Furthermore, we incorporate user interaction history by aggregating click histories with positional encoding to capture temporal order, which enables constant-cost updates of user representations over a fixed history window, allowing new interactions to be incorporated without re-encoding history. As a result, personalization can be updated efficiently using lightweight CPU operations without repeated LLM inference during serving. We evaluate PAIR on LinkedIn's Premium retrieval task. A 0.6B parameter model achieves a MAR@100 of 0.45.

1 Introduction

Professional people retrieval must match complementary, goal-driven intent rather than profile similarity, because retrieval determines which candidates downstream ranking can consider. PAIR uses exposure-conditioned preferences and click history, and a 0.6B model improves offline and online Premium Browsemaps results.

1 Introduction

  • Profile-similarity embeddings can miss complementary professional matches, while candidates excluded during retrieval cannot be recovered by downstream ranking.In offline evaluation, Holistic-171 scores 0.288 MAR@100 and the off-the-shelf 0.6B LLM embedding scores 0.192.Source: §1 · pp. 1–2
  • Production deployment requires preference-based supervision, dynamic intent updates without repeated profile re-encoding, and flexible embedding dimensions for differing latency, memory, and storage budgets.The paper also analyzes training and serving tradeoffs, including mixed-precision stability, temperature sensitivity, history aggregation, and dimensionality.Source: §1 · p. 2
  • PAIR trains from production exposure logs, treating clicked candidates as preferred over exposed-but-unclicked alternatives rather than using arbitrary in-batch negatives.The production retriever defines the exposure distribution, and clicks provide preferences within it.Source: §1 · p. 2
  • A triple-tower architecture combines static profile embeddings with positional-attention embeddings of recent clicks, enabling intent updates through O(Kd) operations without online LLM re-inference.PAIR also uses Matryoshka training to support embeddings truncated to different dimensions for serving budgets.Source: §1 · p. 2
  • 0.450 MAR@100: a 0.6B-parameter PAIR model improves 56% relative to Holistic-171 and 2.3× over the off-the-shelf base encoder; online it lifts click-through rate by 7.5% and Premium retention by 2.1%.The online test reports no measurable latency regression.Source: §1 · p. 2
  • The paper evaluates PAIR through offline retrieval evaluation and an online A/B test in Premium Browsemaps.These evaluations report engagement and Premium retention improvements without measurable serving latency regression.Source: §1 · p. 2

2 Related Work

Prior retrieval and recommendation work addresses embedding quality, production-scale retrieval, logged preferences, and user history, but professional intent alignment remains distinct. PAIR adapts preference learning to exposure-conditioned embedding retrieval and supports low-cost history updates and flexible embedding dimensions.

LLM-Based Embeddings for Retrieval

  • Prior LLM retrieval work improves embedding quality, while recommendation work applies LLMs to generative, augmentation, and hybrid roles; PAIR instead targets preference-aligned professional retrieval.The cited examples include E5 [28], synthetic-data-driven training [29], LLM2Vec [1], and NV-Embed [12].Source: §2 · p. 2

Embedding-Based Retrieval at Production Scale

  • Production retrieval research addresses scalable dual encoders and sampling bias, but semantic or collaborative similarity can misalign with goal-driven professional relevance.The cited approaches include DPR [9], Facebook Search retrieval [7], and YouTube negative-sampling bias correction.Source: §2 · pp. 2–3

Preference Learning from Logged Feedback

  • PAIR draws on preference learning but trains embeddings with clicks and production-exposed negatives, avoiding a separate reference-model forward pass.Exposed-but-unclicked candidates have passed the production retrieval filter; random negatives provide broader coverage.Source: §2 · p. 3

User History Modeling in Retrieval

  • History-aware retrieval can capture multiple or evolving intents, but sequence inference and state updates may be costly; PAIR aggregates history in embedding space for fixed-window updates without online LLM inference.The design trades some sequence-modeling expressiveness for predictable, low-cost serving behavior.Source: §2 · p. 3

Flexible Embedding Representations

  • Matryoshka loss lets one PAIR model serve multiple embedding dimensions, from full 1024-dimensional embeddings to truncated representations, without retraining separate models.The dimensions accommodate different latency and storage budgets.Source: §2 · p. 3

3 System Overview and Problem Formulation

Premium Browsemaps retrieves profiles for Premium subscribers through a candidate-generation stage followed by downstream ranking. This setting motivates retrieval based on complementary professional intent and combines stable profile identity with dynamic click history, trained using production impressions.

3.1 Premium Browsemaps Product Surface

  • Browsemaps recommends Premium profiles on member pages, with profile visits contributing to downstream Premium-retention signals.The product is designed to support professional discovery and broader exposure among Premium members.Source: §3.1 · p. 3
  • Browsemaps first filters the Premium pool and retrieves about 100 candidates; a downstream model ranks them and displays four recommendations.Because retrieval defines the candidate space, later ranking cannot recover intent-aligned candidates it excludes.Source: §3.1 · p. 3

3.2 Retrieval Problem in Professional Discovery

  • Professional discovery is often complementary rather than symmetric: job seekers may seek recruiters or hiring managers, so profile similarity can miss candidates matching current intent.Other cross-role examples include business developers seeking clients or partners and recruiters sourcing complementary skills.Source: §3.2 · pp. 3–4
  • The retrieval objective is to score preferred candidates above non-preferred alternatives, since excluded candidates cannot be recovered by downstream ranking.The formulation treats retrieval as preference-aligned ranking over a query user and candidate pool.Source: §3.2 · p. 4

3.3 Modeling Static Identity and Dynamic Intent

  • User representations combine static profile identity with dynamic intent from the K most recently clicked profiles, whose embeddings are precomputed.The profile text includes career and other profile fields; users without populated profile fields are excluded from the candidate pool.Source: §3.3 · p. 4

3.4 Data Construction

  • Training uses Premium Browsemaps impression logs, treating clicks as positive preferences and sampling a 1:3 positive-to-negative pool of exposed-but-unclicked and random candidates.Preferences are relative to displayed candidates, and a time-based split keeps training interactions before test interactions.Source: §3.4 · p. 4

4 Method

PAIR addresses three production constraints: preference alignment beyond profile similarity, efficient history updates, and flexible embedding dimensions. It combines exposure-conditioned preference learning with intent-aware architecture and serving choices designed for large-scale retrieval.

4.1 Preference-Aligned Intent Retrieval

  • PAIR optimizes preference ordering using clicked candidates and exposed-but-unclicked alternatives, which represent realistic candidates selected by the deployed retrieval policy.The objective has the computational form of multi-negative contrastive learning but uses preference supervision within the serving policy’s exposure support.Source: §4.1 · p. 5

4.2 Triple-Tower Intent Architecture

  • The triple-tower architecture separates query identity, candidate profiles, and behavioral intent, with shared parameters for the query profile and document towers.The final query representation combines profile identity and dynamic intent through normalized additive fusion.Source: §4.2 · p. 5, Figure 1 · p. 6

4.3 Additive History with Positional Attention

  • Position-aware attention aggregates the K most recent clicked-profile embeddings into a history vector, weighting their positions with learned positional encoding.The query-agnostic design lets history updates proceed independently of the query profile tower; with a fixed window, updates require O(Kd) vector operations and no LLM re-inference.Source: §4.3 · p. 5

4.4 Dimension-Flexible Embedding Training

  • Matryoshka training optimizes PAIR across 200, 256, 512, and 1024-dimensional prefixes, so embeddings can be truncated at serving time without retraining.The dimensions support different latency and storage budgets, allowing one model to serve multiple retrieval surfaces.Source: §4.4 · pp. 5–6

4.5 Production Serving Considerations

  • Production serving computes profile and candidate embeddings offline, updates intent through lightweight vector aggregation, and selects embedding dimensions to fit product budgets.Premium Browsemaps uses the full representation, while latency-sensitive surfaces can truncate embeddings without a separately trained model.Source: §4.5 · p. 6

5 Experiments

The experiments assess offline retrieval quality, history and dimensionality tradeoffs, and whether retrieval gains translate into online engagement.

5.1 Experimental Setup

  • The time-based split contains 209K training and 11.8K test instances; evaluation uses one clicked positive and three negatives, with MAR@k on a fixed 100-candidate pool.The base encoder has 0.6B parameters, and models use bf16 after fp16 caused gradient overflow.Source: §5.1 · pp. 6–7, Table 1 · p. 7

5.2 Main Offline Results

  • MAR@100 reaches 0.450 with P-ATT history, improving 56% over Holistic-171 and 2.3× over the untuned 0.6B encoder.The history-free two-tower PAIR model reaches 0.436, a 51% improvement over the production baseline.Source: §5.2 · p. 7, Table 1 · p. 7
  • Adding P-ATT history raises MAR@100 from 0.436 to 0.450, a 3.2% relative gain; ACC above 0.96 is near saturation.The authors therefore treat MAR@k on the 100-candidate pool as the more informative retrieval metric.Source: §5.2 · p. 7, Table 1 · p. 7

5.3 History Augmentation Analysis

  • P-ATT leads GeM modestly; query-agnostic recency and item weighting keeps updates independent of query profiles, while some simpler pooling underperforms history-free retrieval.History methods aggregate precomputed profile embeddings without LLM re-inference during updates.Source: §5.3 · p. 7, Table 2 · p. 7

5.4 Dimension and Serving Tradeoff

  • Full 1024-dimensional embeddings perform best, while the 256-dimensional variant reaches MAR@100 of 0.432 at 25% of full-model storage.One Matryoshka-trained model supports multiple dimensions without retraining separate models.Source: §5.4 · p. 8, Table 3 · p. 8

5.5 Online A/B Test

  • In the controlled A/B test, PAIR raises CTR by 7.5% and Premium retention by 2.1%; latency changes by 0.1% and coverage remains effectively unchanged.Both arms use the same downstream ranker; the treatment uses PAIR with P-ATT history at 1024 dimensions.Source: §5.5 · p. 8, Table 4 · p. 8
  • The online test uses a fixed downstream ranker and measures short-term outcomes on Premium Browsemaps; longer-term effects and other surfaces remain unvalidated.The authors also identify broader comparisons of random-only, impression-only, and mined hard negatives as future work.Source: §5.5 · p. 8

6 Conclusion

PAIR learns professional retrieval embeddings from exposure-conditioned preferences, reaching 0.450 MAR@100 on Premium retrieval. It combines profile identity with dynamic click history and supports dimension-flexible serving.

6 Conclusion

  • A 0.6B-parameter PAIR model achieves MAR@100 of 0.450, 56% above Holistic-171 and 2.3× the off-the-shelf encoder.The conclusion attributes the improvement to task-specific preference supervision reshaping a general embedding space for professional discovery.Source: §6 · p. 8
  • Normalized additive fusion combines static profile identity with dynamic click-history signals, updating history through O(Kd) vector operations without serving-path LLM re-inference.The update cost applies to a fixed history window K.Source: §6 · pp. 8–9
  • Matryoshka training lets one model serve multiple embedding dimensions, trading retrieval efficiency against quality without retraining separate models.The dimensions can accommodate surfaces with different latency and storage budgets.Source: §6 · p. 9

Related papers

Loading 2610.30941v1…