PAIR: Preference-aligned retrieval for complementary professional intent

简体中文

L1 component; diagnostic only. Full Premium recommendation training and production evaluation not reproduced.

Paper information

Field	Details
Paper	PAIR: Preference-Aligned Intent Retrieval for Premium People Recommendation at LinkedIn
Affiliations	First author Zhongkai Shangguan; all authors list LinkedIn.
First public release	2026-10-22 08:57:58 UTC
Author code	No PAIR code release or pinned source/license found in the paper or linked official sources. The paper's CC0 license does not release unpublished code.
Local key	pair; no adapter
Local code	src/auto_research/recommendation_components/pair.py; tests/test_oct23_root_components.py
Local evidence	Local mechanism metrics; A100 validation receipt

What the paper changes

Professional intent can be complementary rather than similar: a job seeker may need a recruiter. Train a shared profile/document encoder and short-term history tower from clicks, unclicked candidates in the same exposure and random members. This is multi-negative contrastive learning with exposure-conditioned supervision, not token-level DPO; there is no reference model or KL.

Original paper figure

Original figure from the reviewed paper, reused without redrawing. See figure provenance.

Core equations

\[ \alpha_i=\operatorname{softmax}_i[w^\top(e_{h_i}+PE(i))],\quad e_H=\sum_i\alpha_i e_{h_i}, \]

\[ e_q^{final}=\operatorname{L2Norm}(\operatorname{L2Norm}(e_q)+\gamma\operatorname{L2Norm}(e_H)). \]

\[ \mathcal L=\sum_{d\in\mathcal D}\left[-s_d(q,d^+)+\log\sum_{c\in\{d^+,D^-\}}\exp s_d(q,c)\right]. \]

Newest click is position one. Learned position embeddings and a global attention vector are independent of the current profile query. Score candidates by temperature-scaled cosine similarity with the positive in slot zero. Sum Matryoshka losses across prefixes rather than averaging them. Paper defaults are \(K=10\), temperature 0.02 and dimensions 200/256/512/1024. History updates are \(O(Kd)\) on a fixed window, not window-independent constant time.

Results reported by the authors

Section 5.5/Table 4 reports a controlled production A/B: CTR +7.5%, Premium retention +2.1% and page latency change 0.1%; offline MAR@100 is 0.450. Exposure bias remains, and results apply to Premium Browsemaps with a fixed downstream ranker, not all recommendations.

Local reproduction and limits

Learned positional attention, normalized fusion, exposure negatives prioritized with random fill, prefix loss and real small-tensor gradients are implemented. Tests compare scalar cosine/CE and gradients, run three updates and verify masks, chronology and exclusion of clicked positives from exposure negatives. These are not 0.6B semantic encoding, model accuracy, online gain or Evolve integration.

PYTHONPATH=src python -c 'from auto_research.recommendation_components.pair import diagnostic; print(diagnostic(42))'
PYTHONPATH=src python -m pytest tests/test_oct23_root_components.py

Full experiment: not reproduced

Private data comprises 209K training instances and 4.4M profiles. The full training uses eight GPUs and production traffic. No data download, full recommender or automatic retry.

Additional references

Full paper · Paper Figure 1