Professional Network Matters: Connections Empower Person-Job Fit

Hao Chen, Lun Du, Yuxuan Lu, Qiang Fu, Xu Chen, Shi Han, Yanbin Kang, Guangming Lu, Zi Li

TL;DR

This paper introduces a novel two-stage approach to improve Person-Job Fit (PJF) by leveraging professional networks. It constructs a "Workplace Heterogeneous Information Network" (WHIN) to capture rich relationships and uses a "Contextual Social Attention Graph Neural Network" (CSAGNN) to filter out noise and extract relevant insights from these networks, leading to superior job matching performance.

What Problem Does This Paper Address?

Online recruitment platforms aim to match job seekers with suitable positions, a task known as Person-Job Fit (PJF). While existing methods primarily rely on textual information from profiles and job descriptions, or historical application data, they often overlook a crucial aspect: the social connections within a professional network. This paper argues that these connections, despite potential noise, hold valuable information that can significantly enhance PJF. The challenge lies in effectively utilizing this noisy social information to improve matching accuracy.

Methods

The proposed approach is a two-stage framework:

  1. Workplace Heterogeneous Information Network (WHIN) Pre-training:
    • A heterogeneous graph is constructed to represent various entities (members, jobs, skills, companies, schools) and their relationships.
  2. This WHIN is used to pre-train entity embeddings using a link-level prediction task. This stage aims to capture rich structural and semantic information, including professional connections.
  3. A heterogeneous graph neural network (GNN) is employed for this pre-training.
  4. Contextual Social Attention Graph Neural Network (CSAGNN):
    • This stage takes the pre-trained embeddings from WHIN and combines them with BERT-based text embeddings for members and jobs.
  5. A key innovation is a job-specific attention mechanism that re-weights professional connections based on the skills required by a specific job. This helps to focus on relevant connections and filter out noise.
  6. A neighbor sampler module further refines this by using similarity between members and required skills to filter out irrelevant professional connections.
  7. The model then uses a GNN architecture with message passing to aggregate information from relevant professional connections, enhancing the member's representation for the specific job.

Datasets

The paper evaluates its approach on three real-world datasets from LinkedIn, categorized by industry:

  • Tech Industry Dataset: Focuses on technology-related jobs and members.
  • Finance Industry Dataset: Focuses on finance-related jobs and members.
  • Hybrid Dataset: Combines members and jobs from various industries (healthcare, education, semiconductors) to test the model's robustness in handling higher social noise and diverse professional networks.

These datasets contain approximately 200,000 entities and 10 million links.

Key Insights and Ideas

  • Professional networks are a powerful, yet noisy, signal for Person-Job Fit: The paper strongly advocates for incorporating social connections, demonstrating that even with noise, the signal is valuable.
  • Job-specific attention is crucial for filtering social noise: The core idea of making the attention mechanism sensitive to the target job's requirements is key to extracting relevant information from connections.
  • Pre-training on a rich heterogeneous graph (WHIN) provides a strong foundation: The WHIN pre-training stage effectively captures diverse entity relationships and provides meaningful initial embeddings, especially for skills.
  • Heterogeneous information is more than just text: The success of the model highlights the importance of integrating structural information (connections, company, school) alongside textual data.

Key Results

  • Superior Performance: CSAGNN consistently outperforms several state-of-the-art baseline models (including context-based, collaborative filtering, and other GNN-based approaches) across all three datasets, particularly on the hybrid dataset which has higher social noise.
  • Significant AUC Improvement: CSAGNN achieved notable improvements in AUC (Area Under the ROC Curve) compared to strong baselines like HGT, demonstrating its effectiveness in ranking job recommendations. For instance, it improved AUC by 1.39%, 1.71%, and 2.37% compared to HGT on the Tech, Finance, and Hybrid datasets, respectively.
  • Ablation Studies Confirm Component Importance: Removing specific components like job-specific attention (w/o A) or professional connections (w/o S) led to performance degradation, confirming their contribution. WHIN pre-training embeddings were found to be crucial, especially for enhancing performance on industry-specific datasets.

Limitations and Future Work

  • Privacy Concerns: The paper acknowledges that using professional networks raises privacy concerns. They emphasize the need for proper consent mechanisms and limiting access to pertinent information.
  • Demographic Bias: The authors note that professional networks can exhibit demographic clustering, potentially leading to biases in recommendations. Future work should audit for demographic equity and prevent the use of protected class information.
  • Computational Overhead: For large-scale applications, reducing the computational overhead of CSAGNN is identified as a key area for future work.
  • Broader Applications: The WHIN pre-training method is suggested to be promising for other applications like skill completion and professional connection recommendations.

Summary updated

Abstract

Online recruitment platforms typically employ Person-Job Fit models in the core service that automatically match suitable job seekers with appropriate job positions. While existing works leverage historical or contextual information, they often disregard a crucial aspect: job seekers' social relationships in professional networks. This paper emphasizes the importance of incorporating professional networks into the Person-Job Fit model. Our innovative approach consists of two stages: (1) defining a Workplace Heterogeneous Information Network (WHIN) to capture heterogeneous knowledge, including professional connections and pre-training representations of various entities using a heterogeneous graph neural network; (2) designing a Contextual Social Attention Graph Neural Network (CSAGNN) that supplements users' missing information with professional connections' contextual information. We introduce a job-specific attention mechanism in CSAGNN to handle noisy professional networks, leveraging pre-trained entity representations from WHIN. We demonstrate the effectiveness of our approach through experimental evaluations conducted across three real-world recruitment datasets from LinkedIn, showing superior performance compared to baseline models.
BibTeX
@inproceedings{chen2023professional,
  author = {Hao Chen and Lun Du and Yuxuan Lu and Qiang Fu and Xu Chen and Shi Han and Yanbin Kang and Guangming Lu and Zi Li},
  title = {{Professional Network Matters: Connections Empower Person-Job Fit}},
  booktitle = {WSDM},
  year = {2023},
  doi = {10.1145/3616855.3635852},
}
Figure