Enhancing Job Recommendation through LLM-based Generative Adversarial Networks

📝 Paper Summary

Job Recommendation Generative Adversarial Networks (GANs) User Profiling with LLMs

LGIR improves job recommendations by using LLMs to infer skills from user interaction history and a GAN to refine the quality of resume embeddings for users with scarce data.

Core Problem

Job recommendation systems struggle with low-quality user resumes (missing skills) and the 'few-shot' problem where users with limited interaction history provide insufficient data for models to generate accurate profiles.

Why it matters:

Users frequently upload incomplete resumes or lack self-awareness of their skills, leading to poor matches in recruitment platforms
LLMs can hallucinate (fabricate) information when generating resumes for users with little data, worsening the recommendation quality rather than improving it
Existing hybrid methods fail to effectively bridge the gap between users with rich interaction histories and long-tail users with sparse records

Concrete Example: A user might upload a sparse resume listing only 'UI Design'. A standard LLM might hallucinate unrelated projects to fill space. However, if the user clicked on 'Senior App Designer' jobs, LGIR infers implicit skills (e.g., 'Mobile UI', 'Prototyping') to prompt the LLM correctly, then uses a GAN to ensure this generated profile matches the statistical quality of highly active users.

Key Novelty

LLM-based GANs Interactive Recommendation (LGIR)

Interactive Resume Completion (IRC): Prompts an LLM not just with the raw resume, but with implicit characteristics inferred from the user's historical job interactions (e.g., agreed interviews)
GAN-based Alignment: Uses a Generative Adversarial Network to transfer knowledge from 'many-shot' users (rich history) to 'few-shot' users, forcing the embeddings of sparse generated resumes to resemble high-quality ones

Architecture

The LGIR framework architecture, illustrating the parallel processes of Interactive Resume Completion via LLM and the GAN-based representation alignment.

Evaluation Highlights

Outperforms state-of-the-art LightGCN by +8.38% on MAP@5 on the Designs dataset
Achieves +7.50% improvement in MAP@5 on the Sales dataset compared to the best baseline
Consistent improvements across three real-world datasets (Designs, Sales, Tech) covering over 80,000 users

Breakthrough Assessment

7/10

Novel combination of LLM-based profile augmentation with GAN-based feature alignment for the long-tail problem. Solid empirical results on real-world data.

⚙️ Technical Details

Problem Definition

Setting: Top-K Job Recommendation using implicit feedback and textual content

Inputs: User set C, Job set J, text documents T (resumes/job descriptions), and interaction matrix R

Outputs: Matching score g(c_i, j_k) predicting the likelihood of user i interacting with job k

Pipeline Flow

Prompt Construction (Resume + Interactions)
LLM Generation (Completed Resume)
Text Encoding (SimBERT)
Quality Classification (Few-shot vs Many-shot)
GAN Refinement (Generator improves Few-shot embeddings)
Recommendation (Pairwise Matching)

System Modules

LLM Generator

Generate enriched resume text based on user self-description and interaction history

Model or implementation: ChatGLM-6B

Text Encoder

Encode generated resumes and job descriptions into fixed-size vectors

Model or implementation: SimBERT

Resume Quality Classifier (GAN Alignment)

Distinguish between high-quality (many-shot) and low-quality (few-shot) resume representations

Model or implementation: MLP Classifier

Generator (GAN) (GAN Alignment)

Refine low-quality resume embeddings to resemble high-quality embeddings

Model or implementation: MLP

Discriminator (GAN) (GAN Alignment)

Distinguish between real high-quality embeddings and refined low-quality embeddings

Model or implementation: MLP

Novel Architectural Elements

Integration of a GAN-based transfer learning module specifically to align 'few-shot' LLM-generated embeddings with 'many-shot' embeddings within a recommendation pipeline

Modeling

Base Model: ChatGLM-6B (LLM) and SimBERT (Text Encoder)

Training Method: Adversarial Training + Multi-task Learning

Objective Functions:

Purpose: Train classifier to identify resume quality based on interaction volume.

Formally: Cross-entropy loss L_C on labels derived from interaction counts.
Purpose: Train discriminator to distinguish high-quality resumes from refined low-quality ones.

Formally: Minimax GAN loss L_D.
Purpose: Train generator to fool the discriminator.

Formally: Generator loss L_G.
Purpose: Optimize recommendation ranking.

Formally: Pairwise loss L_rec = log(sigmoid(score_pos - score_neg)).

Key Hyperparameters:

latent_dimension_d: 768 (SimBERT) / 64 (Latent Space)
learning_rate: 5e-5
batch_size: 1024
+ 3 more
regularization_lambda: 1e-4
hidden_layer_dimension_de: 64
hidden_layer_dimension_de_prime: 128

Compute: Not reported in the paper

Comparison to Prior Work

vs. LightGCN: LGIR incorporates rich semantic text enhancement via LLMs and GANs, whereas LightGCN focuses on interaction graph structure
vs. SGPT+SRC: SGPT+SRC uses LLMs for simple completion; LGIR adds interactive prompting and GAN alignment to fix few-shot quality issues
vs. SHPJF: LGIR generates explicit resume text to fill gaps, rather than just encoding existing search history vectors

Limitations

Relies on proprietary interaction data for the 'Interactive' prompting strategy
LLM generation latency and cost are not analyzed
Requires careful tuning of thresholds to define 'few-shot' vs 'many-shot' users
Evaluation is limited to offline metrics on static datasets

Reproducibility

Datasets are proprietary (BOSS Zhipin). Code availability is not provided. Implementation details (SimBERT, ChatGLM-6B) are standard, but the specific interaction thresholds and proprietary data make exact reproduction difficult.

📊 Experiments & Results

Evaluation Setup

Top-K job recommendation on three industry-specific datasets (Designs, Sales, Tech)

Benchmarks:

Designs (Job Recommendation)
Sales (Job Recommendation)
Tech (Job Recommendation)

Metrics:

map@5 (Mean Average Precision)
ndcg@5 (Normalized Discounted Cumulative Gain)
mrr (Mean Reciprocal Rank)
Statistical methodology: Paired t-test (p < 0.01)

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
LGIR consistently outperforms baselines across all three datasets, with the largest gains in domains where resumes are likely more subjective or variable (Designs/Sales).
Designs	map@5	0.2664	0.2887	+0.0223
Sales	map@5	0.1629	0.1751	+0.0122
Tech	map@5	0.4719	0.5086	+0.0367
Ablation studies demonstrate that both the Interactive Resume Completion (IRC) and the GAN-based alignment contribute to performance.
Designs	map@5	0.2859	0.2887	+0.0028
Designs	map@5	0.2601	0.2859	+0.0258

Experiment Figures

Performance comparison between LGIR (with GAN) and IRC (without GAN) across different user groups sorted by interaction frequency (shots).

Main Takeaways

Pure LLM-based enhancement (Simple Resume Completion) shows limited or no improvement over baselines, likely due to hallucinations.
Including interaction history in LLM prompts (Interactive Resume Completion) significantly boosts performance by grounding the generation.
The GAN-based alignment module specifically helps 'few-shot' users (those with little data) by aligning their embeddings with 'many-shot' users, as evidenced by performance gains in the long-tail user groups.

📚 Prerequisite Knowledge

Prerequisites

Basics of Recommender Systems (Collaborative Filtering)
Generative Adversarial Networks (GANs)
Large Language Models (LLMs) for text generation
Contrastive or Pairwise Loss functions

Key Terms

GAN: Generative Adversarial Network—a framework where a 'generator' creates data and a 'discriminator' tries to distinguish it from real data, improving the generator through competition

Few-shot problem: In this context, refers to users with very few interaction records (clicks/applications), making it hard to infer their preferences or generate accurate profiles

SimBERT: A BERT-based model fine-tuned for semantic similarity tasks, used here to encode text into vector embeddings

Hallucination: A phenomenon where an LLM generates plausible-sounding but factually incorrect or fabricated information

LightGCN: A state-of-the-art graph convolutional network for recommendation that simplifies the design by removing non-linearities

IRC: Interactive Resume Completion—the paper's method of prompting LLMs with both resume text and interaction history

MAP@n: Mean Average Precision at n—a metric evaluating the quality of the top-n recommended items