Recruitment Efficiency Prediction: An End-to-End Recruitment Process Optimization Solution Based on Machine Learning

This article introduces an end-to-end data science project aimed at diagnosing recruitment process bottlenecks using machine learning techniques, building an offer acceptance rate prediction model, and deploying an interactive HR dashboard to support decision-making. The project uses XGBoost models to predict offer acceptance rates, combines SHAP interpretability to analyze key driving factors, and finally delivers a Streamlit dashboard to enable real-time prediction and scenario simulation.

Modern recruitment faces three major pressures: extended cycles, rising costs, and declining offer acceptance rates. Specific pain points include: 33.2% of positions take over 60 days to fill (twice the lean benchmark); 5,000 recruitments result in an annual overspend of $2.66 million; 29% of candidates have an offer acceptance rate below 50%; and there is a lack of pre-offer prediction mechanisms. Most ATS systems do not provide data-driven decision support.

The project sets SMART objectives: diagnose inefficient process links, identify key factors for offer acceptance, build a prediction model (OAR ≥70%), and deploy an interactive dashboard. Core metrics are AUC-ROC ≥80% (actual test set: 0.71, cross-validation: 0.752) and increasing acceptance rate from 65.08% to 80% within five weeks. The dataset contains 5,000 records covering 6 departments, 20 positions, and 4 channels with no missing values; the offer acceptance rate is binary-classified (balanced 50/50) using a threshold of 0.7.

The technology stack includes Python3.10+, Pandas/NumPy (data processing), Scikit-learn/XGBoost (ML), SMOTE (class imbalance), Random Search (hyperparameter tuning), SHAP (interpretability), Streamlit (dashboard), etc. The model uses the XGBoost classification algorithm; after tuning, the test set AUC-ROC reaches 0.71 and cross-validation is 0.752. SHAP analysis is used to explain prediction driving factors, helping HR understand key influencing factors.

EDA reveals: average recruitment cycle is 47.19 days (31% above industry benchmark), single recruitment cost is $5214.83 (11.3% above SHRM benchmark), average acceptance rate is 65.08% (below the healthy level of 80%); there are significant differences in candidate quality and acceptance rates across different channels. The final dashboard RePort features: real-time KPI monitoring, industry benchmark comparison, offer acceptance prediction, scenario simulation, and it has been deployed to Streamlit Community Cloud.

Project value: shift from "post-hoc statistics" to "pre-hoc prediction", optimize cost and resource allocation, guide recruitment strategies, reduce subjective decision bias. Future improvement directions: introduce external data sources (social media, salary data), build real-time recommendation systems, develop mobile applications, etc. This project provides an excellent reference case for data science applications in the HR field.