Loading...
Publication
Explicabilidade em modelos de IA aplicados à seleção de recursos humanos
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.73 MB | Adobe PDF |
Advisor(s)
Abstract(s)
Este trabalho tem como objetivo propor um sistema de apoio à decisão para o processo de recrutamento e seleção de candidatos ao cargo de consultor de vendas, por meio da
predição automatizada de notas atribuídas a currículos. Foram empregados algoritmos de aprendizado de máquina supervisionado, incluindo KNN, Random Forest, SVM e redes
neurais MLP, para prever a avaliação de um recrutador humano a partir de 14 atributos técnicos extraídos de currículos reais. Como diferencial, este estudo incorporou técnicas
de explicabilidade como SHAP, LIME e TreeInterpreter, promovendo a explicabilidade na análise preditiva, possibilitando identificar a importância de cada variável tanto local
quanto globalmente. A base de dados foi padronizada, normalizada e complementada com dados sintéticos a fim de mitigar desequilíbrios. A avaliação dos modelos foi conduzida com base em métricas como MAE, MSE, RMSE e R2, além da análise de resíduos e da matriz de confusão. O melhor desempenho entre as redes neurais foi obtido com a MLP (23-12-6-1), treinada com 50% de dados sintéticos, que alcançou MAE de 0,23, MSE de 0,28, RMSE de 0,53 e R2 = 0,97. No entanto, o modelo com maior desempenho geral foi a Random Forest com 1000 árvores, que atingiu MAE de 0,14, MSE de 0,16, RMSE de 0,40 e R2 = 0,98.
This study proposes a decision support system for the recruitment and selection process of sales consultant candidates through automated score prediction based on resumes. Supervised machine learning algorithms, including KNN, Random Forest, SVM, and MLP neural networks, were employed to predict the evaluation of a human recruiter based on 14 technical attributes extracted from real resumes. As a distinctive feature, this study incorporated explainability techniques such as SHAP, LIME, and TreeInterpreter, promoting explainability in predictive analysis and enabling the identification of the importanceof each variable both locally and globally. The dataset was standardized, normalized, and supplemented with synthetic data to mitigate imbalances. Model evaluation was conducted based on metrics such as MAE, MSE, RMSE, and R2, in addition to residual analysis and the confusion matrix. Among the neural networks, the best performance was achieved by the MLP (23-12-6-1), trained with 50% synthetic data, which reached an MAE of 0,23, MSE of 0,28, RMSE of 0,53, and an R2 = 0,97. However, the overall best-performing model was the Random Forest with 1000 trees, which achieved an MAE of 0,14, MSE of 0,16, RMSE of 0,40, and an R2 = 0,98.
This study proposes a decision support system for the recruitment and selection process of sales consultant candidates through automated score prediction based on resumes. Supervised machine learning algorithms, including KNN, Random Forest, SVM, and MLP neural networks, were employed to predict the evaluation of a human recruiter based on 14 technical attributes extracted from real resumes. As a distinctive feature, this study incorporated explainability techniques such as SHAP, LIME, and TreeInterpreter, promoting explainability in predictive analysis and enabling the identification of the importanceof each variable both locally and globally. The dataset was standardized, normalized, and supplemented with synthetic data to mitigate imbalances. Model evaluation was conducted based on metrics such as MAE, MSE, RMSE, and R2, in addition to residual analysis and the confusion matrix. Among the neural networks, the best performance was achieved by the MLP (23-12-6-1), trained with 50% synthetic data, which reached an MAE of 0,23, MSE of 0,28, RMSE of 0,53, and an R2 = 0,97. However, the overall best-performing model was the Random Forest with 1000 trees, which achieved an MAE of 0,14, MSE of 0,16, RMSE of 0,40, and an R2 = 0,98.
Description
Mestrado de dupla diplomação com a UTFPR - Universidade Tecnológica Federal do Paraná
Keywords
Inteligência artificial Recrutamento e seleção Explicabilidade Aprendizado supervisionado Regressão