HR Analytics for Candidate Selection: Predicting Hiring Decisions with Supervised Learning (SVM, Random Forest, Logistic Regression)

Instituto Tecnológico y de Estudios Superiores de Monterrey Campus Monterrey The Better Hire Predictive HR Analytics and Web Deployment for Outstanding Candidate Selection at Ternium Industry Partner: Ternium Course: Data Science Analysis (TC2004B) Team 3 — “Dinamita” Name Role César Guillermo Vázquez Álvarez Project Manager (PM) Ana Daniela López Dávila Data Scientist, Web/BI Designer Alejandro José Murcia Alfaro ([email protected]) Chief Data Officer (CDO) Camila Navarro Llaven UX/UI Designer Paola Guadalupe Machorro Ortiz Data Engineer Instructors: Ma. Angelida Alarcón Romero; María de los Ángeles Constantino González Monterrey, Nuevo León — May 6, 2022 The Better Hire — Ternium TC2004B — Data Science Analysis Abstract Ternium’s hiring workflow for early talent relied on a manual and fragmented evaluation process, requiring HR teams to review dozens of attributes per applicant and apply rule-based decisions. This project presents an end-to-end Human Resources Analytics (HR Analytics) pipeline to (i) clean and transform a real anonymized dataset of 5,853 applicants with originally 55 features, (ii) train and compare supervised classification models to predict hiring outcomes (Hired: Yes/No), and (iii) deploy a functional web prototype that operationalizes the best-performing model and supports decision making through streamlined input requirements. Eight classifiers were evaluated (K-Nearest Neighbors, Gaussian Naive Bayes, Support Vector Machine, Logistic Regression, Decision Tree, Gradient Boosting, Random Forest, and a Neural Network) using accuracy, precision, recall/sensitivity, specificity, F1-score, and ROCAUC. After hyperparameter tuning, a polynomial-kernel Support Vector Machine achieved the strongest overall performance (accuracy ≈ 0 . 973) while preserving high specificity and competitive ROC-AUC. A Flask-based web application was implemented to allow HR users to input a reduced set of key candidate attributes and obtain an immediate hiring recommendation, thereby reducing operational burden and improving process consistency. Keywords: HR analytics; hiring prediction; supervised learning; SVM; logistic regression; random forest; ROC-AUC; CRISP-DM; web deployment; decision support. 1 The Better Hire — Ternium TC2004B — Data Science Analysis 1 Business Understanding 1.1 Problem Statement The objective of this work is to design and implement a web-based system for Ternium that integrates predictive modeling and analytics to support hiring decisions. The system is required to predict whether a candidate is recommended for hiring while minimizing the number of fields that HR personnel must provide per applicant. The operational intent is to reduce time-to-decision, lower manual effort, and increase consistency relative to purely rule-based screening. 1.2 Business Objectives and Success Criteria From a business perspective, success is defined by the ability to identify outstanding candidates efficiently and at scale while reducing the resources allocated to manual evaluation. In a classification setting, success criteria extend beyond overall accuracy: the solution should reduce the risk of discarding suitable candidates (false negatives) while maintaining strong control over erroneous recommendations (false positives). Therefore, model selection is guided by a balanced assessment across accuracy, precision, recall (sensitivity), specificity, F1-score, and ROC-AUC. 1.3 Current Situation The existing process is based on multiple evaluation stages following initial registration. Candidates are assessed across different areas and classified using labels such as Recommend,Highly Recommend,orDo Not Recommend, culminating in whether the candidate is hired. The decision logic is largely manual and requires reviewing a high number of attributes, which increases workload and can introduce inconsistencies. 1.4 Proposed Solution and Hypothesis The project hypothesis is that outstanding candidates can be identified through supervised learning models that are accurate, measurable, and capable of reducing operational effort. The proposed solution follows the CRISP-DM methodology: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. The final deliverable combines a tuned predictive model with a web prototype that translates model inputs into an HR-friendly interface. 1.5 Social Impact and SDG Alignment The intended social impact is to improve fairness in hiring by reducing purely subjective decisions and supporting more consistent screening. The project is framed as an enabler for equitable access to employment opportunities and as a practical application of data-driven innovation in HR, aligning with goals related to gender equality, decent work, innovation, reduced inequalities, and partnerships for adoption. 2 Data Understanding 2.1 Dataset Overview The analysis uses an anonymized dataset provided by the industry partner, corresponding to hiring-related activity in 2021. After selecting the candidate identifier as index, the dataset contained 5,853 records and originally 55 features, with the majority being categorical fields that describe process stages, academic context, and evaluation outcomes. A key characteristic of the dataset is the presence of substantial missingness, reported as 193,309 missing values across the full table, which materially affects exploratory analysis and motivates careful preprocessing. 2 The Better Hire — Ternium TC2004B — Data Science Analysis 2.2 Variables and Initial Exploration Most variables represent categorical descriptors of the candidate and their progression through evaluation stages, including outcomes in multiple competency areas. Initially, only a small subset of variables were numeric in raw form (e.g., VIPS and NIPS). Early statistical exploration focused on descriptive summaries and basic visual inspection; however, correlation analysis was limited because standard correlation measures are only directly applicable to numerical fields, and missingness patterns introduced additional constraints. 2.3 Data Quality Considerations Two data quality themes were dominant: extensive missingness across several features, and the categorical nature of most attributes. The missingness required decisions about column removal and default values, while the categorical structure required systematic encoding to enable model training. These constraints shaped the preparation pipeline and the definition of a reduced predictor set aligned with stakeholder relevance. 3 Data Preparation 3.1 Standardization, Cleaning, and Filtering Preprocessing began by standardizing column names and ensuring consistent encoding and data types. For the target variable (Hired Yes/No), missing entries were replaced with No , under the operational assumption that nonaffirmative entries correspond to not being hired. Columns with no informational content (entirely missing) were removed, and duplicated columns were resolved by keeping the non-empty version. The dataset was filtered to candidates in Mexico to align the analysis with the intended scope. 3.2 Feature Selection and Stakeholder Constraints Based on partner guidance, VIPS and NIPS were removed because they corresponded to legacy tests no longer used in the selection process. Additional features were removed when they were redundant, constant, or operationally uninformative for the stated objective. The resulting dataset emphasized evaluation outcomes and core indicators that are available in practice and relevant for prediction. 3.3 Categorical-to-Numerical Transformation Given the predominance of categorical variables, the pipeline converted qualitative values into numerical representations suitable for supervised learning. Ordinal mappings were applied where categories have a natural order (e.g., English proficiency from A1 to C2), while common evaluation labels were discretized consistently across multiple columns (e.g., Do Not Recommend=0 , Recommend=1 , Highly Recommend=2 ). Binary flags were normalized ( Yes=1 , No=0 ), and certain missing values in process flags were defaulted to No to ensure complete feature vectors for modeling. 3.4 Final Predictors and Target Definition The target variable was defined as the hiring outcome (Hired Yes/No). Predictors were drawn from the reduced set of evaluation-related features, including area assessments (OperationsQuality, MTTO-DIMA, Commercial-Planning, DIGI-SC, Resto-Soft), group activity and eligibility indicators, English level, and relevant Pymetrics-related fields. This feature set was chosen to balance predictive strength with the practical requirement of limiting the number of inputs required in the deployed interface. 3 The Better Hire — Ternium TC2004B — Data Science Analysis 4 Modeling 4.1 Training and Testing Protocol Models were trained as binary classifiers with classes Yes=1 and No=0 . A standard holdout strategy was used with an 80% training split and a 20% test split. This protocol supports direct comparison across algorithms under consistent evaluation conditions and yields a test set for estimating generalization performance. 4.2 Algorithms Considered Eight supervised learning algorithms were evaluated: K-Nearest Neighbors, Gaussian Naive Bayes, Support Vector Machine, Logistic Regression, Decision Tree, Gradient Boosting, Random Forest, and a Neural Network. The selection reflects common baselines and ensemble methods typically used for tabular classification, as well as linear and kernelized approaches capable of handling complex decision boundaries after appropriate encoding. 4.3 Evaluation Metrics Model performance was assessed using accuracy, confusion-matrix-derived measures, precision, recall (sensitivity), specificity, F1-score, and ROC-AUC. This metric set was used to avoid over-reliance on accuracy alone, especially given the business risk of missing suitable candidates. ROC-AUC provided a threshold-independent view of separability, while sensitivity and specificity captured operational trade-offs between false negatives and false positives. 5 Results 5.1 Comparative Performance and Model Choice Across the evaluated models, results were consistently strong, with baseline experiments yielding high accuracy values in most configurations. However, the final selection emphasized robust performance across multiple metrics, including sensitivity and ROC-AUC, to reduce the likelihood of overlooking strong candidates. Based on these considerations, Support Vector Machine, Random Forest, and Logistic Regression were identified as the strongest candidates for deployment and subsequent tuning. 5.2 Hyperparameter Tuning and Final Performance Hyperparameter tuning was applied to improve generalization performance beyond default settings. After tuning, the three leading models demonstrated similar behavior across metrics; nevertheless, the Support Vector Machine achieved the highest accuracy and maintained strong specificity and competitive ROC-AUC, making it the preferred deployment option. 5.3 Selected Model and Hyperparameters The model selected for deployment was the Support Vector Machine due to its best overall balance of accuracy, error rate, sensitivity, specificity, and ROC-AUC. The tuned configuration used a polynomial kernel with degree 2 and an unconstrained iteration budget. For completeness, the tuned configurations for the other top models were also recorded: Random Forest used an entropy criterion with 90 estimators, and Logistic Regression used an L2 penalty with C= 2. 6 Deployment 6.1 Web Prototype A web application was implemented using Flask to operationalize the predictive pipeline. The prototype was designed to reduce friction for HR users by focusing on the reduced predictor set and offering controlled inputs through predefined options (e.g., dropdowns). Internally, the application performs the same discretization and preprocessing logic used during training, 4 The Better Hire — Ternium TC2004B — Data Science Analysis ensuring consistency between offline evaluation and online inference. 6.2 Operational Validation The prototype was validated using ten candidate profiles with intentionally varied characteristics. Predictions produced by the deployed Support Vector Machine model were compared against the known outcomes from the dataset. In these tests, the system’s recommendations matched the recorded outcomes, providing initial evidence that the pipeline generalizes to realistic user inputs under the constraints of the interface. 7 Recommendations From a business standpoint, the most impactful improvement is to standardize and streamline data capture. A structured web form completed by candidates can reduce manual transcription by HR staff and improve data quality through validation. In addition, limiting the process to a small set of empirically relevant features reduces evaluation time while preserving decision quality as indicated by model performance. From a technical standpoint, the application should be deployed on a stable hosting environment appropriate for internal HR usage. The partner is also advised to evolve certain binary fields into graded scales (for example, multi-level qualitative ratings) to preserve information and potentially improve future model robustness. Finally, the prototype can be extended with candidate search and filtering functions that support HR workflows beyond single-candidate evaluation, enabling cohort analysis and followup actions. 8 Next Steps The CRISP-DM process is iterative, and the model should be revisited as new hiring cycles generate additional data. Future work should include periodic retraining and re-tuning, monitoring for data drift and performance degradation, and collecting structured feedback from HR users to refine the interface and the feature definitions. A governance loop that aligns stakeholders on label definitions and data collection practices will be essential to maintain both predictive performance and fairness objectives over time. 9 References Hunter, J. (2021). New Standards: How Screening Data Is Evolving Hiring Process: Leveraging data to create new talent acquisition standards. Talent Acquisition Excellence, 9(11), 20–21. Pessach, D., Singer, G., Avrahami, D., Chalutz Ben-Gal, H., Shmueli, E., & Ben-Gal, I. (2020). Employees recruitment: A prescriptive analytics approach via machine learning and mathematical programming. Decision Support Systems, 134, 113290. https://doi.org/10. 1016/j.dss.2020.113290 Jaramillo, V. G. (s.f.). Multiserver Service Optimization Based on Mathematical Simulation. Diez, F., Bussin, M., & Lee, V. (2019). Fundamentals of HR Analytics: A Manual on Becoming HR Analytical. Emerald Publishing Limited. Organización de las Naciones Unidas. (2019, August 30). Objetivos de Desarrollo Sostenible (ODS). Objetivos y metas de desarrollo sostenible - Desarrollo Sostenible. Chandler, S. (2017). The AI chatbot will hire you now. Wired. 5