Source-linked AI summary
Determinants of Starting Salaries for Filipino Graduates: An Explainable Machine Learning Approach
Alexander Gabriel A. Aranes, John Michael C. Magpantay, Reginald Neil C. Recario, Jamlech Iram N. Gojo Cruz, Rodolfo C. Camaclang
TL;DR
Philippine research has mainly described graduate employment outcomes without explaining or predicting starting-salary differences. This paper applies explainable machine learning to a noisy, crowd-sourced survey dataset and finds that job role and industry dominate salary prediction over institutional prestige. The finding is supported by SHAP, occupational-text ensemble weighting, and NLI reformulation, with implications for career guidance and future data collection.
Problem
Philippine tracer studies primarily describe employment rates and average pay rather than explaining or predicting why graduates earn different starting salaries.
Method
The paper applies machine learning and SHAP to a crowd-sourced graduate salary survey, then tests feature representations, task formulations, and reasoning approaches.
Results
Job role and industry are the dominant starting-salary determinants, significantly outweighing institutional prestige, with support from SHAP, ensemble weighting, and NLI evidence.
Takeaways & Limitations
Career guidance should prioritize sector and role alignment, while future work should incorporate missing confounders and use more controlled data collection.
Takeaways & Limitations
The salary survey lacks company size, location, and experience information, so the data—not model capacity—constrain predictive performance.
Abstract
from arXiv · showhide
Filipino graduates face a persistent disconnect between educational preparation and labor market outcomes, where starting salary is a key signal of entry-level valuation. Current Philippine research is dominated by descriptive tracer studies that document employment rates but do not explain the determinants of pay. We address this gap using a crowd-sourced survey dataset of graduate responses whose noisy, self-reported nature makes it a challenging prediction target. Applying machine learning to this problem, we identify job role and industry as the dominant determinants of starting salary, significantly outweighing institutional prestige. The strength of this finding is its central contribution: it is corroborated by three independent lines of evidence, namely SHAP attributions, the heavy reliance of the best ensemble on occupational text, and a Natural Language Inference reformulation. These results suggest that career guidance and policy should prioritize sector-specific skills over institutional brand.
1 Introduction
Philippine graduate research has largely described employment outcomes without explaining or predicting differences in starting pay. This paper applies explainable machine learning to identify salary determinants and finds that job role and industry outweigh institutional prestige.
- Motivation: 25.6% of unemployed Filipino adults in 2023 were college graduates, highlighting the scale of education-to-labor-market misalignment.Starting salary is treated as a direct signal of how the labor market values a graduate at entry.
- Research gap: Tracer studies document employment rates and average pay but generally do not explain why graduates earn different salaries or predict outcomes for new cohorts.The paper frames this as a gap between descriptive reporting and predictive, explanatory analytics.
- Approach: Machine learning with SHAP analysis establishes a predictive baseline and identifies influential determinants of graduate starting salary.The baseline also exposes a performance ceiling that motivates analysis of feature representation, model architecture, and task design.
- Central finding: Job role and industry are the dominant starting-salary determinants in the Philippine labor market, significantly outweighing institutional prestige.This conclusion is corroborated by SHAP attributions, ensemble weighting, and NLI reasoning.
- Methodological lessons: Raw occupational text outperforms standard occupational codes, while economically motivated label boundaries outperform statistical quartiles in exploratory extensions.These methodological lessons are presented as transferable to future salary-classification research.
2 Methods
The study uses a two-phase pipeline that turns crowd-sourced graduate salary data into interpretable predictions, then diagnoses the resulting accuracy ceiling. It compares data representations, classifiers, and task formulations, including occupational text and NLI-based reformulations.
- Pipeline: The methodology first builds a classical baseline with SHAP explainability, then diagnoses its performance ceiling across feature, task, and reasoning tracks.The second phase feeds back to corroborate the primary finding.
- Dataset: The “First Pay” survey contains crowd-sourced, self-reported initial employment outcomes from Filipino graduates collected through social media from 2020 through June 2025.The survey originally received 2,934 raw responses across higher education institutions.
- Dataset: Missing company size, geographic region, and experience variables, combined with self-selection, make the dataset a challenging prediction target.These absent determinants constrain the information available to the models.
- Preprocessing: Salary values were converted into four quartile-based income brackets, while university classifications were simplified into private and state categories.Reported salaries were approximately inflation-adjusted using historical Philippine CPI values for comparability across decades.
- Baseline modeling: The data were split 80% for training and 20% for testing, with OHE and CatBoost encoding evaluated across Logistic Regression, SVM, Random Forest, and XGBoost.Hyperparameters were tuned using randomized search with stratified 5-fold cross-validation.
- Exploratory extensions: Exploratory extensions refined the data to 2,763 valid observations and used an 80/10/10 train-validation-test split to examine the ceiling’s source.The extensions replaced PSIC/PSOC codes with raw occupational text and tested transformer representations, economic labels, and NLI.
3 Results and Discussion
Across the experiments, job role and industry emerge as the strongest salary signals, while occupational text and task reformulation improve prediction but do not remove the data-driven ceiling.
- Results: 2,922 observations formed the baseline across four nearly balanced salary quartiles, while exploratory cleaning reduced the dataset to 2,763 observations.The baseline included 25.3% Low, 25.9% Lower-Mid, 23.9% Upper-Mid, and 25.0% High salaries.
- Results: 43.1% accuracy was achieved by Logistic Regression with OHE, matching XGBoost’s F1w of 0.397 and establishing the 4-class accuracy ceiling.CatBoost encoding caused a 67% recall drop for the High salary quartile versus OHE.
- Feature Importance and Economic Determinants: Job Role and Industry were the dominant SHAP drivers, outweighing University Category by approximately 2× in the best-performing model.Year Started also ranked highly, Gender had moderate influence, and Negotiation Status was least influential.
- NLP Extensions: A 6.3-point Weighted F1-score improvement from TF-IDF bigrams over the baseline showed that occupational text preserved salary-relevant specificity discarded by mapped categories.XLM-RoBERTa improved performance further, but its marginal gain over TF-IDF was modest.
- Ensemble Evidence: Approximately 80% of the optimal ensemble weight went to the keyword model, independently converging with SHAP on the importance of role and industry text.The remaining transformer weight was α = 0.2 for the 4-class formulation.
- Extended Results: The 3-class economic formulation improved F1w by 16.7 points, but coarser labels reduce predictive resolution and cannot compensate for absent information such as company size, seniority, location, and experience.The 4-class ensemble and 3-class NLI exceeded their random-guess baselines by +24.1 and +23.7 points, respectively.
4 Conclusion
The study concludes that job role and industry are the primary determinants of Filipino graduates’ starting salaries, outweighing institutional prestige. It also finds that prediction is constrained mainly by task design and noisy survey data, while recommending sector- and role-aligned career guidance and richer future data.
- Job role and industry are the primary determinants of starting salary, significantly outweighing institutional prestige across experimental configurations.This conclusion is supported by SHAP attributions, ensemble reliance on occupational text, and the NLI reformulation.
- Three independent evidence streams support an explainable salary signal rather than a merely predictive result.The evidence comprises game-theoretic SHAP attributions, validation-tuned ensemble weights, and the NLI reformulation.
- The accuracy ceiling reflects task design and data quality, with missing company size, region, and experience leaving dense mid-salary cases difficult to separate.The original quartile boundary falls in the densest salary region, while the survey is noisy and self-reported.
- Future work should use controlled instruments, incorporate missing confounders, and explore NLP-based role clustering while prioritizing sector and role alignment in career guidance.The study also reports that raw occupational text outperforms standard classification codes and economically motivated boundaries outperform statistical quartiles.