Machine Learning Predicts Employment Tenure for Indian Tech Workers Using LinkedIn Data

Researchers analyzing professional networking patterns have demonstrated that machine learning algorithms can predict employment tenure for Indian information technology professionals using public social network data. Published in Discover Analytics by Prajas Naik and Swapnajit Chakraborti, the study examines how career histories, credentials, and networking metrics correspond to job tenure. Unlike conventional retention studies relying on proprietary single-company records, the authors extracted multi-organization professional profiles directly from LinkedIn. By framing employee retention as a continuous regression problem rather than a binary classification task, the team evaluated several predictive algorithms alongside unsupervised clustering techniques. Their findings show that decision tree regressors and artificial neural networks achieve the highest accuracy for estimating tenure length, while revealing that experience, seniority, certifications, and connection counts serve as key variables determining how long an employee remains with an enterprise.

Employee attrition represents a persistent logistical and financial challenge for technology firms worldwide, disrupting project schedules and imposing steep recruitment costs when institutional expertise departs. Human resource departments have historically struggled to anticipate departure timelines reliably, particularly when critical technical personnel vacate specialized assignments. While traditional research has sought to calculate quit probabilities, identifying the specific number of months an employee will remain at an organization provides far greater practical utility for managerial planning. Such forecasts allow leadership to align staff with development cycles whose expected durations do not exceed predicted tenures, permitting targeted interventions before key talent departs. Extant predictive literature, however, has remained heavily anchored in single-firm datasets, most notably the synthetic IBM Human Resources dataset. Because these older models reflect static internal environments, they frequently exhibit severe class imbalances and fail to capture generalizable workforce dynamics.

To overcome the limitations of isolated corporate datasets, the researchers constructed an empirical corpus gathered from public profiles on LinkedIn. The data collection focused on information technology specialists who worked within the Indian tech sector, targeting major metropolitan hubs including Pune, Mumbai, Delhi National Capital Region, Kolkata, Bangalore, and Chennai. To mitigate search visibility constraints inherent to platform connection graphs, the authors initiated queries from established industry accounts possessing over 8,000 followers and filtered returns through third-degree network parameters. Furthermore, the sampling strategy prioritized professionals with a history of at least four career transitions, ensuring that gathered records captured individuals representative of dynamic career progressions. The automated scraping procedure excluded personally identifiable details, discarding names and contact mechanisms immediately upon extraction while omitting demographic attributes such as gender and race from the analytical pipeline.

The resulting raw dataset required comprehensive structural transformation to prepare unstructured resume listings for numerical modeling. Rather than treating each worker as a static entity, the investigators executed a dynamic split that segmented each professional history into discrete, independent job tenure records. This transformation generated 411 unique tenure observations, allowing models to capture evolving career stages where accumulated experience and credentials updated sequentially over time. Active employment positions were omitted from the regression target because definitive departure dates remained unobserved. Degrees were distilled into standardized ordinal hierarchies reflecting highest educational attainment, while technical certifications were mapped to the specific job tenures during which they were completed. Nominal attributes were converted through dummy variable encoding, and ordinal categories were assigned sequential integer ranks. Additionally, to protect algorithmic performance against skewness, the investigators filtered out the top and bottom ten percent of tenure duration outliers.

Exploratory statistical assessments revealed distinctive career patterns across the 411 analyzed employment spells. The sample exhibited a mean overall professional experience of 116 months, roughly ten years, alongside an average completed job tenure of 32 months, or approximately 2.5 years. Company sizes clustered around a median ordinal score of five, corresponding to mid-sized enterprises employing between 1,001 and 5,000 workers, with most records held by senior engineer-level specialists holding master’s degrees. Correlational evaluations using Pearson’s correlation coefficient indicated a weak positive correlation of 0.35 between cumulative experience at resignation and employment duration, with organizational seniority displaying a comparable positive relationship. Technical certifications earned during a tenure also correlated positively with duration. In contrast, educational attainment showed a weak negative correlation with LinkedIn connection counts, indicating that professionals with advanced academic credentials maintained comparatively smaller visible social networks on the platform.

To predict the continuous duration of job tenure in months, the authors compared eight supervised regression algorithms, including Linear Regression, Regression Trees, Random Forest Regressors, Support Vector Regressors, Elastic Net, LightGBM, XGBoost, and Artificial Neural Networks. Hyperparameters were tuned using grid-search cross-validation to prevent overfitting on the modest sample size. Standard Linear Regression produced a cross-validated root mean square error of 15.5912 months and a test set error of 14.2813 months, confirming that variance inflation factors remained below five. Testing polynomial expansions from second through tenth degrees proved counterproductive; although a fifth-degree polynomial drastically lowered training error, test error escalated exponentially, confirming that linear feature relationships preserved far better predictive generalizability. Meanwhile, regularized Elastic Net achieved a cross-validated error of 14.9249 months and a test error of 13.3728 months, with Support Vector Regression yielding closely comparable results.

Among tree-based architectures, the Decision Tree Regressor established the lowest overall error on the evaluation partition, registering a test root mean square error of 12.3899 months alongside a validation error of 14.6014 months. Ensemble configurations likewise demonstrated strong performance, with the Random Forest Regressor yielding a test error of 13.2866 months and attaining the highest coefficient of determination. Gradient boosting implementations converged rapidly, with XGBoost achieving a test error of 13.3131 months and LightGBM recording 13.3340 months alongside superior test mean absolute percentage error metrics. In parallel, an Artificial Neural Network configured with four hidden layers of thirty nodes each, utilizing rectified linear unit activations and stochastic gradient descent under a mean absolute error loss objective, produced remarkable consistency. The neural model yielded a training error of 14.25 months, a test error of 14.05 months, and an average test mean absolute percentage error of 42.8 percent.

Model interpretability techniques, including permutation feature importance metrics and SHAP value analyses applied to the neural network architecture, revealed the underlying factors governing tenure forecasts. Across linear, ensemble, and deep learning algorithms, total accumulated experience emerged as the single most decisive variable, commanding the highest positive regression coefficients and dominant mean absolute SHAP values. High organizational seniority similarly demonstrated strong positive alignment with prolonged employment durations, corroborating earlier findings regarding career stability and compensation. Conversely, social networking metrics introduced an inverse dynamic. Both connection counts and subscriber volumes yielded negative coefficients in linear models and negative SHAP values within the neural network, establishing that professionals with expansive professional networks systematically exhibited shorter tenures. The authors hypothesized that individuals with extensive external networks encounter greater visibility and frequent recruiting solicitations, predisposing them to switch employers more rapidly.

To complement supervised predictions with behavioral categorization, the research team deployed unsupervised clustering algorithms, including Spectral Clustering, K-Means, and Hierarchical Clustering using Gower distances. The analyses converged on a three-cluster solution, validated by an optimal silhouette score of 0.23 and minimized Davies-Bouldin metrics under spectral modeling. The first group, termed Experienced Professionals, comprised 148 tenures characterized by a high average experience of 203.36 months, a mean seniority rank of 4.44 reflecting managerial standing, and an extended average tenure of 47.10 months within larger enterprises. The second cohort, designated Freshers, represented 244 tenures with a mean experience of 68.32 months and an average duration of 24.30 months. The final cluster, labeled Influential Freshers, encompassed 19 tenures representing low-experience individuals averaging 59.84 months in the workforce who commanded massive platform influence, possessing median connection counts of 29,883 and experiencing brief average tenures of 17.73 months.

Synthesizing these empirical observations, the authors proposed the Categorize-Train-Promote framework as an operational retention roadmap for corporate human resource departments. Under this approach, organizations first categorize newly onboarded personnel into retention risk tiers by evaluating cumulative experience, platform connectivity, and educational background without relying on discriminatory personal characteristics. For individuals identified within elevated churn risk profiles, employers implement targeted upskilling initiatives and technical training programs. Because empirical modeling confirmed that technical certifications positively correlate with lengthened job tenure, employer-sponsored certifications provide a viable mechanism to counterbalance turnover risks associated with extensive social networking influence. Finally, the framework emphasizes structured, timely promotions for developing personnel. Since higher organizational seniority systematically predicts prolonged commitment across both regression models and cluster profiles, regular advancement trajectories offer employees enhanced professional safety and career progression, effectively converting flight-prone talent into stable, long-term contributors.

The authors outline vital ethical safeguards to prevent potential misuse of automated tenure prediction systems during hiring workflows. While predictive models offer practical utility for project scheduling and retention planning, utilizing predicted tenure duration as a rigid automated screening cutoff poses substantial fairness concerns. Arbitrary exclusion thresholds risk rejecting capable job seekers who may have exhibited abbreviated past tenures due to unfavorable work environments or uncompetitive compensation rather than personal disloyalty. The researchers insist that machine learning outputs should serve strictly as advisory diagnostics to guide post-hire talent support rather than as gatekeeping filters during recruitment. Because the underlying model incorporated only publicly shared professional achievements and intentionally excluded sensitive biographical factors, ethical organizational deployment requires human managers to examine candidate backgrounds holistically, ensuring distributive justice while targeting proactive mentoring and career development toward high-risk candidates.

Subject of Research: Machine learning prediction of employment tenure among Indian information technology professionals using LinkedIn network data

Article Title: Employment tenure prediction for Indian IT professionals using linkedin network data and machine learning based techniques

Article References: Naik, P., & Chakraborti, S. (2025). Employment tenure prediction for Indian IT professionals using linkedin network data and machine learning based techniques. Discover Analytics, 3(1), Article 20. https://doi.org/10.1007/s44257-025-00052-3

Image Credits: AI Generated

DOI: 10.1007/s44257-025-00052-3

Keywords: employee attrition, employment tenure prediction, machine learning, decision tree regressor, artificial neural network, LinkedIn network data, Indian IT industry, spectral clustering, SHAP analysis, job seniority, professional experience, technology certifications, random forest regressor, elastic net, LightGBM, XGBoost, CTP framework, human resources analytics