Recoverable versus Permanent Disengagement: A Sulfur-Poisoning Model of Employee Attrition

Catalysts exposed to sulfur lose activity in two different ways. Part of the damage is temporary: sulfur molecules sit on the active sites and crowd out the reactants, and they come off when the sulfur source is removed. The rest is permanent. Sulfur slowly reacts with the metal to form a stable sulfate that stays even after the feed is clean, and undoing it requires a harsh, costly regeneration step. This note proposes modeling employee attrition the same way. Work stressors such as overtime, a closed communication climate, or burnout would lower an employee's engagement through two channels. One channel tracks current conditions and recovers when they improve. The other accumulates with total past exposure and does not recover under ordinary interventions.

The practical stakes come from how retention interventions are recommended. Current attrition models score risk from an employee's present features. Counterfactual tools then propose the smallest feature change that flips a prediction from "quits" to "stays." If some disengagement is permanent, those recommendations will promise too much exactly for the employees who have absorbed the most stress. They will also misallocate retention budgets. The claim tested here is concrete. A two-channel model with history dependence should predict who quits about as well as flexible baselines on small, imbalanced HR panels. It should also predict something those baselines cannot: how much an intervention helps an employee, as a declining function of that employee's cumulative past stress.

Core Claim

The mechanism comes from catalysis, and HR is the application domain. In catalysis, the two-pathway picture is well supported. Sulfur dioxide and hydrogen sulfide bind more strongly than the reactants and block noble-metal sites, as a review of carbon monoxide oxidation catalysts describes (Progress in Catalytic Oxidation of Noble Metal-Based Carbon Monoxide). Over time, bulk sulfates form as thermodynamic sinks. In Exploring PtSO4 and PdSO4 phases, DFT calculations show palladium sulfate stays stable against decomposition up to about 775 K. Removing it therefore needs high-temperature regeneration, which itself damages the catalyst. Manganese and cobalt oxide catalysts show the same pattern, with sulfates that permanently destroy acid sites (Catalytic oxidation of NO over MnOx–CoOx/TiO2, Ordered Mesoporous MnAlOx Oxides).

A second catalysis observation matters here. Isolated single-atom sites have no neighbors to fall back on, so one sulfur event kills the whole unit (Activating low-temperature diesel oxidation by single-atom Pt on TiO2). Metal ensembles degrade more gradually.

Each catalysis idea has an HR counterpart that current HR models handle only partly:

  • Mediated effects. HR studies find that working conditions affect retention only through mediators such as job satisfaction, well-being, or motivation. This holds for teachers (Fessehatsion), young Chinese workers (Al Mamun et al.), nurses (Li et al.) and manufacturing employees (Sheng and Dewa). These studies motivate the model, but they are cross-sectional. They cannot tell a mediator that recovers from one that does not.
  • The single-site analogy. Exploiting the Contagious Effect for Employee Turnover Prediction reports that a manager's departure causes an immediate, large shock, while peer departures produce delayed ripples. The model therefore predicts that employees whose attachment rests on one tie will show a step-like rise in risk when that tie leaves. Employees with many ties should show a buffered response.

Formally, let $b_i(t)$ be the reversibly blocked share of employee $i$'s engagement. It relaxes toward a saturating function of current stress $s_i(t)$, following the competitive-adsorption form $b^\ast = Ks/(1+Ks)$ with sensitivity $K$. Let $p_i(t)$ be the permanent share. It grows with exposure only, at rate proportional to $s_i(t)(1-p_i(t)-b_i(t))$, and resets only partly after a costly "regeneration" event such as a role change. The two-channel claim is that quit risk depends on the remaining engagement:

$$h_i(t) = h_0(t)\,\exp\big(\beta\,[1-b_i(t)-p_i(t)] + \gamma^\top x_i\big),$$

Here $h_i(t)$ is the quit hazard, $h_0(t)$ a baseline hazard, $x_i$ static covariates such as tenure and role, and $\beta, \gamma$ learned coefficients. Pulse-survey satisfaction is treated as a noisy reading of $1-b_i-p_i$.

The structure makes one prediction that standard models do not. After an intervention lowers $s_i$, only $b_i$ recovers. The expected retention gain should therefore shrink as $p_i$ grows, and $p_i$ is set by integrated past stress rather than by the current state.

Research Program

The work has three parts.

Estimation. We will write the model as a continuous-time state-space survival model and fit it by variational inference. Estimation is shared across employees, and the catalysis-derived functional forms act as the inductive bias. This targets two open problems the HR file names: deep tabular models overfit small, imbalanced cohorts with roughly 15–19% turnover (A transformer-based deep learning framework to predict employee attrition), and practitioners need interpretable levers (A Comparison of Machine Learning Approaches for Predicting Employee Attrition).

Identifiability check where ground truth exists. Telling a reversible component from a permanent one using only noisy activity traces is hard. Before trusting the method on people, we will fit the same estimator to sulfur-poisoning and regeneration curves. In those curves, the split between reversible and permanent loss is measured independently by surface analysis of sulfate (X-ray photoelectron spectroscopy), and by recovery after the sulfur feed is switched off. A natural system is carbon monoxide oxidation over Pd-Au/FeOx/Al2O3 under SO2 and H2S (The Enhancement of CO Oxidation Performance and Stability in SO2 and H2S Environment). This is the only role catalysis plays as data. If the estimator cannot recover a known reversible/permanent split from catalyst traces, it should not be used on employees.

HR deployment. The model needs longitudinal panels: repeated satisfaction or burnout measures, logged stressors such as overtime and manager changes, recorded interventions, and quit dates. Most public HR data are cross-sectional. The IBM HR Analytics dataset is also synthetic, so it can test only the static part of the model. The main test therefore needs a partner organization's pulse-survey panel, plus the enterprise data behind CEHNN and the 87,070-employee, 753-firm data behind CATCICE if their authors share them.

Baselines

Naive reference. The organization's base turnover rate, and a logistic regression on current features. This is the floor any model must beat in calibration and discrimination.

Classical classifiers. The interpretable set compared by Guerranti and Dimitri (A Comparison of Machine Learning Approaches for Predicting Employee Attrition): logistic regression, classification trees, random forests and voting ensembles. Random forests are the realistic deployment standard in HR analytics.

Tabular transformer. The deep model from A transformer-based deep learning framework to predict employee attrition. It tests whether flexible feature interactions already capture whatever the two-channel structure adds.

Network models. CEHNN (Exploiting the Contagious Effect) and CATCICE. These are the strongest existing methods on panels with relational data. The single-tie prediction must be compared against them specifically, because they already encode manager and peer contagion.

Strongest simple statistical competitor. A Cox model with time-varying current stress, a cumulative-exposure covariate, and an exposure-by-intervention interaction.

Ablations. Two ablations isolate the key ingredient. One removes the permanent channel, leaving only reversible blocking, which amounts to a current-state model. The other replaces the mechanistic forms with a generic two-state hidden Markov model of the same size.

Fairness and evaluation. All models get the same train/test splits by time and the same features. Each gets an equal hyperparameter search budget of 50 configurations on a validation fold. Each is evaluated by time-dependent concordance, by integrated Brier score, and by intervention-response error on held-out intervention events.

Reported numbers. No reported result was found on Lacuna for discrimination metrics of any of these baselines on IBM HR, the CEHNN data, or the CATCICE data. The file reports dataset properties only: the 91% contagion share, the 15–19% turnover rates, and the 87,070-employee cohort. Every baseline number will therefore be produced in this study under the shared protocol.

What counts as beating them. The model wins if it matches the best baseline's concordance within its 95% confidence interval, and also reduces intervention-response error by a margin whose interval excludes zero. Prediction accuracy alone is not the target.

Why not just…?

Why not just add cumulative stress as a covariate in a Cox model? This is the real threat. The permanent channel is, to first order, a saturating function of cumulative exposure. If the Cox baseline with an exposure-by-intervention interaction matches our intervention-response error, the catalysis structure adds nothing except interpretability, and we will report that. The model differs in its predicted shape: saturating growth, partial reset after a regeneration event, and a recovery fraction bounded by $1-p_i$. The held-out intervention analysis below is designed to separate a linear interaction from that shape.

Why not just use a flexible sequence model, such as a recurrent network or generic hidden Markov model, on the employee histories? It can in principle learn any history dependence. In practice it faces the small-cohort overfitting already documented for tabular deep models (transformer attrition model). Nothing forces its latent states to mean "recoverable" versus "not recoverable," so it gives no lever an HR team can act on. The data-size sweep in the experiments will show where, if anywhere, the flexible model overtakes the structured one.

Why not just measure burnout directly and treat it as the mediator? The mediation studies above do roughly this (Fessehatsion, Lee et al.). A single burnout score at one time point cannot distinguish an employee who will rebound from one who will not. Two employees with equal burnout but different exposure histories are the decisive case. We will test whether the model separates their post-intervention outcomes when burnout alone does not.

Why not just use counterfactual explanations to pick interventions? FCEGAN generates minimal, realistic changes to mutable features that flip a quit prediction. It is a valuable tool, but it inherits the assumption of the classifier underneath it that the present state is all that matters. We will run FCEGAN on top of the best baseline. We will then check whether its predicted gains are overstated in proportion to the employee's estimated $p_i$.

Proposed Experiments

Synthetic identifiability. First, we simulate employee panels from the model with known $b$ and $p$. We vary panel length, survey frequency and turnover rate between 10% and 20%, and report how well the posterior recovers $p_i$ (rank correlation with the true values).

Catalyst ground truth. Second, we run the same estimator on poisoning and regeneration traces. Success means the inferred permanent fraction matches the fraction that does not recover after the sulfur feed is removed, and agrees with sulfate measured by surface spectroscopy. Failure stops the HR application.

Prediction on a real panel. Third, on the partner panel and any shared enterprise data, we compare all baselines on time-dependent concordance and integrated Brier score. We use temporal splits and subsample the training cohort from 500 to the full size, to find the regime where the structured model helps.

The decisive check: intervention response. We collect held-out intervention events such as schedule reductions, raises, manager changes and transfers. For each, every model predicts the change in quit hazard over the following six months, and we compare against observed outcomes, using matched comparison employees to adjust for selection into interventions. The hypothesis fails if the observed gain does not decline with estimated cumulative exposure, or if the Cox interaction model fits the decline equally well.

The single-tie test. Fourth, on data with manager links, we test whether quit hazard after a manager departure is step-like for employees with one strong tie and buffered for those with many ties. The test asks whether this pattern holds beyond what CEHNN's contagion terms already explain.

Static sanity check. Finally, we report static-model results on IBM HR. Because that dataset is synthetic, these results serve only to confirm the implementation.

References

  1. Progress in Catalytic Oxidation of Noble Metal-Based Carbon Monoxide: Oxidation Mechanism, Sulfur Resistance, and Modification
  2. Exploring PtSO4 and PdSO4 phases: an evolutionary algorithm based investigation
  3. Catalytic oxidation of NO over MnOx–CoOx/TiO2 in the presence of a low ratio of O3/NO: activity and mechanism
  4. Ordered Mesoporous MnAlOx Oxides Dominated by Calcination Temperature for the Selective Catalytic Reduction of NOx with NH3 at Low Temperature
  5. Activating low-temperature diesel oxidation by single-atom Pt on TiO2 nanowire array
  6. Examining the Relationship Between the Aspects of School Working Conditions and Teachers’ Intent to Stay as Mediated by Job Satisfaction
  7. Work design, employee well-being, and retention intention: A case study of China's young workforce
  8. Relationship Between Sleep‐Related Worry, Work Well‐Being, and Retention Intention Among Nurses: A Cross‐Sectional Multicenter Study
  9. Organizational Drivers of Turnover Intention in China’s Manufacturing Industry: Evidence of the Mediating Role of Employee Motivation
  10. Exploiting the Contagious Effect for Employee Turnover Prediction
  11. A transformer-based deep learning framework to predict employee attrition
  12. A Comparison of Machine Learning Approaches for Predicting Employee Attrition
  13. The Enhancement of CO Oxidation Performance and Stability in SO2 and H2S Environment on Pd-Au/FeOX/Al2O3 Catalysts
  14. Employee Turnover Prediction: A Cross-component Attention Transformer with Consideration of Competitor Influence and Contagious Effect
  15. The influence of communication climate, organizational identification, and burnout on real estate agents’ turnover intention
  16. Flexible Counterfactual Explanations with Generative Models