Predicting
21st Century
Workplace Demands
Multi-Class Machine Learning Approach
for Employee Engagement
Understanding Workforce Analytics
From Work from Home (WFH) survey data extraction to business-ready insights
Business Context
A Return-to-Office (RTO) proposition for our stakeholders
Data Preparation and Exploratory Data Analysis
Feature Engineering
Variable selection, skip logic, and correlation mapping
Multi-Class Predictive Modelling
Target vector, dealing with class imbalance, and
model choice
Validation & Business Deployment
Performance metrics and Employee Archetype scoring
A Return-to-Office Proposition
Physical real estate is a variable resource, not a fixed asset.
Organisations must quantify worker compliance versus disruption before committing capital.
Stakeholder Concerns
- Executive Management: Optimise space allocation (e.g. decentralisation and flexible co-working spaces), reduce lease exposure, leverage rotational hybrid schedules
- HR Leadership: Identify flight-risk cohorts, design targeted retention structures
- Facilities & Finance: Align headcount forecasts with footprint planning, introducing sensor tracking for space utilisation
People enjoy the first day of no COVID restrictions at the central business district in Singapore on April 26, 2022. (REUTERS/Edgar Su)
Profiling, Cleaning & Munging
Dataset's primary challenges
- Scale:
36078 rows x 34 columns
- Heterogeneity:
dtypes, NaNs
- Wave misalignment
Wave 2 Isolation
Extract 2022 survey parameters to establish a clean baseline
Validation
Interrogate variable types, NaN distributions, and cross-wave identifier integrity
Dimensionality Check
Class count for 'return_office' target variable, and dependent variables, against sample size
Sifting through the data structure
Handling the Void: Deterministic Skip Logic
Systemic Missingness as Signal
- Survey skip patterns (or NaNs) encode employee behavioural information:
- No WFH experience = 92% RTO compliance
- WFH experience = 2.5% Quit, 23% WFH-seek
- For predictive modeling, to use only 'wfh_missing' (created flag) for its impact on RTO
Cross-Referencing
Map skip patterns against WFH variables such as 'ever_WFH' and 'n_work_home'
Re-Engineer Nulls
Encode systemic skips as binary feature signals rather than imputing or discarding
Validate Coverage
Confirm re-engineered features preserve integrity across three target classes (Quit, WFH-seek, Comply)
Feature Correlation Heatmap
Surfacing Possible Behaviourial Drivers
- Prior WFH experience leads to warmer views 'daysemployee_work_home' (ρ = -0.31), 'value_WFH_rawpercent25' (ρ = -0.17), 'WFHperceptions'
(ρ = -0.16)
- Age may have mild influence on RTO compliance
'Age' (ρ = +0.09) sits above family columns (ρ < 0.05)
- Commute duration
'commute_time' (ρ = -0.11)
- Graduate status
'graduate (ρ = -0.17)
Feature Ranking
Age
Education Laddder
Compliance vs WFH days
Geographical scan
Protecting the Minority Class
Compliant Return
82.5% of workforce
Majority
WFH Seekers
15.0%
Consider engagement
Flight Risk
1.9%
Retention conversation
Mitigation Strategy
- Class weighting
- Inverse-frequency penalties applied inside training pipeline (leakage-safe)
- Manual weighting on minority class to confirm 'balanced' is optimal
- Macro-F1 peaks at weight 10 (0.403)
The Multi-Class Predictive Modelling
Candidate Model Algorithms
- Dummy Classifier
Majority-class dummy as the floor
- Gaussian Naive Bayes
Fast probabilistic baseline (independence + normality assumptions)
- KNN (nearest 15)
Instance-based, captures local behavioural clusters
- Logistic Regression
Interpretable coefficients for stakeholder communication
Logistic Regression model performs "the best" with balanced accuracy of 0.557
For Quit class
6% Precision
(i.e. 94% false alarms)
54% Recall
For WFH-seek class
28% Precision,
51% Recall
For Comply class
94% Precision
62% Recall
From Model to Business Tiers
Moderate Risk
Baseline worker, no WFH experience
Higher Risk
Same worker, WFH-experienced
