🤖 AI

A time-series classification framework for individual-level absenteeism prediction under severe class imbalance

This paper proposes a time-series classification framework using an LSTM-FCN architecture and optimized loss functions to enable genuinely proactive, individual-level absenteeism prediction in highly imbalanced datasets, overcoming the limitations of existing methods that merely reproduce realized outcomes.

Original authors: Kwong Ho Li, Matthew Roughan, Wathsala Karunarathne

Published 2026-07-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kwong Ho Li, Matthew Roughan, Wathsala Karunarathne

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a busy hospital or a delivery company. Your biggest headache? Absenteeism. When staff don't show up, everything grinds to a halt. You need to know before they are gone who is likely to call in sick, so you can prepare.

This paper is about building a "crystal ball" for employee attendance, but with a very specific twist: it tries to predict the future based on the past, rather than just looking at what happened right now.

Here is the story of their research, broken down into simple concepts.

1. The Problem: Looking in the Rearview Mirror

Most current computer programs for predicting absences are like driving a car while only looking in the rearview mirror. They take data from "today" (like an employee's age or how far they live) and try to guess if they are absent today.

The authors say this is useless for planning. By the time you know the data for "today," the employee has already made the decision to stay home or come in. You can't fix a staffing shortage that already happened.

The Solution: They propose a Time Series Classification framework. Think of this as looking out the windshield. Instead of asking, "Is John absent right now?" they ask, "Based on John's attendance pattern over the last 40 days, is he likely to be absent in the next 5 days?" This allows managers to be proactive, not reactive.

2. The Data Hurdle: Building a "Digital Twin"

Real employee attendance records are private and sensitive (like medical records). You can't just download them from the internet.

To get around this, the researchers built a simulated dataset. Imagine they took a small, real dataset from a courier company in Brazil and used it as a "seed." They then used math to grow a massive forest of 1,000 "digital employees" who behave statistically exactly like the real ones. This allowed them to test their system without violating anyone's privacy.

3. The "Class Imbalance" Trap

Here is the tricky part. In almost any company, 97% of the time, people show up to work. Only about 3% of the time, they are absent.

In machine learning, this is called severe class imbalance. It's like trying to teach a dog to bark only when it sees a lion, but 99% of the time, you show it a cat. The dog will just learn to stay silent (predict "no lion") because it's right 99% of the time.

In their case, a computer model could achieve 97% accuracy simply by guessing "Everyone is present" every single day. But that's a disaster for a manager who needs to know who is actually missing. The goal isn't just "accuracy"; it's Specificity—the ability to correctly spot the rare "absent" days without crying wolf.

4. The Loss Function Showdown: The "Focusing" vs. The "Self-Correcting"

To fix the imbalance, the researchers tested two different "teachers" (mathematical formulas called loss functions) to train their AI.

Teacher A: Binary Focal Loss (BFL).
- The Metaphor: Imagine a teacher who tries to focus only on the students who are failing. However, this teacher needs a specific instruction manual (a parameter called $\alpha$ ) to know how much to focus.
- The Discovery: The standard instruction manual (used by most people) actually made things worse! It told the teacher to focus on the wrong group. The researchers had to rewrite the manual using a specific formula based on how rare the absences were. When they did this, the teacher got much better at spotting absent employees. But, it required the manager to know the exact math beforehand.
Teacher B: Geometric Mean (G-Mean) Loss.
- The Metaphor: This teacher has a self-correcting instinct. If the class starts ignoring the failing students, this teacher automatically shifts attention to them without needing a manual.
- The Discovery: This teacher performed just as well as the perfectly tuned "Focal Loss" teacher, but it didn't need any complex math adjustments. It just worked automatically.

The Winner: Both worked well, but G-Mean was the "plug-and-play" champion because it didn't require the user to calculate complex balancing numbers first.

5. The Best Architecture: The Hybrid Engine

They tested three different "engines" (AI models) to process the data:

LSTM: Good at remembering long stories (like a diary).
CNN: Good at spotting patterns in images (or in this case, patterns in time).
LSTM-FCN: A hybrid that combines both.

The Result: The Hybrid (LSTM-FCN) was the clear winner. It was like having a detective who can both read the whole diary and spot the specific handwriting patterns. It achieved about 80% balanced accuracy, meaning it was very good at predicting absences without making too many false alarms.

6. The Fine-Tuning: How Much History Do We Need?

They also tested how far back the AI should look:

Too short (5 days): The AI forgets the context.
Too long (160 days): The AI gets confused by old, irrelevant history.
The Sweet Spot: Looking back 40 to 80 days was the perfect window. It gave the AI enough history to see patterns (like "John gets sick every time he works a double shift") without getting bogged down in ancient history.

Summary of Findings

Don't look at the rearview mirror: Predicting future absences based on past sequences is the only way to be truly proactive.
The "Standard" settings fail: If you use the default settings for handling rare events in AI, you will fail to spot absences. You must adjust the math or use a self-correcting method.
The Hybrid Model wins: Combining different AI techniques (LSTM-FCN) works best for this specific problem.
The Magic Window: Looking back 40–80 days gives the best results.
No Magic Wand: They had to use fake data because real data is private, but they proved their method works on the simulation, which is calibrated to real-world stats.

In short, this paper provides a blueprint for building a system that can look at an employee's recent attendance history and say, "Hey, based on the pattern, this person is likely to be absent next week," allowing managers to fix the problem before it happens.

Technical Summary: A Time-Series Classification Framework for Individual-Level Absenteeism Prediction Under Severe Class Imbalance

Problem Definition

The paper addresses the operational challenge of predicting individual staff absenteeism in high-demand environments (e.g., healthcare, emergency services, logistics). Existing machine learning approaches suffer from two critical structural limitations:

Temporal Misalignment: Current regression and classification methods map features observed at time $t$ to labels at the same time $t$ . This results in models that reproduce already-realised outcomes rather than predicting future events, rendering them reactive rather than proactive.
Loss of Sequential Structure: Existing methods often aggregate individual behavior into population-level summaries, discarding the sequential patterns within an employee's attendance history (e.g., accumulating fatigue or recurring health episodes) that contain predictive information.

Furthermore, the problem is characterized by severe class imbalance. In the studied context, attendance (present) constitutes approximately 97.69% of records, while absence is the minority class (2.31%), yielding an imbalance ratio $\rho \approx 42.31$ . Standard loss functions bias models toward the majority class, leading to poor specificity (the ability to correctly identify absent employees), which is the critical failure mode for workforce planning.

Methodology

Data Simulation

Due to the lack of publicly available longitudinal individual-level attendance data, the authors constructed a reproducible simulated dataset. This dataset was calibrated to the distributional properties of the UCI "Absenteeism at Work" benchmark.

Expansion: Monthly aggregate data from 36 employees was expanded into daily binary sequences (1,027 days).
Synthesis: 964 additional synthetic employees were generated via distribution-preserving sampling, resulting in a total of 1,000 employees.
Formulation: The problem was reformulated as a Time Series Classification (TSC) task. A sliding window approach was used where the input sequence $X$ consists of $n$ consecutive past days of attendance, and the target label $y$ represents the absence status over the subsequent $m$ days (prediction horizon).

Model Architectures

Three deep learning architectures were evaluated:

LSTM: A recurrent network processing the sequence to capture temporal dependencies.
CNN (FCN): A fully convolutional network using 1D convolutions to extract hierarchical features.
LSTM-FCN: A hybrid architecture combining a convolutional branch (for local patterns) and a recurrent branch (for sequential dependencies), with representations concatenated before the output layer.

Loss Functions and Theoretical Analysis

The study rigorously analyzed two loss functions under severe imbalance:

Binary Focal Loss (BFL): Introduces a focusing parameter $\gamma$ $γ$ and a class-balancing factor $\alpha$ $α$ .
- Theoretical Finding: The authors derived that the initial gradient ratio between present and absent classes is $\rho \alpha / (1-\alpha)$ . They proved that the focusing parameter $\gamma$ has zero effect at initialization ( $p=0.5$ ).
- Calibration: To achieve gradient balance, $\alpha$ must be set to $\alpha^* = 1/(1+\rho) \approx 0.023$ . Standard parameterizations (e.g., $\alpha=0.75$ ) amplify majority-class dominance, causing model collapse.
Geometric Mean (G-Mean) Loss: Optimizes the geometric mean of True Positive Rate (TPR) and True Negative Rate (TNR).
- Theoretical Finding: G-Mean exhibits a self-correcting gradient structure. The ratio of total gradient contributions depends on $\widehat{TNR}/\widehat{TPR}$ , which is independent of the imbalance ratio $\rho$ . As the model collapses toward predicting "present" for everyone, the absent-class gradient naturally dominates, providing a restoring force without manual parameter tuning.

Experimental Setup

Metrics: Specificity (primary), Balanced Accuracy, Precision, Recall, and F1-score.
Hyperparameters: Systematic sweeps were conducted on batch sizes, window sizes, and loss function parameters.
Environment: Training on an NVIDIA RTX 4090 GPU using PyTorch.

Key Results

1. Loss Function Performance

Standard BFL Failure: The standard BFL configuration ( $\alpha=0.75, \gamma=2$ ) resulted in a specificity of only 0.124, as the model predicted "present" for almost all employees.
Calibrated BFL: When $\alpha$ was set to the theoretically derived $\alpha^* \approx 0.023$ and $\gamma=0$ , specificity improved to 0.813 with a balanced accuracy of 0.888.
G-Mean Performance: G-Mean achieved a specificity of 0.844 and balanced accuracy of 0.734 without any parameter calibration.
Conclusion: G-Mean is preferable when $\rho$ is unknown or when avoiding manual calibration is desired. Calibrated BFL ( $\alpha^*, \gamma=0$ ) is competitive but requires precise knowledge of the imbalance ratio.

2. Architecture and Hyperparameters

Architecture: The LSTM-FCN hybrid model demonstrated the most consistent precision and specificity across configurations.
Batch Size: Performance stabilized for batch sizes $\ge 64$ . Smaller batches failed to reliably include minority class samples, causing the G-Mean self-correction mechanism to collapse. A batch size of 1,024 yielded the highest specificity.
Window Size: For a 5-day prediction horizon, window sizes between 40 and 80 days provided the best trade-off between performance and stability. Larger windows (e.g., 160 days) led to declines in F1-score and balanced accuracy.
Background Information: Incorporating static employee background information (e.g., age, BMI) did not significantly improve performance compared to using attendance sequences alone.

3. Test Set Evaluation

On held-out test data, the LSTM-FCN model with G-Mean loss, a window size of 40–80 days, and a batch size of 1,024 achieved a balanced accuracy of approximately 80%. The model generalized well, indicating that the learned temporal patterns were not overfitted.

Significance and Contributions

The paper claims four principal contributions:

TSC Formulation: It reframes absenteeism prediction as a sequential knowledge problem, explicitly separating historical sequences from future labels to enable genuine proactive decision support, contrasting with existing reactive models.
Theoretical Analysis of Loss Functions: It provides a rigorous theoretical derivation of gradient dynamics under severe imbalance. It proves that standard BFL parameterizations fail when the positive class is the majority and derives the exact analytical value for $\alpha$ required for balance. It also characterizes the self-correcting nature of G-Mean loss.
Systematic Evaluation: It establishes practical design guidelines for imbalanced sequential prediction, identifying LSTM-FCN as the robust architecture and defining optimal ranges for batch sizes and window sizes.
Reproducible Simulation Framework: It addresses the data scarcity barrier by releasing a calibrated simulation methodology that generates synthetic longitudinal attendance data, providing a baseline for future research in the absence of public individual-level records.

The authors conclude that while the study is limited by the use of simulated data, the framework offers practical guidance for severely imbalanced binary classification in workforce management, healthcare, and fraud detection. Future work is directed toward validating these findings on real organizational data and extending the framework to multivariate sequences.