Machine Learning-Based Prediction System for Optimized Faculty Exam Duty Allocation

Machine Learning-Based Prediction System for Optimized Faculty Exam Duty Allocation

Sharath Naik, Gururaja S

Department of Computer Applications

St Joseph Engineering College, Mangalore, Karnataka, India

Abstract – Invigilation duties for faculty members are often assigned manually during exams in many colleges, which can lead to issues like uneven workload, scheduling conflicts, or inefficiencies. To improve this process, this study suggests a data-driven approach to help in distributing exam duties more fairly and efficiently. By considering past records such as previous duty assignments, teaching experience, subject expertise, and availability on specific dates, we developed a system that uses basic classification methods to assist in scheduling. Using a sample dataset that reflects real college environments, models like Random Forest and Decision Tree were applied to understand key factors that influence duty allocation. This method not only improves the overall scheduling process but also ensures transparency and fairness in managing faculty responsibilities during examinations.

KeywordsExam Duty Allocation, Faculty Scheduling, Machine Learning, Random Forest, Decision Tree, Prediction Model, Educational Optimization

1. INTRODUCTION

   In schools and colleges, managing exam duties for faculty members is often a repetitive and complex task. When done manually, it can lead to issues like uneven distribution of work, favouritism, and scheduling conflicts. As institutions grow and exam schedules become more complex, there is a growing need for a smarter, data-based approach to handle duty assignments. This study introduces a system that uses past recordssuch as previous duty assignments, subject expertise, availability, and leave historyto help make better and fairer decisions. Using straightforward predictive methods like Decision Tree and Random Forest, the system can suggest suitable faculty members for invigilation duties. A custom dataset modelled after real college settings was used to train and test this system. The goal is to reduce the burden of manual scheduling, promote fairness, and make the whole process more efficient and transparent. The model attained high predictive accuracy in assigning duties, and Preference Level and Experience were the most predictive features. The suggested solution can greatly lower administrative effort, enhance fairness, and act as a scalable base for future smart examination administration systems.

2. RELATED WORK

   The scheduling and resource allocation challenge has been solved in many areas of application with machine learning and optimization methods. For the purpose of education exam management, different techniques have been suggested to maximize scheduling efficiency and fairness.

   Hanafi et al. [1] proposed a Mixed Integer Linear Programming (MILP) model to resolve the exam invigilation assignment problem at University Pertahanan Nasional Malaysia. They considered staff availability, room capacity, and institutional limitations and achieved optimal assignment via mathematical optimization. Their model, though, did not have prediction capabilities on the basis of historical data, which our ML-based model integrates.

   Ben Said and Mouhoub [2] developed a hybrid ML + Constraint Programming approach to scheduling nurses, illustrating the potential of combining data-driven prediction with rule-based reasoning. Though their application was healthcare, their approach is a close match to the exam duty allocation problem. Their model accurately predicted shift appropriateness and imposed operational constraints, just like this paper does within the academic environment.

   Sun et al. [3] improved constraint programming efficiency by employing supervised learning to inform variable ordering in scheduling problems. Their learning-augmented constraint solving methodology showed enhanced solver performance in job shop schedulinglessons that can be applied to educational duty planning, where workload fairness and constraint are essential.

   Our work extends these foundations with the introduction of a fully ML-based, interpretable model that incorporates real- world motivated faculty features. In contrast to purely optimization-based solutions, our system learns from past trends, enabling it to keep pace with institution dynamics and tastes with little human action.

3. PROPOSED SYSTEM

   1. System Overview

      The proposed system uses machine learning to forecast which faculty members would be best suited for duty assignments during exams. It takes into account factors such as availability, prior duty number, teaching experience, leave status, and expressed preferences. It forecasts whether a faculty member should be assigned to duty on a particular exam day based on these.

      This method decreases workload by hand, evades human prejudice, and provides fair allocation of invigilation tasks. The system can be implemented on any existing academic ERP or examination management systems for real-time scheduling.

   2. Architecture Diagram

4. MACHINE LEARNING MODEL IMPLEMENTATION

   Fig.1: System architecture of the proposed ML-based faculty duty allocation model.

   1. Feature Descriptions

   Feature	Description
   Experience_Year s	Total years of teaching experience
   Specialization	The faculty member's field of expertise
   Availability_Date	The proposed exam duty date
   Previous_Duties	Number of prior exam duties completed in the session
   Is_On_Leave	Indicates if the faculty is on leave
   Preference_Level	Faculty's stated preference for the time/date
   Duty_Assigned	Target variable (Yes/No) for supervised learning

   Sample Rows from the Synthetic Faculty Dataset

   1. Dataset

      There were 100 faculty records in the simulated dataset, which included both numerical and categorical features. Whether or not a duty should be assigned was the binary target label.

   2. Data Preprocessing

      • To convert categorical fields like Specialization, Is_On_Leave, and Preference_Level, label encoding was used..

      • Irrelevant columns like Faculty_ID and Name were excluded from training.

      • Data was split into 80% training and 20% testing using train_test_split().

   3. Algorithm and Code

      Two models were trained:

      1. Decision Tree Classifier simple and interpretable

      2. Random Forest Classifier Using an ensemble model to improve generalization

      The model was trained using:

      The accuracy and performance were evaluated using:

      • accuracy_score()

      • classification_report()

      • Feature importances via.feature_importances_

   4. Results and Visualization

      Classification metrics and scores for accuracy on a test dataset of 20 records were used to evaluate the trained models' performance. With an overall accuracy of 90% as opposed to 75% for the Decision Tree model, the ensemble-based Random-Forest classification performed better than the Decision Tree model. This demonstrates the benefit of group learning in this situation.

      Here is a summary of the Random Forest model's classification report:

      Class 0 – Precison: 1.00 | Recall: 0.82 | F1-score: 0.90

      Class 1 – Precision: 0.82 | Recall: 1.00 | F1-score: 0.90

      Overall Accuracy: 0.90

      Random Forest Accuracy: 0.9

      Decision Tree Accuracy: 0.75

      Classification Report (Random Forest):

      	precision	recal l	f1- score	support
      0	1.00	0.82	0.90	11
      1	0.82	1.00	0.90	9
      accuracy			0.90	20
      macro avg	0.91	0.91	0.90	20
      Weighted avg	0.92	0.90	0.90	20

      These outcomes highlight the model's strong predictive ability, particularly in maintaining balanced performance between both assigned and non-assigned duty classifications.

      Feature Importance Analysis

      The feature importance chart (Fig. 2) displays the top five variables that influence the Random Forest model's predictions

      • Is_On_Leave Highest influence; ensures that unavailable faculty are excluded

      • Experience_Years Preference given to more experienced staff

      • Previous_Duties Balances faculty workload

      • Availability_Date Ensures correct mapping to exam schedules

      • Specialization Helps in subject-aligned duty allocation

        Fig. 2: Top Feature Importances in Exam Duty Prediction using Random Forest

        The Random Forest classifier's prediction performance across two classes is shown in the confusion matrix in Fig 3:

        Table: Confusion Matrix Random Forest Classifier

        Actual / Predicted	No Duty	Duty Assigned
        No Duty	9	2
        Duty Assigned	0	9

      • True Positives (TP): 9 correctly assigned duties

      • True Negatives (TN): 9 correctly identified as not assigned

      • False Positives (FP): 2 mistakenly assigned

      • False Negatives (FN): 0 duties missed

      The classifier's ability to correctly predict both outcomes with little misclassification is reinforced by the matrix.

      Fig. 3. Confusion Matrix of Random Forest classifier showing classification performance on the test dataset.

5. RESULT ANALYSIS

   A test dataset with 20 samples was used to assess the with a 90% accuracy rate on the test set, the Random Forest approach demonstrated a notable improvement over earlier iterations. With a 75% accuracy rate, the Decision Tree classification likewise did fairly well. With precision, recall, and F1-score of roughly 0.90, the Random Forest model showed consistent performance in both classes.

   Feature Insights:

   • Experience_Years emerged as the most dominant predictor, suggesting that faculty with greater experience are more likely to be assigned duties.

   • Availability_Date and Previous_Duties significantly influenced predictions, reflecting practical considerations such as availability and workload balancing.

   • Preference_Level had the least impact, but still contributed to fine-tuning decisions.

   • Customizing duty assignments according to subject relevance was made easier by Specialization.

     Interpretation:

     The model can replicate human reasoning if it comes to task delegation. It reflects institutional fairness by accounting for availability, workload history, and experience.

     Limitations:

   • The dataset used was synthetically generated and lacks the complexity and variability of real institutional data.

   • The current model may not generalize well to real- world scenarios without training on actual faculty records.

   • Label Encoding was used for categorical variables, which may oversimplify relationships between features.

   • No hyperparameter tuning or cross-validation was performed, which could further optimize performance.

   • The current version does not account for faculty preferences over multiple exam dates or long-term fatigue.

6. CONCLUSION AND FUTURE WORK

   This paper presents a practical system to help colleges and academic institutions assign exam duties to faculty members more fairly and efficiently. Instead of relying on manual planning, which can often be inconsistent or biased, the system uses past records such as a teacher's availability, previous workloads, teaching experience, and personal preferences to guide the assignment process. Using basic prediction techniques like Decision Tree and Random Forest, the system was able to suggest duty allocations with reasonable accuracyabout 90% with Random Forest and 75% with Decision Tree. These results show that the system can effectively recognize patterns that match common scheduling practices. The analysis also highlighted which factors, such as faculty leave records, experience level, and previous duty counts, played a major role in the decisions. By automating what has traditionally been a time-consuming and subjective task, this system aims to reduce the administrative load and ensure duties are shared more equally among staff..

   Future Work:

   • Web Integration: For a smooth deployment, integrate the prediction models into a full-stack web- based app with a real-time backend like Node.js and MongoDB.

   • Dataset Expansion: Gather and incorporate actual faculty information from institutional sources to enhance model performance and generalizability.

   Usability Considerations

   The proposed system can be easily added to a college's existing ERP (Educational Resource Planning) platform for real-time use. By connecting the model to the backend as a small service, institutions can make the exam duty assignment process smoother and more efficient. This integration allows for automatic updates based on teacher availability, exam schedule changes, and leave records, reducing the need for manual work. It makes the scheduling process more flexible and responsive, helping staff manage duties with less hassle and greater accuracy in real academic settings.

   arXiv:2409.07547, 2024. [Online]. Available: https://arxiv.org/pdf/2409.07547

   [3] Y. Sun, S. Nguyen, D. Thiruvady, X. Li, A. Ernst, and U. Aickelin, "Enhancing Constraint Programming via Supervised Learning for Job Shop Scheduling," arXiv preprint arXiv:2211.14492, 2022. [Online]. Available: https://arxiv.org/pdf/2211.14492

7. REFERENCES

1. M. A. I. Hanafi, S. A. Syed Ali, R. M. Jusoh, and N. A. Rahman, "A Mixed Integer Linear Programming for Exam-Invigilator Assignment Problem: A Case Study at Universiti Pertahanan Nasional Malaysia," Appl. Sci., vol. 14, no. 23, pp. 117, 2024. [Online]. Available: https://www.researchgate.net/publication/379444212

2. A. Ben Said and M. Mouhoub, "Machine Learning and Constraint Programming for Efficient Healthcare Scheduling," arXiv preprint