Functional Data Analysis for Prediction of Employee Presence

Abstract

As smart workplaces become more commonplace, predicting employee attendance is important in order to use space and resources efficiently. This paper demonstrates an intelligent forecasting technique through a dataset of employees' presence incorporating temporal and contextual variables. The data is cleaned and nulls and duplicates are removed, followed by exploratory analysis and visualization of correlation to identify behavioural trends. The prediction approaches used: AdaBoost, SARIMAX, FTSA, LSTM, DFMM, RobFTS, CNN and the combination of XGBoost, Gradient Boosting and KNN – Voting Regressor. Functional Data Analysis captures temporal dependence and dynamic occupancy patterns. The experimental validation shows that the DFMM model is very accurate, with a MAPE of 4,68% and a R² value of 91,3%. The Voting Regressor has an R² of 93.1% to produce better generalisation. We use approaches of Explainable Artificial Intelligence such as LIME and SHAP to understand the predictions and find influential elements. To implement the system in practice, the system is built using the light weight framework of flask which offers a web interface to log users in and out securely through the use of a SQLite database, user input from a file, backend processing, and user interactive visualisation. Submitting the test data to the system, the user defines the prediction horizons and will be provided with the weekly and daily forecasts of the presence of employees (e.g. four weeks, eight weeks) to aid decision making in the future planning of working places and optimal use of resources.