.png){kind=icon}
.png){kind=icon}
Return
Explainable domain adaptation for imbalanced occupancy estimation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jobe.2024.110613.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The recent emergence of IoT and sensor technology has led to the development of numerous tools for enhancing energy efficiency in smart homes. Accurate occupancy estimation is pivotal for sustainable energy management and reducing carbon emissions. Previous research used mostly traditional machine learning paradigms and assumed that the domain of the training and test set are the same. However, very often this is not the case. This discrepancy often results in biased models when deployed, leading to compromised model performance. The impracticality of creating a trained model for every environment stems from the scarcity of data, rendering data collection time-consuming, costly, and infeasible in certain scenarios. Traditional machine learning methods may yield suboptimal performance, failing to offer satisfactory solutions. Furthermore, occupancy datasets frequently exhibit imbalance, with models more often predicting absence due to the sporadic nature of individual presence, which negatively impacts model performance. Previous research has addressed these challenges using Domain Adaptation methodologies, leveraging deep learning to transfer knowledge from data- rich to data-poor domains. However, these black-box approaches often suffer from a lack of transparency, compromising model interpretability and accountability. This paper aims to integrate explainable AI with Domain Adaptation in the context of imbalanced Occupancy Estimation. We adapt and compare five domain adaptation methodologies based on Decision Trees - SER, STRUT, SER*, STRUT*, and GOSDT-DA - along with a random forest-based technique, SER-STRUT. We also introduce a new method: IMB-SER which modifies the tree structure of the domain adaptation model to mitigate the effects of imbalanced data. We compare the generated IF-ELSE rules and employ SHAP to interpret model behavior on test data, assessing similarity and robustness. Our empirical evaluation, using real occupancy estimation data, demonstrates accuracy levels of up to 93% for two-class and 85% for three-class scenarios, highlighting the potential of these approaches. The code implementations are accessible via our GitHub repository provided below: https://github.com/nmahamoodally/Transfer-Learningusing-Decision-Trees-for-Occupancy-estimation-with-Imbalanced-Dataset
Keywords:
Smart building
Occupancy estimation
Domain adaptation
Decision trees
Imbalanced data
Explainable AI
Journal
IF:
7.4
Papers:
1.6W
Citations:
6.6W