Decoding Learner Trajectories: a Hybrid Deep Ensemble for Dynamic Performance Assessment

Abstract

The proper profiling of the performance of information technology (IT) staff in adaptive e-learning contexts represents a critical need to the facilitated development of individual skills. Nevertheless, this project is full of problems due to the heterogeneity of the learning behaviour, the high dimensionality of interaction data and complex spatiotemporal dependencies of the activity of learners. In that regard, we propose an Ensemble Spatiotemporal Deep Learning (ESTDL) model that is able to conduct a strong multi-classification of performance levels (high, medium, low). Our architecture has a four-stage pipeline, which includes data acquisition, a vigorous preprocessing and a duplicate elimination step, attribute selection which is performed using an Information Gain criterion to isolate the most discriminative features, and a heterogeneous ensemble classifier. The ensemble is a synergetic combination of Xception to extract hierarchical spatial features, ShuffleNet to introduce computational efficiency on lightweight processing, Simple Recurrent Unit (SRU) to model temporal sequences, and Hybrid ShuffleSRU was designed to implement joint spatiotemporal learning. These constituent models are then synthesised by a majority voting scheme to provide a final classification which is robust and reliable. The results of empirical analysis of a real-world dataset of IT professional learning sessions provide the classification accuracy of 97.3%, F1-score of 0.974, and Matthews Correlation Coefficient of 0.981, significantly exceeding the performance of individual baseline models and traditional machine-learning applications. The results obtained in this research paper confirm that the combined use of feature selection, a variety of architectural inductive biases, and ensemble decision-making may lead to an incredibly accurate, scalable, and interpretable performance analytics model. This contribution will provide a methodologically sound solution to fit the adaptive learning systems with direct implications concerning the design of individualised training pathways and competency management of the working force in the IT industry.