An effective differentiable clustering approach for human activity recognition

We introduce DIFFUSE, a novel differentiable unsupervised soft clustering approach ideally suited for datasets with complex behavioral patterns, such as those found in Human Activity Recognition (HAR). DIFFUSE employs a fully differentiable loss function compatible with gradient-based optimization, enabling efficient execution on GPU-accelerated environments and addressing key challenges, including dependency on numerous sensitive hyperparameters. Central to DIFFUSE is a parameterized function, typically instantiated by a neural network, that generates soft cluster assignments, facilitating efficient learning of cluster structures from batched data inputs. Its dual-loss function combines clustering fidelity and cohesion losses: the fidelity loss ensures the clustering solution accurately reflects the data’s structure, while the cohesion loss promotes regularization by encouraging similar data points to share similar cluster assignments. Furthermore, DIFFUSE supports the integration of raw data representations that, although not directly suitable for clustering, enhance the precision of cluster assignments by extracting complex patterns inherent in diverse human activities. Experimental results on various datasets demonstrate that DIFFUSE consistently outperforms existing methods by effectively capturing nuanced patterns inherent in complex datasets. These outcomes highlight DIFFUSE’s ability to deliver precise, adaptable, and robust clustering solutions, significantly advancing unsupervised learning in scenarios with initially unclear data structures.