In many high-dimensional big data clustering methods, subspace learning is a commonly used technique. Traditional subspace-based methods project high-dimensional data into low-dimensional space to perform dimensionality reduction and clustering. Dimensionality reduction can reduce computational complexity, but it also leads to the loss of some key features. To deal with this issue, we propose a novel clustering algorithm based on a hyperdisk representation that provides a tighter approximation of sample regions. Specifically, the hyperdisk is defined as the intersection between the affine packet and a hypersphere, forming a disk-like region that offers a more compact representation of the class boundaries. This model achieves a balance between the loose approximation of the affine packet and the strict constraints of the convex hull, thereby enhancing the stability and reliability of the algorithm in low-sample high-dimensional classification scenarios compared to traditional hyperellipse models. For optimization, a standard quadratic programming algorithm is utilized to solve the proposed formulation. The performance of the algorithm is comprehensively evaluated from multiple perspectives, and its effectiveness is demonstrated through extensive experimental results.

In the realm of low-consumption computing technologies for big data analytics, target tracking often involves processing massive and continuous video streams. The discriminative correlation filter is a resource-efficient technique that learns filters through regression and generates response maps by convolving filters with feature representations, thus pinpointing targets accurately. However, most tracking models predominantly optimize temporal and spatial information, overlooking variations in the relative importance and correlations across feature channels. To address these issues within an energy-conscious computing framework, we introduce a correlation-based attention learning model. First, we calculate a similarity matrix for feature channels and use $K$-nearest neighbor to identify each channel’s neighborhood, allowing the filter to harness shared information among similar channels. Then, we construct a $p$-Laplacian matrix from the similarity data to guide the update of channel weights, capturing comprehensive feature relationships while enabling flexible adjustment of the order $p$ for different datasets. By doing so, our model more accurately redistributes each channel’s contribution, improving adaptability under diverse conditions while aligning with the efficiency constraints of big data analytics. Experiments on prominent public datasets confirm that our method not only delivers higher discriminative power and superior tracking performance compared to existing techniques but also upholds the principles of low-consumption computing, making it particularly suitable for large-scale, resource-aware scenarios.