Machine learning is increasingly applied to tackle classification challenges in various real-world contexts. Recently, there has been a growing interest in tensor-based modeling techniques within the machine learning community. In this work, we combine classification and tensor decomposition methods to reformulate the classification problem as a tensor completion task. Specifically, a tensor of scores of the samples is learned, where the sample values correspond to indices in the tensor, containing the scores for each class. Subsequently, when we encounter new data points, we can classify them by extracting the score values from the learned tensor at the indicated sample position, assigning them to the class with the highest score. Given that only a fraction of tensor entries can be obtained from a limited training set of samples, we utilize Tucker decomposition in conjunction with the hinge loss function to complete the score tensor, considering the discrete nature of the predicted class variable. Our experimental results across various real-world classification tasks reveal that this proposed tensor-based learning approach enhances classification performance, especially when dealing with a constrained number of training samples, outperforming state-of-the-art methods.
@inproceedings{aidini2023classification, title = {Few-Shot Classification Using Tensor Completion}, author = {Aidini, Anastasia and Tsagkatakis, Grigorios and Sidiropoulos, Nicholas D. and Tsakalides, Panagiotis}, booktitle={57th Asilomar Conference on Signals, Systems, and Computers}, year = {2023}, organization={IEEE}, abstract = {Machine learning is increasingly applied to tackle classification challenges in various real-world contexts. Recently, there has been a growing interest in tensor-based modeling techniques within the machine learning community. In this work, we combine classification and tensor decomposition methods to reformulate the classification problem as a tensor completion task. Specifically, a tensor of scores of the samples is learned, where the sample values correspond to indices in the tensor, containing the scores for each class. Subsequently, when we encounter new data points, we can classify them by extracting the score values from the learned tensor at the indicated sample position, assigning them to the class with the highest score. Given that only a fraction of tensor entries can be obtained from a limited training set of samples, we utilize Tucker decomposition in conjunction with the hinge loss function to complete the score tensor, considering the discrete nature of the predicted class variable. Our experimental results across various real-world classification tasks reveal that this proposed tensor-based learning approach enhances classification performance, especially when dealing with a constrained number of training samples, outperforming state-of-the-art methods.}, keywords = {Classification, Tensor Completion, Tucker Decomposition, Hinge-loss Function, Learning}, }