FSL-LFMG: Few-shot Learning with Augmented Latent Features and Multitasking Generation for Enhancing Multiclass Classification on Tabular Data: Existing and New Concepts

Nur, Aviv

Etd

FSL-LFMG: Few-shot Learning with Augmented Latent Features and Multitasking Generation for Enhancing Multiclass Classification on Tabular Data: Existing and New Concepts

Public Deposited

In my Master's thesis, I propose advancing Prototypical Networks that employs augmented latent features (LF) by an autoencoder and multitasking generation (MG) by STUNT in the few-shot learning (FSL) mechanism. Specifically, the achieved contributions to this work are sixfold. First, I propose the FSL-LFMG framework for few-shot multiclass classification on tabular data. This framework incorporates sample-level data augmentation using autoencoders, task-level data augmentation via an enhanced STUNT framework, and Prototypical Networks to capture generalized knowledge. Second, I design the latent features learning and augmentation process that employs autoencoders to extract significant features, which are then used to enhance the quality and diversity of the training data. Third, I employ the enhanced STUNT Multitasking Generation framework that uses K-medoids instead of K-means to generate more accurate tasks. Fourth, I implement an advanced Prototypical Networks with Manhattan distance as a classifier effectively address the multiclass classification problem. Fifth, I conduct an extensive experimental study on four diverse domain datasets—Net Promoter Score segmentation, Dry Bean type, Wine type, and Forest Cover type—to prove that my FSL-LFMG approach on the multiclass classification outperforms the Tree Ensemble models and the One-vs-the-rest classifiers by 7.8% in 1-shot and 2.5% in 5-shot learning. Finally, I demonstrate the adaptation of the new concept task in the model obtained from the FSL-LFMG framework — from the NPS segmentation (the existing concept) and obtain a level of customer’s loyalty (the new concept) — to assess the power of generalization of this framework by significant results of the mean test accuracy in both 1-shot setting (83.95%) and 5-shot setting (103.52%).

Creator