Semi-supervised Sequence Classification through Change Point Detection

Authors: Nauman Ahad, Mark A. Davenport6574-6581

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide extensive synthetic simulations and show that the learned representations are better than those learned through an autoencoder and obtain improved results on simulations and human activity recognition datasets.
Researcher Affiliation Academia Nauman Ahad, Mark A. Davenport School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA {nahad3, mdav}@gatech.edu
Pseudocode Yes Algorithm 1 SSL via change point detection
Open Source Code Yes Code: https://github.com/nahad3/semi sup cp
Open Datasets Yes HCI: Gesture Recognition (Forster, Roggen, and Troster 2009) and WISDM: Activity Recognition (Kwapisz, Weiss, and Moore 2010)
Dataset Splits Yes In all synthetic simulations, we split the data in a 70/30 ratio where we use the larger split for training and the smaller split as a test dataset. We further split the training data in a ratio of 10/60/30. We use the smallest of these splits to obtain labeled data, the largest as unlabeled data for the semi-supervised setting, and the last split for validation.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions types of networks and general libraries but does not provide specific software dependencies with version numbers.
Experiment Setup Yes The neural network consists of 6 convolutional layers (or 3 temporal blocks as defined by (Bai, Kolter, and Koltun 2018)) followed by 1 linear layer. We use a RELU activation function after every convolutional layer. We use a 2-layer feedforward neural network followed by a softmax function to obtain a distribution over the different classes. Our loss function is constructed applying a hinge loss (with margin parameter ρ) to this KL divergence...Here, PL and PU denote the sets of sub-sequence pairs in PS Y PD formed from the labeled and unlabeled data, respectively, and λr is a tuning parameter which controls the influence of the unsupervised part of the loss function. Here, λC is a tuning parameter, LCE is the cross entropy loss, and LNE is the negative entropy loss.