PROJECT TITLE :

Representation Learning from Limited Educational Data with Crowdsourced Labels

ABSTRACT:

It has been demonstrated that representation learning plays a significant part in the unprecedented success of machine learning models in a wide variety of tasks, including machine translation, face recognition, and recommendation, amongst others. The vast majority of the currently available strategies for representation learning frequently call for an extensive number of labels that are free of noise. But in many real-world situations, the use of labels is extremely restricted for a variety of reasons, including financial restraints and concerns about individuals' right to privacy. When standard representation learning approaches are directly applied to small labeled data sets, it is easy to run into problems with over-fitting, which in turn leads to solutions that are less than optimal. Even worse, the limited labels in some fields, such as education, are typically annotated by multiple workers with varied areas of expertise. This results in noises and inconsistency in crowdsourcing settings. In this paper, we propose a novel framework with the goal of learning effective representations from limited data using crowdsourced labels. The framework can be found at the end of the paper. To be more specific, we design a grouping-based deep neural network to learn embeddings from a limited number of training samples and present a Bayesian confidence estimator to capture the inconsistency among crowdsourced labels. Both of these things are presented in this paper. In addition, in order to speed up the process of training, we develop a hard example selection procedure, which allows us to pick up training examples in an adaptive manner that the model incorrectly categorizes. Extensive experiments that were run on three different real-world data sets have shown that our framework is superior to other state-of-the-art baselines when it comes to learning representations from limited data using crowdsourced labels. These experiments were conducted on the data sets. In addition, we present a comprehensive analysis on each of the primary components of our proposed framework, and we also introduce the promising results that it achieved in our actual production, in order to ensure that the proposed framework is understood in its entirety.