Large-scale machine learning algorithms for big data

Publication Type:
Thesis
Issue Date:
2018
Metrics:
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf132.31 kB
Adobe PDF
02whole.pdf788.95 kB
Adobe PDF
Machine learning is a research area in artificial intelligence which aims to learn a model from data. On one hand, the target is to learn a model yielding superior performance. On the other hand, as the rapid increase of the size of the collected data, there emerges a demand for machine learning algorithms to deal with large-scale problems. Recent years have witnessed a sharp increase of the scale of the collected data. Taking recommender systems as an example, the Yahoo Music dataset includes more than 262 million ratings. In image classification, Imagenet contains more than 100 million images from the Internet. Such a large scale brings a great challenge to machine learning algorithms: how could the machine learning algorithms achieve satisfactory performance with less computational cost? In this dissertation, I mainly focus on several specific machine learning tasks and their scalability issues in either computation or storage aspects. Computational cost plays a crucial role in machine learning algorithms. For instance, iteration complexity is a commonly-used theoretical metric to evaluate how fast an optimization algorithm converges. An example is the full singular value decompositions (SVDs) in the nuclear norm minimization for low-rank matrix completion. Its computational complexity can be 𝑂(𝑛³) where 𝑛 is the size of the matrix. It would be computationally unfordable when n scales up. Memory cost is also a typical concern in machine learning. Recently deep neural networks have captured much attention and been successfully applied to a variety of applications. These deep models are known to be hungry for data, so training them usually requires a large number of training samples. When the entire training set cannot be loaded into the memory simultaneously, online (stochastic) learning can be applied. In such a memory-restricted scenario, both theoretical analysis and empirical investigation are expected. Targeting on the above two aspects in large-scale machine learning tasks, in this dissertation, I investigate a variety of machine learning tasks and analyze their specific characteristics. Specifically, I mainly focus on four tasks, i.e., matrix factorization for ordinal ratings, semi-supervised learning, active learning for image classification, online learning for imbalanced streaming data. For the first three tasks, I analyze the specific characteristics of the underlying problems and design new algorithm to optimize the objective. Theoretical verification such as computational complexity is provided. For the last task, I propose an online learning algorithm to deal with imbalanced problems under the strict memory constraint.
Please use this identifier to cite or link to this item: