Information Retrieval with Dimensionality Reduction using Deep Belief Networks

Slot, V. R. (2016) Information Retrieval with Dimensionality Reduction using Deep Belief Networks. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text thesis_vincent_slot_04-2016.pdf - Published Version Download (1MB) | Preview
	Text Toestemming.pdf - Other Restricted to Backend only Download (593kB)

Abstract

The most common method to achieve dimensionality reduction of data would be to perform matrix decompositions, as is the case in Latent Semantic Analysis (LSA) and Principal Component Analysis (PCA). However, dimensionality reduction can also be achieved by layered Restricted Boltzmann Machines (RBM), called Deep Belief Networks (DBN). A primary application of LSA is information retrieval (IR), in this context often referred to as Latent Semantic Indexing (LSI). Dimensionality reduction in IR is used to interpret data on an abstract level to represent the data as patterns in more informative, fewer dimensions. Through these patterns meaning can be attributed to words and documents. This research investigates whether or not neural networks can compete with these ‘traditional’ methods of dimensionality reduction in IR in the setting of a small database. Experiments are conducted with a search engine constructed for the purpose of comparison of the different algorithms. It compares vector space search, LSI search, DBN search and different variations of DBN search. These variations are based on so-called phantom document querying, where queries are being replaced by the entire text body of the top search results of a different search engine. For each search algorithm different parameter configurations are being researched and their optimal implementations are compared with each other. The results show that the DBN shows slightly inferior results to the traditional methods in the application of deep learning in a relatively small database.

Item Type:	Thesis (Master's Thesis / Essay)
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	15 Feb 2018 08:11
Last Modified:	15 Feb 2018 08:11
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/13801

Actions (login required)

View Item