Tag based Bayesian latent class models for movies : economic theory reaches out to big data science

Amez, Lucy (2017) Tag based Bayesian latent class models for movies : economic theory reaches out to big data science. Doctoral thesis, London Metropolitan University.

Abstract

For the past 50 years, cultural economics has developed as an independent research specialism. At its core are the creative industries and the peculiar economics associated with them, central to which is a tension that arises from the notion that creative goods need to be experienced before an assessment can be made about the utility they deliver to the consumer. In this they differ from the standard private good that forms the basis of demand theory in economic textbooks, in which utility is known ex ante. Furthermore, creative goods are typically complex in composition and subject to heterogeneous and shifting consumer preferences. In response to this, models of linear optimization, rational addiction and Bayesian learning have been applied to better understand consumer decision- making, belief formation and revision. While valuable, these approaches do not lend themselves to forming verifiable hypothesis for the critical reason that they by-pass an essential aspect of creative products: namely, that of novelty. In contrast, computer sciences, and more specifically recommender theory, embrace creative products as a study object. Being items of online transactions, users of creative products share opinions on a massive scale and in doing so generate a flow of data driven research. Not limited by the multiple assumptions made in economic theory, data analysts deal with this type of commodity in a less constrained way, incorporating the variety of item characteristics, as well as their co-use by agents. They apply statistical techniques supporting big data, such as clustering, latent class analysis or singular value decomposition.

This thesis is drawn from both disciplines, comparing models, methods and data sets. Based upon movie consumption, the work contrasts bottom-up versus top-down approaches, individual versus collective data, distance measures versus the utility-based comparisons. Rooted in Bayesian latent class models, a synthesis is formed, supported by the random utility theory and recommender algorithm methods. The Bayesian approach makes explicit the experience good nature of creative goods by formulating the prior uncertainty of users towards both movie features and preferences. The latent class method, thus, infers the heterogeneous aspect of preferences, while its dynamic variant- the latent Markov model - gets around one of the main paradoxes in studying creative products: how to analyse taste dynamics when confronted with a good that is novel at each decision point. Generated by mainly movie-user-rating and movie-user-tag triplets, collected from the Movielens recommender system and made available as open data for research by the GroupLens research team, this study of preference patterns formation for creative goods is drawn from individual level data.

Documents
1263:6862
[img]
Preview
LucyAmez_PhD_Thesis_Sept2017.pdf - Published Version

Download (5MB) | Preview
Details
Record
Statistics

Downloads

Downloads per month over past year



Downloads each year

View Item View Item