Open Collections

Investigations on deep learning techniques for property assessment

University of British Columbia

Property assessment plays a vital role in not only people's daily life but also the establishment of policies in North America. Therefore, the robust and accurate assessment is thus required. However, there are several challenges to prevent improving performance in property assessment. For example, the to-be-valued property and its neighbors can influence their values mutually, which is named peer-dependence. Besides, the property values also depend on their topography around the house. Moreover, various urban data in the city, i.e., criminal activities and point of interests, also reflects the property values. The accuracy of property assessment can be further improved if the process can fuse the urban data for assessment. Nowadays, deep learning models become the most powerful computational models in the subjects of both electrical engineering and computer science. Therefore, the thesis aims to investigate deep learning techniques to address three specific challenges in property assessment, which are considering peer-dependence considering topography and considering urban data for improving property assessment. In the first study, a recurrent neural network (RNN), peer-dependence valuation model (PDVM), is proposed to converts the peer-dependence based valuation into sequence prediction. Firstly, K-Nearest similar house sampling (KNSHS) algorithm generates sequences from the to-be-valued property and nearby houses. Then, an RNN estimates the property values from given sequences. The experimental results indicate that PDVM outperforms the other novel computational models being used for property assessment. In the second study, a convolutional neural network (CNN) based method, attention-based multi-modal fusion (AMMF), is proposed to estimate property values with house attributes and corresponding street maps which contain features of topography. The experimental results indicate the proposed method outperforms the contemporary approaches. In the third study, to investigate the ways of urban data fusion to property values, a novel deep learning framework, multi-source urban data fusion (MUDF), is proposed to estimate property values from metadata and street-view images of properties. The experimental results with the data collected from the city of Philadelphia demonstrate the accuracy of the proposed model.

Text

2019-01-16

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/68253

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/