 Joint multi-task modeling of multiple attributes for the enrichment of real estate dataMiroslav Despotovic, Stumpe Eric, Matthias Zeppelzauer and Roberto Labadie-Tamayo ERES from European Real Estate Society (ERES) Abstract: In recent years, there has been an increasing focus on addressing the problem of feature extraction and/or imputation of missing real estate data for training Automated Valuation Models (AVM), where Multi-Task Learning can be used as a convenient solution. Multi-task learning aims to learn multiple different tasks simultaneously while maximizing performance on certain or all the tasks. By learning several tasks in parallel, the information learned for one task can be used to better generalize for another task. The related information from other tasks thereby serves as a domain-specific inductive bias. This inductive bias is the pre-requisite for the classifier's or regressor's ability to generalize.Thereby, a more general representation should be learned that takes relationships between the different tasks into account and that enables transferring knowledge learned for one task to another (shared concepts). As a result, on task can build upon the information leaned for another task.To cope with the feature extraction and missing information in the data we leverage the framework of multi-task learning. Thereby, we aim at exploiting mutual information between similar real estate attributes for more accurate prediction of the extracted features. We aim at stacking and connecting task-specific layers on top of the multimodal representation. We jointly fine-tune the individual models with the information available as ground truth and connect the individual predictors to enable the sharing of concepts between real estate attributes. The main challenge is the different nature of the attributes that shall be modeled. Some attributes may be categorical (e.g. type of floor), others binary (e.g. balcony vs. no balcony) and others continuous (e.g. living area). Thus, different target functions must be optimized that require for different cost and regularization functions. This makes the learning task more difficult than typical multi-label learning tasks where all target functions are of the same type.The results show improved robustness of prediction by jointly modeling multiple attributes (including real estate properties' unstructured description text, images, numerical and categorical meta data) via multi-task learning, i.e. by leveraging information learned for one attribute to predict another one. We were able to achieve a particularly good prediction accuracy rate for external building characteristics such as year of construction, building condition and even heating demand.As real estate related attributes can be related (e.g. bulding condition is very likely to correlate with heating demand or building size is very likely to correlate with the number of floors) the joint modeling of multiple attributes in a multi-task fashion is a promising direction. We believe that the present study provides new insights into the use of alternative data and state-of-the-art analysis techniques for real estate analysis and provides a promising basis for further research in this area. Keywords: Automated Valuation Models; Data Enrichment; Multi-Task Machine Learning (search for similar items in EconPapers) Downloads: (external link) Related works: Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text Persistent link: https://EconPapers.repec.org/RePEc:arz:wpaper:eres2024-023 Access Statistics for this paper More papers in ERES from European Real Estate Society (ERES) Contact information at EDIRC. |
Is your work missing from RePEc? Here is how to contribute.
Questions or problems? Check the EconPapers FAQ or send mail to .
EconPapers is hosted by the
School of Business at Örebro University.