EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
1998. Harnessing spatial granularity in boosted tree models for house price prediction
Invited abstract in session WA-31: Analytics for Decision Making, stream Analytics.
Wednesday, 8:30-10:00Room: 046 (building: 208)
Authors (first author is the speaker)
| 1. | Margot Geerts
|
| Decision Sciences and Information Management, KU Leuven | |
| 2. | Seppe vanden Broucke
|
| Department of Decision Sciences and Information Management, KU Leuven | |
| 3. | Jochen De Weerdt
|
| Decision Sciences and Information Management, KU Leuven |
Abstract
Spatial relations in data such as house prices are known to vary strongly depending on their location. Several modeling techniques have been proposed from the domain of spatial statistics such as geographically weighted regression. This technique addresses this problem by allowing to vary the parameters over space. Nevertheless, more recent machine learning methods such as gradient boosted trees perform better in terms of prediction and improve scalability towards large datasets significantly. Nevertheless, global tree-based models, trained over the complete spatial region of the dataset, present large local errors which indicate their flawed ability to model spatially varying relations in the data. A solution consisting in learning local models on spatial partitions of the dataset is not preferred due to the loss of information shared by the partitions. In this study, we confirm the presence of this phenomenon in a real-life house price dataset and explore pathways towards a unification of global and local models to address this problem.
Keywords
- Analytics and Data Science
- Machine Learning
Status: accepted
Back to the list of papers