RCA
Elite Member
- Joined
- Jun 27, 2017
- Professional Status
- Certified General Appraiser
- State
- California
Here is a nice and easy to understand presentation on Random Forest by Killian Weinberger, a Professor at Cornell and expert on Deep Learning and Regression Trees.
Unlike MARS, which produces value models, Random Forests is considered a "Black Box" method in the context of machine learning. But Random Forests and MARS are similar in these respects:
1. Both methods are non-parametric.
2. Both are non-linear
3. Both can model interactions between features.
4. Both perform implicit feature selection, i.e. decide which features have significant importance in predicting the target variable (e.g. Sale Price)
5. Both are flexible in handling different kinds of data
Differences.
1. Random Forests builds a large number of decision trees based on bootstrap sampling and bagging.
2. MARS does not use bootstrap sampling and bagging as inherent components. Although, bootstrap sampling and bagging can be used with MARS. However, in the latter case, you wind up with many MARS models and predictions, where the predictions are aggregated into a single value that may be more reliable than the value produced by a single MARS model.
3. Conclusion: Random Forests is easier to setup and use than MARS. It is more robust to changes in its parameters. It could be used by a review appraiser to quickly check the approximate validity of the values returned by an appraiser who uses MARS. That is its best use. Also, the appraiser himself, could easily run Random Forests against his data as a check on his MARS model. But, again, it is a Black Box approach that does not provide an understandable value model to explain the "Why and Wherefore" of property price differences.
www.cs.cornell.edu
Unlike MARS, which produces value models, Random Forests is considered a "Black Box" method in the context of machine learning. But Random Forests and MARS are similar in these respects:
1. Both methods are non-parametric.
2. Both are non-linear
3. Both can model interactions between features.
4. Both perform implicit feature selection, i.e. decide which features have significant importance in predicting the target variable (e.g. Sale Price)
5. Both are flexible in handling different kinds of data
Differences.
1. Random Forests builds a large number of decision trees based on bootstrap sampling and bagging.
2. MARS does not use bootstrap sampling and bagging as inherent components. Although, bootstrap sampling and bagging can be used with MARS. However, in the latter case, you wind up with many MARS models and predictions, where the predictions are aggregated into a single value that may be more reliable than the value produced by a single MARS model.
3. Conclusion: Random Forests is easier to setup and use than MARS. It is more robust to changes in its parameters. It could be used by a review appraiser to quickly check the approximate validity of the values returned by an appraiser who uses MARS. That is its best use. Also, the appraiser himself, could easily run Random Forests against his data as a check on his MARS model. But, again, it is a Black Box approach that does not provide an understandable value model to explain the "Why and Wherefore" of property price differences.
Welcome | Kilian Q. Weinberger
www.cs.cornell.edu