Since I am using the Random Forest model in R to predict LST (land surface temperature) from actual LST, I realized that model performance is better in lower Average LSTs, thus I wanted to extract the probable effects behind this phenomenon. one idea that I come up with is spread of the dataset. In other words, I thought that a lower standard deviation in lower surface temperatures would cause a better performance in the RF model since the probable violation of variables would be reduced. However, I am trying to confirm such kind of claim within the literature. I really wanted to know how the spread of data would affect the final performance of a machine learning model. For example, lower STDEV leads to better performance? if so, how can I find relative literature for this?
thanks for comment
sir the performance of the model would increase by taking the following parameters into the consideration
1.try to take the balanced dataset.ie having equal positives and negatives samples.
2.tune all the parameters of the RF classifier before training the model.
Low spread in data is always good for the performance of any learning algorithm/method. In case of your example, it is possible that it is less likey to have a "high" LST and when that happens, values can vary highly. That will result with less accurate predictions with RF (note that there are no probability assumptions in RF thus we can't say there is a violation).
One possible explanation is that maybe, in high LST, there is/are some excluded variable(s) that influence temperatures .
Another possibility is that as Mushtaq Ahmad Wani mentioned, you have an imbalanced set of response variables (LSTs). But I do not know if you use LSTs as continuous variables or class variables (e.g. 0s or 1s). In either case, you can create several sub-datasets by equalizing low and high LST observations and check Rf performance over these different datasets by using standard statistical metrics. If you don't have enough observations, you can use Bootstrapping resampling method to create these datasets.
Lastly, you can change your performance metric. Instead of using accuracy only, precision, recall or F1-score can be checked too. That will give you a better idea about the real performance of your RF model. You can use these metrics to tune your RF model parameters also.
- Badges
- Science topic