I am trying to identify a method to formulate a hybrid data mining for regression analysis. I have trained different regression algorithms and have obtained set of RMSEs (Algorithms that I have used; Linear Regression, Ridge Regression, Decision Tree Regression, Random Forest Regression and Multi Layer Perceptron Regression). Considering RMSE, the Decision Tree Regression has scored the lowest. Now I don't have an idea on how to proceed for the further implementations.
Very much appreciating a valuable insight. Thank you in advance.
Tharika Weerakoon
I'm not sure I'm followed your question. But here what I would recommend you to do.
+ You can extend your experiments by comparing above methods based on multiple error metrics such as RMSE, MAE, MAPE, SMAPE, R2,...
+ Or you can implement more method in comparing such as a new neural network (functional-link neural network, radial basis function nets, recurrent neural network, LSTM, GRU or CNN and so much more...)
+ Or you can compare the convergence characteristic or the stability (run the model n times) among above models.
+ Or you can analyze the effect of parameters in your proposed models.
That's all I can think of right now. Hope It can help you.
Not exactly sure of your question either :-). So I am making some assumptions here. Ask yourself these questions:
1) Why would you use these different regression approaches? What makes them different, and, most importantly, how do these differences in the technique fit my data for worse or for better? This is the top-down question.
2) Apart from fit (or error) measures, can you understand how these models describe your data? What are they telling you (strongly linked to question 1). Is this a view you can comprehend, and does it make sense.
3) How many regressors are in each model/approach. Remember, the more degrees of freedom you add, the better you can make it fit the data. But, if you have a great deal of variables, the point is to try and reduce them to a dimensionality that is meaningful in two ways: from our point of understanding, and from the point of predictability.
4) On a more technical side, look at the model stability over the input space (for each approach). Create some novel data set (for example, sample for each variable within reasonable bounds of the variable distribution) and run it again. Do this for each variable independently, keeping the others as they were. In essence, this is a Monte-Carlo like approach. Look for the impact on the RMSE's (or other measures). This can tell you how 'robust' each model is against changes in the data.
Hope this helps.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining