When doing machine learning, do we normally use several algorithms for comparison? For example, if the RMSE of SVM is 0.1, how do I come up with the conclusion that this model performed well? Just based on saying RMSE value is low, so the result is good? But if there is no comparison, how do I say it is low?
Or shall I include other algorithms e.g. random forest etc to do a comparison of the value? I intended to use only SVM regression, but now I am a bit stuck at the interpretation of the results.. Thank you in advance!
You have to compare performance of your algorithm with the previous state of the art techniques based on similar data sets and similar task (e.g., classification, clustering, tracking, detection etc).
SVM is applied in a large number of domains. Even though the research niche you want to tackle (the psychological part of tourists) has no previous use of SVM, you still can use several regression models/classifiers to compare between.
Based on my understanding, you are using the regression version of SVM. In this case, the coefficient of determination R-score could be useful. If it is near to one, that means the prediction is good. If less than 0, then it isn't. This measure is available in recent packages such as scikit-learn.
Dear Chung-En Yu,
It seems to me that you are using this data for the first time on your proposed method. In such case you do not need any comparison, however you need to add sufficient explanation/justification in discussion section for why is your method good? and how machine learning research will be benefited? i.e., to explicitly describe your contribution. Here your work may be considered as first state of the art on your data and task for your proposed method/algorithm. The next/future researchers will than compare their results with that of your results. ...
Sorry to differ, even if one's contribution is a unique method, it is still better that he/she compares different regression models to complete the picture (recommendation of the best regression model that can accompany his/her method, cannot be done without comparison).
General practice is to use a number of features and classifiers and then decide on the best one, which would hence justify its use for a specific application.
You cam measure the performance using precision, recall, F-score and Rand Index for class labelled data. These measures can be computed using TP(true positive), TN(true negative), FP(false positive), FN{false negative). [Supervised Learning]
For unlabelled data you can only use Rand Index. [Unsupervised Learning]
Cite
3rd Mar, 2020
Abbas Cheddad
Blekinge Institute of Technology
It is apparently regression not classification.
Cite
3rd Mar, 2020
Joanne Yu
Fachhochschule Salzburg
Thank you all for the response! I would take into account all your suggestions.
Abbas Cheddad I appreciate a lot with your comment. Indeed, I wanted to confirm if I should include more models. I would then add different regression models for comparison in this case. (Of course, will also look for relevant literature for justification)
Cite
1 Recommendation
4th Mar, 2020
Pinaki Mitra
Indian Institute of Technology Guwahati
For the regression we usually minimize the sum of squared of the distance from the regression hyperplane. Other metrics can be minimizing the maximum distance.
Cite
7th Mar, 2020
Mohammed Abdallah Bakr Mahmoud
Beijing Institute of Technology
Read the following paper https://arxiv.org/pdf/1902.05731.pdf
Cite
Can you help by adding an answer?
Answer
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms