I have studying the size of my training sets. I am wondering if there is an "ideal" size or rules that can be applied. I am thinking of a generative hyper-heuristics that aim at solving np-hard problems that require a lot of computational resources.
Dear Patricia Ryser,
The following data partitioning methods have been suggested in several literatures in the field of Machine learning/ Pattern recognition:
a). 70% of the entire Dataset for training (Training data)
b). 15% of the entire Dataset for validation (Validation data)
c). 15% of the entire Dataset for testing (Testing data)
Alternatively, you can partition the Noisy Dataset into two part:
a). 75% of the entire Dataset for training (Training data)
b). 25% of the entire Dataset for testing (Testing data)
I hope this information will be helpful for you
Good luck.
I usually use the following trade-offs: the Test set is 10 - 15% of the training set. Confirming the lot is 5 to 10 percent of the training set.
Normally 70% of the available data is allocated for training. The remaining 30% data are equally partitioned and referred to as validation and test data sets. Partitioning ratio is an important aspect but, apart from this one must ensure that the population statistics of these data sets are marginally different from that of the overall data. It should also be ensured that the training dataset should include all possible patterns used for defining the problem and should extend to edge of the modeling domain.
Vinay,
Suppose we are generating some algorithms, then we suppose the patterns are the one for problem to solve. Not the patterns of instructions that make the algorithms.
Am I correct?
Many thanks
Patricia
Dear Patricia Ryser,
we normalize all data and divide the data into two classes: training data (seventy of all data), and testing data (thirty of all data)
Article in this link (Article Modeling of groundwater level fluctuations using dendrochron...
) will guide you.
One more thing i want to ask, whether the training and test sets collection should be equal? or Training dataset should be more than test set?
Thank you very much for this answer. I really appreciate it.
The data I work with are solutions of the Traveling salesman problem and other NP-hard problems. Finding a tour can take up to 10 seconds with a short runs. Finding solutions for more demanding problems can double or triple this time. As a result, it can become infeasible to run 100 instances even on a cluster. The computations are highly intensive for the generative hyper-heuristics. Unlike neural network, this can take quite a bit of time ...
i have developed an new algorithm which solves this partition problem. Contact me.
Hi Mr. @Emmanuel Sakala, I would like to know how your algorithm works. Thanks in advance!
its a algorithm that compares the results of using a stratified 10-fold cross validation and a leave-one-out and decide the best model.
Thanks for your prompt answer!
Do you have any paper or code/application repository that I could have a look? I'm interested in using this cross-validation to help me achieve a better training of a Neural Network.
Thanks again!!
The 80/20 rule is governed by Pareto principle (https://en.wikipedia.org/wiki/Pareto_principle)
Ankush Bansal please paste the link to the original source. Right now your answer looks like a text book example of plagiarism.
https://stackoverflow.com/a/13623707
https://towardsdatascience.com/train-validation-and-test-sets-72cb40cba9e7 Maciek Konopka
I don't think there is anything 'ideal' ratio for splitting a dataset. It depends on the type of dataset. I would suggest to try different ratios i.e 80-20, 70-30, 65-35 etc and to pick the ratio that gives the best performance result.
I think this link will be useful
https://glassboxmedicine.com/2019/09/15/best-use-of-train-val-test-splits-with-tips-for-medical-data/
There is no fixed rule for separation training and testing data sets. Most of the researchers were used 70:30 ratio for separation data sets. It is also depends on data characters, data size etc. You can used 70:30; 80:20; 65:35; 60:40 etc. anything which suits of your data characters.
I've published two journal articles where I split my data in 7 groups: 5 for training, 1 for validation and 1 for generalization. In short: about 71,4% of data for training and 28,6% for validation and generalization. It worked very well for me!!!
For further interest, please refer to:
Article Hybrid Adaptive Computational Intelligence-based Multisensor...
Article Assessment of Noise Impact on Hybrid Adaptive Computational ...
So why not 90:10, Does it not suitable? What will it cause? If the number of my dataset is relatively small, can I choose 90:10 to train and test?
- Badges
- Science topic
- Similar topics
- Computer Science
- Computing