If the three species have different characteristics, you should include samples from all species. You may combine the three groups and then pick randomly 70% for training and the remining 30% for testing.. Hope this is helpful..

Dear Anh Tram Nguyen,

In order to perform classification or clustering analysis task, you need to split your Clusters A, B, C into training and testing splits.

If your classifier/model is parametric, you may 10% to 20% for training (from each A, B & C,...) and the remaining for testing purposes via few sample trials.

If you using nonparametric methods, you may use 50% for training and the remaining for testing purposes. Here, also some people using up to 70% for training and the remaining for testing/validation or accuracy measurements of the prosed model/method/algorithm.

Hi Hamdi Al-Jamimi , I cannot add species C into the training data set because I want my model to predict 3 classes on unseen data, even I have species D, E or etc. in the future, I want to generalize my model

You are right .. you need to test the model using unseen data.. however, the data you used for training (A & B) might have characteristics that are not existing in the testing samples (C).. You can generalize the model by having random selection of samples for training, not by hiding a whole group that might have special features not included in training groups.. Think about it..

Hello

It is my understanding that you are building a neural network to classify an unknown dataset into three categories.

Can i instead suggest you first use an unsupervised learning approach, i.e. clustering. This will separate your uknown dataset into a number of clusters (in your case three), depending on the features. As this doesnt rely on the label, it is more robust than a classification effort. And you can also do it in a species to species way, so you get more "test" cases

Now , you dont know which cluster corresponds to which category. However, here you can utilise some knowledge on the subject to differentiate them, (i.e. by picking a really obvious example from each cluster) or use these separated datasets for classification. A way to classify would be, (and without knowing the implementation i cant offer better help) Cluster centroid relative distances + features + centroids + statistical properties such as std of clusters fed into a classification algorithm.

Hi Anh Tram Nguyen , it is general to face this issue for training perfectly fit dataset for making prediction. Please refer these links attached, hope it helps you to train your data model for getting better accuracy rate then before! 1)https://towardsdatascience.com/problems-in-machine-learning-models-check-your-data-first-f6c2c88c5ec2 2)https://www.google.com/amp/s/www.geeksforgeeks.org/underfitting-and-overfitting-in-machine-learning/amp/ 3)https://www.analyticsvidhya.com/blog/2015/12/improve-machine-learning-results/ 4)https://www.researchgate.net/post/What_could_be_causing_low_classification_results

Hi Anh Tram Nguyen ,

Firstly, you have to keep in mind that a machine learning model does a matching between the input to predict an output. Therefore, it can't predict something it was not trained for. So if you want it to predict any class you have to provide that class first. Moreover, training ML models is continuous and needs to be updated as more data mean better performance (generally). And when you want to add new species (classes) you have to "inform" your model and give enough data and retrain it so it can distinguish between them.

And about generalization, it is not predicting automatically new unseen classes. Generalization means that your model can predict correctly your classes even if they come from slightly different distributions (using an unreal example: imagine your species A, B, and also C are living in different environments such as mountains and forests. Then a good model should generalize well and predict correctly the A, B, and C classes when they come from a different "unseen" environment, say a farm)

So to sum it up, you should use all your classes, but split your data points into different sets, training, testing or you can also use cross-validation as well. Then when you want your model to recognize new classes you have to train it again with all the classes and repeat the split / cross-validation. I hope this helps

Hi Anh Tram Nguyen,

It is not clear which classifier you are using and the proportion of observations used in the training sample. So. try some good classifier and take the training versus ratio as 90:10. Hope you will get the desired results.