Hi, I'm trying to do some IoT device classification, the dataset I'm using looks like this:
100 classes
5000~15000 samples for each class
10~15 features per sample
Can you suggest the most suitable model for this dataset? I'm looking to use RNN based models such as LSTMs but are there other algorithms more efficient(in terms of classification time) with about the same accuracy?
Thanks!
We need to look for a mathematical model and software tools that implement it, which do not impose any requirements on the data at all. you need to understand THAT any data requirements distort the picture. As a result, the more data requirements you make, the less you learn about reality, and the more you learn about your own requirements
I developed this model.
I would probably suggest you to start with simpler algorithms (including conventional ML classifiers, specially decision tree based or multi-class SVM or KNN models etc.), might be helpful. The "No free lunch theorem" holds here too- we don't really know what would work (or wouldn't). Though, your data seems like a time-series (probably streaming in real time), firstly start with simpler models, then move on to the specialised ones.
https://www.quora.com/What-are-the-best-machine-learning-models-for-extremely-small-datasets-20-instances
I would suggest checking the following paper. It has discussed the same issue.
"Efficient Voting Prediction for Pairwise Multilabel Classification".
http://www.ke.tu-darmstadt.de/publications/papers/neucom10.pdf
Is there a particular reason you've decided to use a RNN? Is there a temporal component to your data? Are the features continuous or discrete - a mix both? How are the data distributed?
I wouldn't say you're working with an 'extremely small dataset' but you still need to be careful about overfitting your model to data. I'd start with something simpler (multinomial logistic regression or random forest), and scale model complexity as required.
Also, here's an interesting read:
https://towardsdatascience.com/the-fall-of-rnn-lstm-2d1594c74ce0
- Badges
- Science topic