I am currently working on rare event prediction, which I have never done before (I used to work with simple prediction problem), and I looked up on this article about using LSTM for time series rare event classification.
It was very exciting to read since I think my case is a little similar: to predict calving time of cow, with historical activity feature, where calving is only happen once in the end of data collection.
However... aside from my data is way smaller than the example, I got too much confusion using LSTM. The good point about LSTM I am looking forward to is the "look back" feature that can let you decide for each output how many input in previous time you look back to.
My question will be: is there any "easier" or more simple machine learning method that works like LSTM for time series classification?
I tried to use simple ML such as decision tree, random forest, but I don't think it represents the problem well (many historical data into one output).
Dear Niken Prasasti Martono ,
Have your try Multi-layer Perceptron, Functional-link Neural Network, Radial Basis Function Networks, Extreme Learning Machine.
https://en.wikipedia.org/wiki/Multilayer_perceptron
https://en.wikipedia.org/wiki/Radial_basis_function_network
Conference Paper A Resource Usage Prediction System Using Functional-Link and...
https://en.wikipedia.org/wiki/Extreme_learning_machine
Hi Niken.
Our friend has done a research to solve time series prediction problem. Their method is able to look back to the previous sample which considered relevant to predict current sample. Unlike LSTM, it is easier to train as it does not have so many parameters. You can find their paper here Article Learning Elastic Memory Online for Fast Time Series Forecasting
.Apart from that, you can also try to use the following method https://arxiv.org/pdf/1803.01271.pdf . Basically, they tries to use a simple convolutional architecture to perform sequence modeling task. They claim that the result is better than LSTM.
Hopefully this can help.
Your goal should be to derive very high quality features. This allows you to reach your goal with simpler models and less data.
Furthermore LSTMs are not the final solution. See about that:
https://towardsdatascience.com/the-fall-of-rnn-lstm-2d1594c74ce0
https://bair.berkeley.edu/blog/2018/08/06/recurrent/
- Badges
- Science topic