Hello everyone;
I'm preparing my master thesis it's about influenza prediction using deep learning. The data that I had is the rate of dangerous cases and the suspicious ones per week and per region around the country.
for now I implemented the Random Forest, KNN, DNN, LSTM, CNN, CNN-LSTM and Deep Belife Network. I concluded that I'm faced with time series forecasting so I used the window method to make my problem surpervised with window_size=3, 2 and 1. Calculating the r2_score I got 10 regions under the 70% (which I had read that it's the acceptable threshold).
So I'm writing this question hoping to find a solution or get some idea or another deep learning technique and maybe special architecture of the technique used above to improve my prediction in this region.
and thank you in advance
(You will found attached some picture of the regions that I want to improve and my LSTM model.)
In time series, there are various methods derived from ARIMA. These methods required that your serie (or some n-derivate of it) be stationary. This book is a good reference about time series and ARIMA.
https://www.amazon.com/Time-Analysis-Forecasting-George-Box/dp/0470272848
On the other hand, deep learning methods work in most general cases. RNN neuronal network is used in short time series and LSTM in long short series. Furthermore, there is another recurrent architecture known as GRU (gated recurrent united) that works well in long time series.
GRU: https://arxiv.org/pdf/1412.3555.pdf
Comparing Models for Time Series Analysis: https://repository.upenn.edu/cgi/viewcontent.cgi?article=1166&context=wharton_research_scholars
Spatial autoregressive models are statistical models A good paper is : ``About predictions in spatial autoregressive models: Optimal and almost optimal strategies’’.
The following link for the R codes used to obtain the simulation results included in this paper :
http://www.thibault.laurent.free.fr/code/GLT/GLT.html
Hello everyone ;
thank you all for your reply I tried GRU and RNN as mr Roger Figueroa Quintero but without any improvement. About the answer of mr Mohammad Amin Motamedi thanks I had already following these likns. Mr Ayman Amin I will try to built this model thank you but I'm looking for Deep learning technique.
Did someone had some other tips to apply?
For many problems LSTM should outperform other common techniques but it may be quite difficult to train. I suggest to have a look:
https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
https://towardsdatascience.com/time-series-analysis-visualization-forecasting-with-lstm-77a905180eba
All the necessary changes, parameteric variation and code sample can be found here at this great tutorial ... https://medium.com/open-machine-learning-course/open-machine-learning-course-topic-9-time-series-analysis-in-python-a270cb05e0b3
However, i do not recommend to use LSTM for this application ...it generally is good for NLPs but with this task at hand..you gonna face a lot of problem in its training and data preparation..
All the best
Go for LSTMs after performing suitable preprocessing of data. It's better than GRU.
Chapter Deep Learning for Time Series Forecasting
Chapter Deep Learning-Based Time Series Forecasting
In the past few years, many notable architectures have been published such as the Multi-Horizon Quantile Recurrent Forecaster (MQRNN) and the Deep Space-State Models (DSSM). All these models have leveraged Deep Learning to contribute many novelties to the time series forecasting domain.
LSTM stands for Long short-term memory. LSTM cells are used in recurrent neural networks that learn to predict the future from sequences of variable lengths. Note that recurrent neural networks work with any kind of sequential data and, unlike ARIMA and Prophet, are not restricted to time series.
- Badges
- Science topic