Predicting individual-level voting behavior with neural nets and HMM clustering
-
train HMM on sample (parallelized)
-
score all obs on HMM, rank by log likelihood (not parallel yet, unfortunately)
3. split obs into ~10 clusters based on log likelihood ranking
4. Train separate RNN's for each cluster, using sequential data exclusively (parallelized)
5. Train 3 global feed forward neural networks for the 08, 10, and 2012 elections (parallelized)
6. Combine weights (average) for the 3 nets; higher weights for 2012 and 2010
7. Pass respective outputs from RNN and FFNN to a single-hidden-layered FFNN to ensemble intelligently
8. To test
*Pass test data to HMM, calculate log likelihood, rank and cluster
*Pass sequential data to appropriate RNN (RNN that belongs to the individual's cluster), get out predictions
*Pass all individuals data to FFNN, get out predictions
*Pass predictions from RNN and FFNN to ensembler
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent