xz76/dna-barcode-classification

README

Methods

In this competition, I used seven methods to approach the prediction.

1. 1D CNN

   • File location: Methods/1D_CNN.ipynb
   • The final accuracy at Kaggle is 94.004%.

2. 2D CNN trained on the one-hot encoder

   • File location: Methods/2D_CNN.ipynb
   • The final accuracy at Kaggle is 95.208%.

3. LSTM/Recurrent neural networks

   • File location: Methods/LSTM_Recurrent_neural_networksipynb.ipynb
   • The final accuracy at Kaggle is 94.943%.

4. Variational auto encoder (VAE) fed into a supervised learning algorithm

   • File location: Methods/Variational_auto_encoder_(VAE).ipynb
   • The final accuracy at Kaggle is 94.317%.

5. Feed-forward neural network

   • File location: Methods/FNN.ipynb
   • The final accuracy at Kaggle is 94.341%.

6. Natural language processing(NLP) embeddings of k-mers + multinomial

   • File location: Methods/NLP_with_Multinomial_Classifier.ipynb
   • The final accuracy at Kaggle is 96.990%. (Due to the kaggle entry issue, I did not submit this result successfully.)

7. Natural language processing(NLP) + CNN

   • File location: Methods/NLP_with_CNN.ipynb
   • The final accuracy at Kaggle is 96.291%.

8. Consider new species.(Submission failed)

   • File location: Data/mv_predict_3.csv
   • The final accuracy at Kaggle is 97.592% on the public leaderboard, and 97.432% on private leaderboard.
   • The imperfect thing is that there should be five entries each day, but after four entries, the system did not allow me to submit the prediction. I am new to kaggle so I have no idea regarding this issue. This is the best prediction among all my results after I taking new species into account. The main idea is combining all seven results together and training to figure out the new species.

Data

• All methods can be run directly by using the original data sets at Data folder. No external data is needed.

• Since different methods may use a different types of data forms, such as character to integer, one-hot encoder, or k-mers, thus, different methods were organized separately and can be run independently.