Basically, my approach here is as follows:

• Train individual models for symbols and numbers
• Segment larger test image into fifths, predict symbols and numbers separately
• For each fifth, generate the set of symbols/numbers where that fifth is that symbol with >= .2 probability.
• Find the cartesian product of each set of symbols/numbers
• For each combination of predicted symbols/numbers (or each unique expression), predict 1 if any of the expressions evaluate to true.
• Predict 0 if none of the predicted symbol/number combinations (expressions) evaluate to true

To run

• Make a folder called data and put train.csv, train_labels.csv, and test.csv in the folder.
• Run python process_data.py and python score.py

Thanks!