training.py: two tweaks to feature selection (#226)

1. Include posting amounts as a feature. This allows us to distinguish
different classes of payments to the same payee (e.g. recurring membership
fees, which often have a constant amount, from individual purchases).

2. For example key/value pairs, include the key by itself (with no substring
of the value) as a feature. This is useful because different account types
often have non-overlapping sets of example keys, and including the bare key as
a value allows the decision tree to be effectively segmented by account type
fairly close to the root.

These two very small changes significantly improve training accuracy on my
journal, from 94.81% to 99.32% (an 86% reduction in error rate!).

Author: jktomer

beancount_import/training.py

@@ -30,10 +30,11 @@
 def get_features(example: PredictionInput) -> Dict[str, bool]:
     features = collections.defaultdict(lambda: False)  # type: Dict[str, bool]
     features['account:%s' % example.source_account] = True
-
-    # For now, skip amount and date.
+    features['amount:%s' % example.amount.currency] = example.amount.number
+    # For now, skip date.
 
     for key, values in example.key_value_pairs.items():
+        features[key] = True
         if isinstance(values, str):
             values = (values, )
         for value in values:

beancount_import/training_test.py

@@ -1,6 +1,7 @@
 import datetime
 
 from beancount.core.data import Amount
+from beancount.core.number import D
 from . import test_util
 from . import training
 
@@ -21,7 +22,10 @@ def test_get_features():
                 'a:hello': True,
                 'b:foo': True,
                 'b:bar': True,
-                'b:foo bar': True
+                'b:foo bar': True,
+                'a': True,
+                'b': True,
+                'amount:USD': D(3)
             }