Commits on Jan 8, 2024

1. training.py: two tweaks to feature selection …

   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!).

   

   jktomer committed Jan 8, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for 14248ed
   • Browse repository at this point

   Copy the full SHA

   14248ed View commit details

   Browse the repository at this point in the history