[SOLR-13903] Classification Model Confusion Matrix Discrepancy

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 8.2
• Component/s: streaming expressions
• Labels: classification
• Assignee: Unassigned
• Reporter: Ahmed Adel
• Created: 08/Nov/19 04:00
• Updated: 21/Jan/20 16:59

Description

Using features and train stream sources generate a model with TP, TN, FP, FN fields. For some reason, the summation of the values of these fields is sometimes less than the training set size.

 How to regenerate:

1. Create two collections: cellphones and cellphones-model

2. Indexing the attached dataset into cellphones

3. Run the following expression:

commit(cellphones-model,update(cellphones-model,batchSize=500,}}
   train(cellphones,
     features(cellphones, q=":", featureSet="featureSet",
 field="title_t",
 outcome="brand_i", numTerms=25),
 q=":",
 name="cellphones-classification-model",
 field="title_t",
 outcome="brand_i",
 maxIterations=100)))

4. Run the following query to retrieve confusion matrix:

search q=:&collection=cellphones-model&fl=name_s,trueNegative_i,truePositive_i,falseNegative_i,falsePositive_i,iteration_i&sort=iteration_i%20desc&rows=100

The summation of the metrics TP, TN, FP, FN is always less than the training set size by one in this instance for all iterations.