This is a general maintenance release of Kraken, so these notes are going to be a quick read for you. 🙂
Updates in this release
- Kraken now correctly handles datasets with reserved characters in the target column name. For the complete list of reserved characters, see "section 2.2 Reserved Characters (January 2005)" of this article.
- When evaluating columns to automatically exclude based on high cardinality, Kraken now bases the decision solely on whether the number of unique values exceeds 90% of the total number of values in a column. Previously, Kraken would also exclude a column if the total number of unique values exceeds 1,000.
- Prediction Influencers (a.k.a. SHAP values) calculations and documentation have been improved.
- SHAP values for Random Forest are now correctly used to indicate the probability of the positive class, instead of negative class.
- SHAP values for XGBoost are now more accurate and they are no longer based on sampling.
- See the new article on How Kraken calculates Prediction Influencers for details on the SHAP calculation methods used.
Reminder: If Kraken is misbehaving, check KIKI first.
Registered Kraken users have access to KIKI, a.k.a. "Kraken Important Known Issues". If you're experiencing an issue with Kraken, KIKI is your first stop to determine if we're aware of the issue. Where possible, we provide suggested workarounds while we work to eliminate issues. KIKI will be updated as we eliminate issues through Kraken product updates, as well as when new issues are identified.
Reminder: SONAR© Guide content is just a click away!
As a Kraken user, you have access to SONAR© Guide, the proven training methodology from Big Squid and your key to learning everything you need to know to use Kraken effectively and become a citizen data scientist hero in the process. A link to the SONAR© Guide content library is accessible whenever you access Kraken help screens; just look for the SONAR© Guide logo on the right-hand side. It's also available in the SONAR© Guide section of support.bigsquid.com.