Need clarification on multiple input fields in Predictive Intelligence Similarity solution
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
49m ago
We are reviewing the OOTB Major Incident Recommendation Predictive Intelligence Similarity Definition (ml_sn_global_major_incident_recommendation).
In our instance, the OOTB configuration currently has:
- Table: Incident
- Fields: Short description
- Test Table: Incident
- Test Fields: Short description
- Filter: Active = true
We understand that the current configuration uses Short description when determining similarity between incident records.
We are considering adding Service as an additional field, so that the configuration would become:
Fields
- Short description
- Service
Test Fields
- Short description
- Service
We need clarification on the exact internal behavior of the Predictive Intelligence Similarity model when multiple fields are configured, specifically:
- How are Short description and Service used together when calculating the similarity score?
- Does the model calculate similarity independently for each field and then combine the results, or are the values from the fields combined/represented together before similarity is calculated?
- Is there any field-level weighting? For example, does Short description contribute more or less than Service, or are all configured fields treated equally?
- Since Service is a reference field, what information from the Service field is actually used by the model, for example the display value, sys_id, or some other representation?
- If two incidents have highly similar Short descriptions but different Services, how does the difference in Service affect the final similarity score?
- Conversely, if two incidents have the same Service but different Short descriptions, how does that affect the score?
- Does adding Service act only as additional similarity context, or can it effectively enforce/make stronger recommendations for incidents belonging to the same Service?
- Is there any official documentation or KB article explaining the algorithm/behavior when multiple fields, particularly reference fields, are used in a Predictive Intelligence Similarity Definition?
We are specifically looking for the behavior of Predictive Intelligence Similarity, rather than Task Intelligence.
This clarification will help us determine whether adding Service to the OOTB Major Incident Recommendation solution is appropriate, and what effect it would have on similarity results.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
17m ago
Hi @Suggy
From my understanding of how Predictive Intelligence Similarity works:
Fields are combined, not scored separately. The values of all configured fields are combined into one text representation per record, turned into a vector using the word corpus, and compared with cosine similarity. There's no separate per-field score.
There's no field-level weighting. Each field's influence depends on how many meaningful words it adds.
Reference fields like Service contribute their display value (for example "Email Service") as words, not the sys_id.
Effect on scores:
- Similar short description but different Service: the score drops a little. Short descriptions are short, so a multi-word service name can have a noticeable effect.
- Same Service but different short description: the shared service words raise the score, so you may see weaker recommendations that only share the service.
- Adding Service only adds context. It doesn't filter or guarantee recommendations from the same Service. If that's the requirement, keep Short description as the only field and filter the results by Service in the logic that consumes the predictions.
Recommendations:
- Train and test both versions in a sub-production instance with real major incident examples before changing production.
- Changing the OOB solution definition marks it as customised, so document the change or consider using a copy.
- For confirmation of the exact internal behaviour on your release, check the Predictive Intelligence documentation or raise a case with ServiceNow Support.
If this answer helped you, please mark it as Helpful and Accept as Solution so others with the same question can find it.