The appropriate threshold will vary from one application to another and will depend on the data / model, so it is wise to adjust the initial value via trial-and-error.
Setting a threshold of 0
will translate to recommendations that are "unfiltered" regardless of confidence score, while a threshold of 95 or higher will only return recommendations with extremely high confidence scores.
We also encourage you to revisit threshold values over time for further improvements. As your catalog evolves and as the model continues to improve, a previously configured threshold might need some changes. For example:
- You set a threshold of 75 for a model that is initially returning inconsistent recommendations with lower confidence. After recent improvements from additional training, the model now returns recommendations in the higher confidence range [70-99%] . In this case, the old threshold is now too timid, and you should raise it to a higher value.
- You set an initial threshold of 60 for a very diverse catalog which was quite diverse. After adding new categories in your e-shop, you now see few recommendations meeting that threshold criteria. In this case, the old threshold doesn't account for the variations in the new catalog, so it would be wiser to lower the threshold to allow a wider range of recommendations to pass through.