The CSV file upload is intended to be a temporary solution until you set up your events ingestion. We do however recommend using CSVs with samples of real events for optimal training and performance of the models.
If your CSV has successfully uploaded but the model does not appear to be trained correctly, here are a few possible reasons:
ObjectIDs:
The objectIDs included in your CSV file must exist in the index the model is training on.
Timestamps:
The timestamps should cover a period of at least 30 days. For example, if you have dates starting from 19.01.25, the dates should cover a period up to 18.02.25.
The timestamp of each event should be in a standard format: ISO8601 or RFC3339 (with or without the time).
Events older than 90 days are ignored.
User Tokens:
It's important to consider the distribution of userTokens as these will have an impact on Frequently Bought Together and Related Items model training.
Since these models mostly look at user behaviour and interaction patterns to make recommendations, there is a likelihood that they might not learn anything if every event has a different user token.
Event Types:
The eventType field can only take one of 2 specific values: "click" or "conversion". Recommend models do not take into account eventSubType.
More information on the format of your CSV file can be found here.
Useful Guides:
- Uploading events with multiple ObjectIDs via CSV for Recommend models
- [Recommend] Why do I get an error when uploading my events from a CSV file?
- Can I upload more than 100MB in a CSV of historical events for Recommend?