Showing all feedback negative

[generalomosco]

Don't know why it is showing all feedback back negative when some values is 0 and 1 in feedback table, what i did is passing "read" as feedback with zero value after queried to prevent fetching same items over and over (it will be nice passing exclude items IDs as query to avoid this), so after when the user scrolled to each post feed i set the "read" feedback to -0.5 if not engaged with the post or value to 1 if engaged with post but still all feedback is showing negative in dashboard.

[zhangzhenghao]

This is most likely expected behavior rather than a dashboard bug.

In Gorse, there are three different concepts here:

1. positive_feedback_types
   • used as positive samples in ranking / CTR training
2. negative_feedback_types
   • used as explicit negative samples
3. read_feedback_types
   • used as implicit negative samples for ranking / CTR training

So if your read feedback type is configured in read_feedback_types, then entries of type read are treated as negative samples in the training dataset unless the same user-item pair is already positive.

From the code path in master/tasks.go:

• explicit negative feedback is added first (highest priority)
• positive feedback is added next
• read_feedback_types is loaded after that and inserted as implicit negative
• finally both explicit negative and read feedback are counted into ctrDataset.NegativeCount

That means this setup:

• send read with 0 to mark exposure
• later change it to -0.5 if not engaged
• later change it to 1 if engaged

will not automatically make dashboard treat all read records as positive, because the dashboard's negative statistics are based on how Gorse classifies feedback types for training, not simply on whether the raw value column is positive or negative.

A few important notes:

• For built-in feedback semantics, feedback type matters much more than raw value.
• If read is in read_feedback_types, it is considered an implicit negative signal for ranking.
• A positive value on a read record does not make it a positive training sample by itself.
• If you want an engaged item to become positive, you should usually write another feedback type that is included in positive_feedback_types (for example like, click, engaged, complete, etc.).

So for your scenario, I would recommend:

• keep read only for exposure / de-dup / seen-item filtering
• configure read in read_feedback_types if you want it to act as an implicit negative when there is no engagement
• write a separate feedback type such as like / engaged / click for real positive engagement
• include that engagement type in positive_feedback_types

If helpful, I can also suggest a concrete config example for feeds like:

• exposed but skipped
• clicked
• liked
• fully consumed

That would make the training behavior much easier to reason about.

[generalomosco]

You mean Gorse doesn’t care much about value = 1 or -0.5
It cares about whether the feedback type is classified as positive / negative / read.
So can I use “read” as neural then “engaged“ or “watched” as positive and “skip” as negative feedback on viewing each item and later delete the “read” item after updated of engaging feedback