Recommendation Engine - How to give click through rate or number of likes in reranking

[NajiAboo]

We are using Haystack semantic search to search the documents. We also provided the option to thump up and thump down based on the recommendation of the system. We are keeping the count of thump up and thump down and next time when they search we are also sorting with these counts and trying to bring the result to the top [ Here a lower accuracy result may come at the top if it gets more thump ups]. My question is, There is any way to include these counts as a parameter when i do retraining so that this will automatically update the accuracy score?

[vblagoje]

Hey there, what kind of retraining do you plan to do? It is possible but not trivial. I am assuming you want to improve the retriever. If that is the case, a better option would be to use GPL to improve your retriever.

[NajiAboo]

Consider the query "How to solve the overheat in a refrigerator? ". Our recommendation engine will bring recommendations and be listed based on accuracy. If the user finds the 4th one is more accurate, then there is an option to upvote it. Currently, we are saving the upvote count and sorting with the upvote when next time somebody asks similar questions (sort by upvote and then accuracy). We are thinking to retrain the model with this upvote count also, so that next time when similar questions comes we will get high accuracy. But still not sure its a good idea to do so and possible to do ?