Our assumption is that we can improve Learning to Rank (LTR) model training by using abandoned queries data.
I've started to collect some references and ideas at https://wikitech.wikimedia.org/wiki/User:GModena_(WMF)/Notes/MLR

Some approaches we could investigate include:

• Abandoned queries as negative examples: the idea is to incorporate abandoned queries as negative training signals in mjolnir, indicating rankings that failed to satisfy user intent.
• Dwell time and abandonment features: extract features from abandonment patterns (quick abandonment vs. abandonment after viewing results) to differentiate between different types of negative feedback.
• Query reformulation analysis: we could study how users modify abandoned queries to better understand query intent and potential ranking failures.
• Contextual abandonment modeling: can we use session context when using abandonment signals, to estimate the reason for abandonment?
• Multi-objective optimization: the idea is to balance traditional relevance metrics (ndcg) with abandonment reduction goals in mjonlir taining objective function.

Depending on how far we want to go, we could consider also consider Counterfactual Learning in conjunction with abandonment data. The idea here is that abandoned queries provide implicit negative feedback
that standard supervised learning approaches might not properly utilize. Counterfactual learning could help address this by allowing us to estimate "what would have happened if a different ranking had been shown".