ListMLE Loss

The original paper link is down :’( range:

summary

• Listwise Maximum Likelihood Estimation.
• Use for learning to rank:
  • https://www.auai.org/uai2014/proceedings/individuals/164.pdf
  • 
  • The above probability is defined according to PlackettLuce model.
• the loss is minimized when we maximize the probability of the observed ranking.
• more info: https://notesonai.com/ListNet+and+ListMLE
• nice github repo: https://github.com/allegro/allRank/tree/masterlink:
• pytorch implmentation: https://github.com/Obs01ete/kaggle_latenciaga/blob/b6b55b16fd444d423946eab3d5c8b293f7dc8939/src/losses.py#L62

pros

Cons

• ListMLE cannot well capture the position importance, which is a key factor in ranking
  • https://www.auai.org/uai2014/proceedings/individuals/164.pdf
    • the paper mentions:
    • It views the ranking problem as a sequential learning process, with each step learning a subset of parameters which maximize the corresponding stepwise probability distribution. To solve this sequential multi-objective optimization problem, we propose to use linear scalarization strategy to transform it into a single-objective optimization problem, which is efficient for computation