From a7d0ae5f97b51c38103fc4672909f07f22d32516 Mon Sep 17 00:00:00 2001 From: Pavel Metrikov <46672636+metpavel@users.noreply.github.com> Date: Thu, 13 Jul 2023 00:30:02 -0700 Subject: [PATCH] Update Advanced-Topics.rst --- docs/Advanced-Topics.rst | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/Advanced-Topics.rst b/docs/Advanced-Topics.rst index 5fe16f677320..d2beaacb6838 100644 --- a/docs/Advanced-Topics.rst +++ b/docs/Advanced-Topics.rst @@ -99,4 +99,4 @@ In this case, LightGBM will load the position file automatically if it exists. Also, you can include position column in your data file. Please refer to the ``position_column`` `parameter <#position_column>`__ in above. -Currently, implemented is an approach to model position bias by using an idea of Generalized Additive Models (`GAM `_) to linearly decompose the document score ``s`` into the sum of a relevance component ``f`` and a positional component ``g``: ``s(x, pos) = f(x) + g(pos)`` where the former component depends on the original query-document features and the latter depends on the position of an item. During the training, the compound scoring function ``s(x, pos)`` is fit with a standard ranking algorithm (e.g., LambdaMART) which boils down to jointly learning the relevance component ``f(x)`` (it is later returned as an unbiased model) and the position factors ``g(pos)`` that help better explain the observed (biased) labels. Similar score decomposition ideas have previously been applied for classification & pointwise ranking tasks with assumptions of binary labels and binary relevance (a.k.a. "two-tower" models, refer to the papers: `Towards Disentangling Relevance and Bias in Unbiased Learning to Rank `_, `PAL: a position-bias aware learning framework for CTR prediction in live recommender systems `_, `A General Framework for Debiasing in CTR Prediction `_). In LightGBM, we adapt this idea for the general pairwise Lerarning-to-Rank with arbitrary ordinal relevance labels. Besides, GAMs have been used in the context of explainable ML (`Accurate Intelligible Models with Pairwise Interactions `_) to linearly decompose the contribution of each feature (and possibly their pairwise intercations) for subsequent analysis and interpretation of their effects in the trained models. +Currently, implemented is an approach to model position bias by using an idea of Generalized Additive Models (`GAM `_) to linearly decompose the document score ``s`` into the sum of a relevance component ``f`` and a positional component ``g``: ``s(x, pos) = f(x) + g(pos)`` where the former component depends on the original query-document features and the latter depends on the position of an item. During the training, the compound scoring function ``s(x, pos)`` is fit with a standard ranking algorithm (e.g., LambdaMART) which boils down to jointly learning the relevance component ``f(x)`` (it is later returned as an unbiased model) and the position factors ``g(pos)`` that help better explain the observed (biased) labels. Similar score decomposition ideas have previously been applied for classification & pointwise ranking tasks with assumptions of binary labels and binary relevance (a.k.a. "two-tower" models, refer to the papers: `Towards Disentangling Relevance and Bias in Unbiased Learning to Rank `_, `PAL: a position-bias aware learning framework for CTR prediction in live recommender systems `_, `A General Framework for Debiasing in CTR Prediction `_). In LightGBM, we adapt this idea for the general pairwise Lerarning-to-Rank with arbitrary ordinal relevance labels. Besides, GAMs have been used in the context of explainable ML (`Accurate Intelligible Models with Pairwise Interactions `_) to linearly decompose the contribution of each feature (and possibly their pairwise interactions) for subsequent analysis and interpretation of their effects in the trained models.