I have been wondering about a particular technique to denoise position bias in learn-to-rank.I am aware of inverse propensity weighting techniques. During a discussion, it was suggested to me , however, to use position as a feature during training from offline data. On inference, position would then be set to a constan value t, for example highest position (1).
I do not understand why this technique makes sense or would work in general. It does in the case of simple linear models, as you set the constant value to 0. Could someone provide me papers which supports or cracks down this method/