CG
cumulative gain
for each query result, it has a relevant score ,
DCG
discounted cumulative gain
take the position of each result into account,
it is calculated as following in industral environment
where , the numerator using exprenation seems to winden the gap between top scores and low scores, and minus one is to reback to [0, 1]. Add 1 is to set the value of denominator into
NDCG
normalized discounted cumulative gain
The results size of each query is not the same, which means the max DCG of each query is not the same. So, DCG needs normalizing.
where the (ideal) is the best search result by human sorting.