Re: What equation do NDCG used ?
It's the same formula, what do you think is different? On Wed, Apr 4, 2012 at 9:35 PM, ziad kamel ziad.kame...@gmail.com wrote: Hi , I checked the code for NDCG and it seems not same as http://en.wikipedia.org/wiki/Discounted_cumulative_gain How that formula was derived ? Thanks
Re: What equation do NDCG used ?
DCG should check the order and the formula seems not , just check if it contains the item if (relevantItemIDs.contains(item.getItemID())) { cumulativeGain += discount; } Second thing is that it have a relevance number rel which the formula don't use. On Wed, Apr 4, 2012 at 3:54 PM, Sean Owen sro...@gmail.com wrote: It's the same formula, what do you think is different? On Wed, Apr 4, 2012 at 9:35 PM, ziad kamel ziad.kame...@gmail.com wrote: Hi , I checked the code for NDCG and it seems not same as http://en.wikipedia.org/wiki/Discounted_cumulative_gain How that formula was derived ? Thanks
Re: What equation do NDCG used ?
It does use the order: double discount = i == 0 ? 1.0 : 1.0 / log2(i + 1); The score you gain for recommending a relevant item decreases as you go down the list of recommendations. As the comments in the code note, relevance is 1 for relevant items and 0 for others. There is no point in showing a term multiplied by 1, or 0. On Wed, Apr 4, 2012 at 10:19 PM, ziad kamel ziad.kame...@gmail.com wrote: DCG should check the order and the formula seems not , just check if it contains the item if (relevantItemIDs.contains(item.getItemID())) { cumulativeGain += discount; } Second thing is that it have a relevance number rel which the formula don't use. On Wed, Apr 4, 2012 at 3:54 PM, Sean Owen sro...@gmail.com wrote: It's the same formula, what do you think is different? On Wed, Apr 4, 2012 at 9:35 PM, ziad kamel ziad.kame...@gmail.com wrote: Hi , I checked the code for NDCG and it seems not same as http://en.wikipedia.org/wiki/Discounted_cumulative_gain How that formula was derived ? Thanks
Re: What equation do NDCG used ?
It seems that having a recommended list that is for example 9, 23, 8 or 8 , 9 , 23 will give same NDGC , since it just have to be within the relevant list . NDGC usually being used in ranking lists , that is why I think relevant 0 or 1 won't work. On Wed, Apr 4, 2012 at 4:23 PM, Sean Owen sro...@gmail.com wrote: It does use the order: double discount = i == 0 ? 1.0 : 1.0 / log2(i + 1); The score you gain for recommending a relevant item decreases as you go down the list of recommendations. As the comments in the code note, relevance is 1 for relevant items and 0 for others. There is no point in showing a term multiplied by 1, or 0. On Wed, Apr 4, 2012 at 10:19 PM, ziad kamel ziad.kame...@gmail.com wrote: DCG should check the order and the formula seems not , just check if it contains the item if (relevantItemIDs.contains(item.getItemID())) { cumulativeGain += discount; } Second thing is that it have a relevance number rel which the formula don't use. On Wed, Apr 4, 2012 at 3:54 PM, Sean Owen sro...@gmail.com wrote: It's the same formula, what do you think is different? On Wed, Apr 4, 2012 at 9:35 PM, ziad kamel ziad.kame...@gmail.com wrote: Hi , I checked the code for NDCG and it seems not same as http://en.wikipedia.org/wiki/Discounted_cumulative_gain How that formula was derived ? Thanks
Re: What equation do NDCG used ?
No, re-read my last message. The ordering matters, since the discount changes at each position. On Wed, Apr 4, 2012 at 10:36 PM, ziad kamel ziad.kame...@gmail.com wrote: It seems that having a recommended list that is for example 9, 23, 8 or 8 , 9 , 23 will give same NDGC , since it just have to be within the relevant list . NDGC usually being used in ranking lists , that is why I think relevant 0 or 1 won't work.