For each query there exist a class of highly relevant items and a class of
marginally relevant items. The relevance assessment per query can be found
here . Because of this subdivision, most of the evaluation measures
have been split up as well. The following names are given: evaluation
measures(highly relevant) and evaluation measures(relevant). The evaluation
measures(higly relevant) are based on the highly relevant items only, while the
evaluation measures(relevant) are based on all relevant items (highly relevant
items + marginally relevant items).
The uploaded ranked lists are turned into a gain vector by replacing items IDs
by their relevance scores. A highly relevant retrieved item corresponds to
relevance score 2, a marginally relevant retrieved item corresponds to
relevance score 1 and a non-relevant retrieved item corresponds to relevance
score 0.
| * |
True Positives(highly relevant) |
= Vh |
| * |
False Positives(highly relevant) |
= Va - Vh |
| * |
True Negatives(highly relevant) |
= Ds + (Vh - Va) - Ch |
| * |
False Negatives(highly relevant) |
= Ch - Vh |
| * |
First Tier(highly relevant) |
= (# visible highly relevant items within the first d1 items of the rankedlist/ d1 ) * 100% (if Va < Ch then d1 = Va
else d1 = Ch)
|
| * |
Second Tier(highly relevant) |
= (# visible highly relevant items within the first d2 items of the rankedlist/ d2 ) * 100% (if Va < 2 * Ch then d2
= Va else d2 = 2*Ch)
|
| * |
Precision(highly relevant) |
= Vh / Va
|
| * |
Recall(highly relevant) |
= Vh / Ch |
| * |
Average Precision(highly relevant) |
= average of the Precision(highly relevant) scores after each highly relevant
retrieved item |
| * |
True Positives(highly relevant) |
= 5 |
| * |
False Positives(highly relevant) |
= 14 - 5 = 9 |
| * |
True Negatives(highly relevant) |
= 1814 + 5 - 14 - 6 = 1799 |
| * |
False Negatives(highly relevant) |
= 6 - 5 = 1 |
| * |
First Tier(highly relevant) |
= (4/6) * 100 % = 66.667% |
| * |
Second Tier(highly relevant) |
= (5/12) * 100% = 41.667% |
| * |
Precision(highly relevant) |
= 5 / 14 = 0.35714 |
| * |
Recall(highly relevant) |
= 5 / 6 = 0.83333 |
| * |
Average Precision(highly relevant) |
= (1/1 + 2/2 + 3/4 + 4/5 + 5/11) / 5 = 0.80090 |