Gain Cumulative Discounté Normalisé (NDCG) est une métrique populaire utilisée dans la récupération d'informations and systèmes de recommandation to evaluate the quality of ranked lists. NDCG takes into account the position of relevant documents in the ranked list, emphasizing the importance of higher-ranked items. The metric is especially useful in scenarios where the relevance of items is not binary but graded (e.g., on a scale from 0 to 3).
Le calcul de la NDCG implique deux étapes principales : d'abord, computing the Gain Cumulatif (CG), which sums the relevance scores of the retrieved items based on their ranks. The second step is applying a Fonction de Discount, often logarithmic, to reduce the weight of lower-ranked items, reflecting the principle that users are more likely to engage with items presented earlier in the list.
Pour normaliser le CG, la NDCG utilise un Gain Cumulatif Discounté Idéal (IDCG), which represents the maximum possible gain for an ideal ranking des éléments. Le score final de la NDCG est calculé comme suit :
NDCG = (DCG) / (IDCG)
This normalization ensures that NDCG values range from 0 to 1, making it easier to compare the effectiveness of different algorithmes de classement. A higher NDCG score indicates better retrieval performance, as it means that relevant items are ranked higher.
NDCG is widely used in various applications, including search engines, recommendation systems, and apprentissage automatique models, to assess how well a system meets user expectations and retrieves pertinent information.