Novelty

Novelty metrics assess the extent to which a recommender system suggests items that are new or unexpected to the user, beyond what is already popular or frequently consumed. These metrics are important for fostering exploration and serendipity, as highly novel recommendations can lead to delightful discoveries.

API Reference

For class signatures and source code, see the Novelty Metrics API Reference.

EFD

Expected Free Discovery (EFD@K). Estimates the likelihood that users discover relevant but less popular (unexpected) items in their top-K recommendations, promoting serendipity.

\[ \text{EFD@}K = \frac{1}{|\mathcal{U}| \cdot C} \sum_{u \in \mathcal{U}} \sum_{i=1}^{K} \frac{r_i \cdot (-\log_2 p_i)}{\log_2(i + 1)} \]

where \(p_i = n_i / |\mathcal{U}|\) is the popularity of item \(i\) and \(C = \sum_{i=1}^{K} 1/\log_2(i+1)\).

For further details, please refer to this link.

evaluation:
    top_k: [10, 20, 50]
    metrics: [EFD]

Extended-EFD is also available, meaning you can compute the EFD score using a discounted relevance value, as follows:

evaluation:
    top_k: [10, 20, 50]
    complex_metrics:
        - name: EFD
          params:
              relevance: discounted

EPC

Expected Popularity Complement (EPC@K). Measures the average complement of item popularity in the top-K recommendations, encouraging exposure to less popular content.

\[ \text{EPC@}K = \frac{1}{|\mathcal{U}| \cdot C} \sum_{u \in \mathcal{U}} \sum_{i=1}^{K} \frac{r_i \cdot (1 - p_i)}{\log_2(i + 1)} \]

For further details, please refer to this link.

evaluation:
    top_k: [10, 20, 50]
    metrics: [EPC]

Extended-EPC is also available, meaning you can compute the EPC score using a discounted relevance value, as follows:

evaluation:
    top_k: [10, 20, 50]
    complex_metrics:
        - name: EPC
          params:
              relevance: discounted