Consider Recall Computation

[apdullahyayik]

Computation of recall at evaluation/metric.py does not consider "k" value which is passed as an arg. Whereas at the same module computation of precision uses "k" value. I think, there may be smthg wrong here.

def recall_at_k(y_true_list, y_reco_list, users, k):
    recall_all = list()
    for u in users:
        y_true = y_true_list[u]
        y_reco = y_reco_list[u]
        common_items = set(y_reco).intersection(y_true)
        recall = len(common_items) / len(y_true)
        recall_all.append(recall)
    return np.mean(recall_all)

[massquantity]

Yep, k of recall in evaluation/metric.py is redundant, but for the sake of consistency with other metrics, I still add it into the recall function. In my opinion, the k of recall_at_k is already included in y_reco, since it has length of k. But I'm not sure about this, any suggestions?

[apdullahyayik]

No I don't have since it is cool. Two more questions here. (1) Computation of recall looks like hit rate: "computation of intersections by discarding the order". Is it true? I may be wrong, maybe hit rate needs considering the order. (2) Where is the difference between precision and recall on the code?

def precision_at_k(y_true_list, y_reco_list, users, k):
    precision_all = list()
    for u in users:
        y_true = y_true_list[u]
        y_reco = y_reco_list[u]
        common_items = set(y_reco).intersection(y_true)
        precision = len(common_items) / k
        precision_all.append(precision)
    return np.mean(precision_all)

Answer is appreciated.

[massquantity]

1. I don't know about hit rate before. After searching on the web, they does look similar.
2. There is only one difference. The denominator in precision is k, whereas that of recall is len(y_true)