auc.py

import numpy as np
from lightfm import LightFM
from sklearn.metrics import roc_auc_score
from lightfm.datasets import fetch_movielens

def auc_score(model, ratings):
    """
    computes area under the ROC curve (AUC).
    The full name should probably be mean
    auc score as it is computing the auc
    for every user's prediction and actual
    interaction and taking the average for
    all users
    
    Parameters
    ----------
    model : BPR instance
        the trained BPR model
        
    ratings : scipy sparse matrix [n_users, n_items]
        sparse matrix of user-item interactions
    
    Returns
    -------
    auc : float 0 ~ 1
        auc score of the model
    """
    ratings = ratings.tocsr()
    
    auc = 0.0
    n_users, n_items = ratings.shape
    for user in range(n_users):
        y_pred = model.user_embeddings[user].dot(model.item_embeddings.T) + model.item_biases
        y_true = np.zeros(n_items, dtype = np.int32)
        y_true[ratings[user].indices] = 1
        auc += roc_auc_score(y_true, y_pred)

    auc /= n_users
    return auc


movielens = fetch_movielens()
train = movielens['train']
test = movielens['test']

model = LightFM(learning_rate=0.05, loss='bpr')
model.fit(train, epochs=10)

auc_score(model, test)