metrics.py

from sklearn.metrics import roc_auc_score
from sklearn.metrics import log_loss
import numpy as np

def MSE(preds, true):
    squaredError = []
    for i in range(len(preds)):
        val = true[i] - preds[i]
        squaredError.append(val * val)  # target-prediction之差平方
    return sum(squaredError) / len(squaredError)


def MSE_ips(preds, true, user_num, item_num, inverse_propensity):
    squaredError = []
    globalNormalizer = 0
    for i in range(len(preds)):
        val = true[i] - preds[i]
        squaredError.append(val * val * inverse_propensity[int(true[i]) - 1])
        globalNormalizer += inverse_propensity[int(true[i]) - 1]
    # aaa = sum(squaredError) / globalNormalizer # SNIPS
    return sum(squaredError) / (user_num * item_num)


def MAE(preds, true):
    absError = []
    for i in range(len(preds)):
        val = true[i] - preds[i]
        absError.append(abs(val))  # 误差绝对值
    return sum(absError) / len(absError)


def MAE_ips(preds, true, user_num, item_num, inverse_propensity):
    absError = []
    for i in range(len(preds)):
        val = true[i] - preds[i]
        absError.append(abs(val) * inverse_propensity[int(true[i]) - 1])  # 误差绝对值
    return sum(absError) / (user_num * item_num)


def RMSE(preds, true):
    squaredError = []
    absError = []
    for i in range(len(preds)):
        val = true[i] - preds[i]
        squaredError.append(val * val)  # target-prediction之差平方
        absError.append(abs(val))  # 误差绝对值
    from math import sqrt
    return sqrt(sum(squaredError) / len(squaredError))


def RMSE_ips(preds, true, user_num, item_num, inverse_propensity):
    squaredError = []
    for i in range(len(preds)):
        val = true[i] - preds[i]
        squaredError.append(val * val * inverse_propensity[int(true[i]) - 1])
    from math import sqrt
    return sqrt(sum(squaredError) / (user_num * item_num))


def sigmoid(x):
    return 1.0 / (1.0 + np.exp(-x))


def AUC(true, preds):
    return roc_auc_score(true, preds)


# def NLL(true, preds):
#     import math
#     s = 0
#     for i in range(len(true)):
#         s += math.log(1 + math.exp(- true[i] * preds[i]))
#     return - s / len(true)

def NLL(true, preds):
    return -log_loss(true, preds, eps=1e-7)