BPTF.py 文件源码

def _update_time_parameters(self, user_factors, item_factors, time_factors, ratings, factors_mu, factors_variance, time_id):
        index = ratings.keys()
        QQ, RQ = 0.0, 0.0
        for dim0, dim1 in index:
            Q = user_factors[dim0, :] * item_factors[dim1, :]
            QQ += np.mat(Q).transpose() * np.mat(Q)
            RQ += (ratings[dim0, dim1] - self.mean_rating) * Q

        RQ = np.reshape(RQ, newshape=(RQ.shape[0], 1))
        if time_id == 0:
            mu = (time_factors[1, :] + factors_mu) / 2
            sigma_inv = np.linalg.inv(2 * factors_variance + self.rating_sigma * QQ)
        elif time_id == self.time_num-1:
            sigma_inv = np.linalg.inv(factors_variance + self.rating_sigma * QQ)
            Tk_1 = np.reshape(time_factors[self.time_num-2, :], newshape=(time_factors.shape[1], 1))
            mu = sigma_inv * (np.dot(factors_variance, Tk_1) + self.rating_sigma * RQ)
        else:
            sigma_inv = np.linalg.inv(2 * factors_variance + self.rating_sigma * QQ)
            Tk = time_factors[time_id-1, :] + time_factors[time_id+1, :]
            mu = sigma_inv * (np.dot(factors_variance, np.reshape(Tk, newshape=(Tk.shape[0], 1))) + self.rating_sigma * RQ)

        return np.random.multivariate_normal(mu, sigma_inv)