隐式/显式反馈

推荐系统中用户对物品的反馈分为显式和隐式反馈，’式‘字容易写错，特意将字体加粗…
显式反馈 (如评分、评级) 或单一的隐式反馈 (如浏览、点击、加入购物车)

隐式反馈 (implicit feedback)：
CTR (Wangchong)
CDL (Wang Hao)
Neural Collaborative Filtering (Xiangnan He)

显式反馈 (explicit feedback)
ConvMF

显式反馈转换为隐式反馈
转换的方式主要有以下几种：
以 movielens 数据集为例，电影评分范围为：
1，2，3，4，5(observed) and missing value(unobserved)
(1) rating>=3 : r = 1(正样本，observeed, positive sample) otherwise r = 0(负样本，unobserved, negative sample)
(2) rating不为空 : r = 1(正样本，observeed, positive sample) otherwise r = 0(负样本，unobserved, negative sample)，这种方式将所有缺失项都看成是负样本。
(3) rating不为空 : r = 1(正样本，observeed, positive sample) otherwise 通过负采样 r = 0(负样本，unobserved, negative sample)，这种方式对缺失项都进行采样，选取一部分作为负样本，如论文Neural Collaborative Filtering (Xiangnan He)。

特别注意：
隐式反馈中，r = 1并不代表用户喜欢该物品(电影)，因为不是每个打分都是5分，而且打分因人而异，有的用户比较宽容，打分偏高，有的用户比较严格，打分偏低；r=0也不意味着用户不喜欢该物品，也许用户喜欢该物品只是没有发现该物品而已。r的值仅仅代表用户对该物品有无评分操作，除此之外，别无他意。

隐式反馈的作用——显著提高推荐系统的预测准确率
In other words, a user implicitly tells us about her preferences by choosing to voice her opinion and vote a (high or low) rating. This reduces the ratings matrix into a binary matrix, where “1” stands for “rated”, and “0” for “not rated”. Admittedly, this binary data is not as vast and independent as other sources of implicit feedback could be. Nonetheless, we have found that incorporating this kind of implicit data – which inherently exist in every rating based recommender system – significantly improves prediction accuracy.

Reference
[1] Factorization Meets the Neighborhood: a Multifaceted Collaborative Filtering Model
[2] 段玉超. 面向显式反馈和隐式反馈的协同推荐算法研究[D]. 深圳大学, 2016.