推荐系统概念解释 and 一个好的推荐系统
用户满意度
- 用户满意度是推荐系统测评的重要指标,但是实际上,用户满意度数据获得的方式十分有限,因为这是一种用户的主观情感。
- 设计合适的方式对于用户的满意度进行回收分析,是改进推荐系统的一个很好的方式。这样的的方式包括但不限于,设计合适的调查问卷,在物品的购买结束后附上一份满意度调查。
- 满意度在一些程度上可以细分为更加具体的信息。例如点击率,用户停留时间,转化率,完播率,或者是哔站视频点赞,三连的比例。
预测准确度
- 召回率 (Recall)
其中
是根据用户在训练集上的行为给用户做出的推荐列表, 是用户在测试集上的行为列表。召回率的意义?可以参考机器学习中留下的定义进行理解
精确率
其中
是根据用户在训练集上的行为给用户做出的推荐列表, 是用户在测试集上的行为列表。精确率的意义?
覆盖率
- 描述了一个系统对于物品长尾的发掘能力。
- 覆盖率的一个定义可以是:
- 其中,用户集合为
,系统为每位用户推荐一个长度为 的物品列表 - 覆盖率的意义:覆盖率越高,以为这系统中被推荐给用户的物品,占所有物品的比例越大,对于一个好的推荐系统,不仅需要有较高的用户满意度,还需要有较高的覆盖率。
- 当然对于覆盖率的定义,不止以上的这一种,甚至说,在实际使用上,上述简单的覆盖率不足以支撑大规模复杂系统的覆盖率计算,所以如何对于覆盖率进行修正和更新?信息熵与基尼系数!
- 推荐了解,马太效应,一个强者更强,弱者更弱的效应,在推荐系统中也同样存在。
多样性
- 假设,
定义了物品 i 和 j 之间的相似度,给用户 的推荐列表 的多样性定义: - 推荐系统整体的多样性可以定义为,用户推荐列表多样性的平均值:
- 假设,
信任度
- 用户对于该系统的信任程度
实时性
- 系统对于数据更新的时效性
健壮性
- 系统对于外来攻击的防护性