【Week 1】Preferences and Ratings

内容来源：“Recommender Systems Specialization” course of 明尼苏达大学 (University of Minnesota) at Coursera

一、Learning Objectives

Understand what data recommenders can use to learn what users like
Identify types of data collected from users
Understand when different data types are possible and appropriate
Be able to identify types of preference data likely used in a system
(怎么收集数据、怎么使用、以及它们意味着什么）

二、Preference Model

推荐过程的核心是学习得到用户的偏好信息(preference information)，即用户喜欢什么。
用户在互联网中的各种行为或数据信息能够体现其偏好信息，例如评分、购买、浏览等行为和年龄、性别等数据。这些数据和行为大体可分为两类，即explicit data 和 implicit data。

2.1 Explicit ratings

user直接表明对某个item的喜好信息，例如可以收集用户对电影的打分、对商品的评论等。

How to collect data?

怎么从用户那收集信息？

Star Ratings:

Widely-used interface
Several design decisions
- 5? 7? 10?（5-stars or 7-stars）
- Half-stars?（是否提供半星，即打0.5星）
- Provide meaning/calibration?（是否为每个评级提供说明）
- More not necessarily better（因为用户可能不能很好的利用这些信息，所以有时提供更多的说明信息并不代表能收集到质量更好的数据，但有利于用户表达他们的观点）

Thumbs and Likes

Vote up/down
Or just ‘Like’/‘+1’
Common with ephemeral items （常见于短暂的项目，用户对信息快速进行浏览的场景，例如知乎、微博）
- News aggregation (Reddit, Digg)
- Q&A (StackOverflow)
- YouTube
Very low cost to rate

Other Interfaces

Continuous scales （users can rate anywhere on a fairly wide ranging scale.）
Pairwise preference （to judge whether they like one item better than another）
Hybrid (e.g. 1-100 + never again)
Temporary (e.g. Pandora 30-day suspend，即30天内不再播放该歌曲)

When are ratings provided?

什么时候让用户进行评分？

Consumption — during or immediately after experiencing the item （例如刚看完电影，就让用户进行评分，其感受不受记忆影响）
Memory — some time after experience
Expectation — the item has not been experienced （适用于一些high cost low volume items，例如调研对某片区域房子的偏好信息，实际上拥有房子的用户很少，因此需要对房子进行详细描述，使用户基于该描述对房子进行评分，得到Expectation ratings）

Problems

用户的评分信息是否准确，用户本身是否可靠？
用户的偏好会随着时间改变，但是评分数据不会因此更新。
评分的具体含义不明确，给出相同评分的用户所表达的喜好程度可能不同。

2.2 Implicit data

Data collected from user actions
Key difference: user action is for some other purpose, not expressing preference
Their actions say a lot!

How to collect data?

Reading Time

Early implicit data: how long did user read?
Listening and watching （用户听音乐/观看影片的时间，例如秒切和秒关表示对该内容不感兴趣）
- IMMS
- Video services

Binary actions

Click on link (ad, result, cross-reference)
Don’t click on link
Purchase
Follow/Friend

Subtleties and Difficulties

What does the action mean?
- Purchase: they might still hate it
- Don’t click: expect bad, or didn’t see （没点击有可能仅是没看见，而不是不喜欢）
How to scale/represent actions?
when you are combining multiple types of actions, how you represent them and combine them in order to come up with the recommendations becomes important.
Lots of opportunity to be creepy
- Education may help
- So can respecting privacy
  （行为有歧义，但通过教育用户这些行为将如何影响后续推荐过程，能够获取质量更好的行为。该过程为达到更好推荐效果提供机会的同时也尊重了用户的隐私）

2.3 Key difference between implicit data and explicit ratings

As in Explicit ratings, the user is intending to say how much they like something. In implicit data, the user is just doing something that’s based on their preference but they’re not doing it to say what they prefer. Their actions can say a lot and they have a lot of actions.
（前者是间接表达偏好信息，而后者是直接表达）

三、Conclusion

Recommenders mine what users say (explicit data) and what they do (implicit data) to learn preferences
Ratings provide explicit expressions of
preference
Implicit data benefits from greater volume （相较于explicit ratings，更易获得大量的用户行为数据）