数据挖掘笔记第六节 推荐系统

Posted 2021-04-27 是坠坠啊

写在前面

• 本文为《数据挖掘》课程推荐系统部分重要的知识点，为相对简略的知识点汇总。文末点击 阅读全文获取思维导图，后期复习的时候可以借助其快速地过知识点。
• 教材：《数据挖掘导论》：Tan, Pang-Ning, et al. Introduction to data mining, 2018.

Basics of Recommendation Systems

Formal Model

U: the set of users

I: the set of items

U×I→R: the set of ratings

– Explicit 显式：Ask people to rate items (like 1-5 stars)

– Implicit 隐式：Learn from user behaviors (like click or not).

Goal:

– To predict the rating a user would give to an item.

Key Steps of Recommendation

1. Gather known ratings

2. Extrapolate unknown ratings from the known ones

• Mainly interested in high unknown ratings

3. Evaluate extrapolation methods

Mainstream Methods

• Content-based recommendation

• Collaborative-based recommendation

• Memory-based methods

• User-oriented methods

• Item-oriented methods

• Model-based methods

• Latent-factor-based methods

• Hybrid recommendation

Evaluation

Offline 离线评价，数据切分成训练集和测试集

User Survey

Online

Evaluation Metrics

• Compare predictions with known ratings 类似回归问题的评价指标

• Root-mean-square error (RMSE)

• Precision at Top 10:

• % of those in Top 10 （10个中预测中了几个）

• Rank Correlation:

• Spearman’s correlation between system’s and user’s complete rankings

• Another approach: 0/1 model (click or not) 类似分类问题的评价指标

• Coverage: 一般能达到100%

• Number of items/users for which system can make predictions

• Precision:

• Accuracy of predictions (predictions &recall)

• Receiver Operating Characteristic (ROC)

• Tradeoff curve between false positives and false negatives

• More Thoughts on Evaluation

• Narrow focus on accuracy sometimes misses the point

• Prediction Coverage 小众的也要推，对于平台发展有好处

• Prediction Diversity 要看用户对于新鲜事物的态度，抓住尝鲜动机强的用户

• Prediction Novelty

• Real-time

• Robustness 海量数据时要求稳健

• In practice, we care only to predict high ratings:

• RMSE might penalize a method that does well for high ratings and badly for others （RMSE 在大规模/大距离数据时表现不好）

Representative Recommendation Techniques

Content-based Recommendation

Main idea:

• Recommend items to customer i similar to previous items rated highly by i

Item & User Profile

• For each item, create an item profile

• Profile is a set (vector) of features

• Movie: author, title, actor, director,…

• Text: the set of “important” words in document

• User profile possibilities:

• Weighted average of rated item profiles

• Variation: weight by difference from average rating for item

Prediction

cosin similarity

A&D

Collaborative-based Recommendation 协同过滤

Basic Assumption

• Those who agreed in the past tend to agree again in the future

Two Mainstreams

1.Memory-based Methods

修正，纳入固定效应

2.Model-based Methods

• Cluster users and then recommend items the users in the cluster closest to the active user like.

• Mine association rules and then use the rules to recommend items.

• Learn a latent factor model from the data and then use the discovered factors to find items with high expected ratings.

• 优化问题

• Overfitting：加入正则项

• Stochastic Gradient Descent
• GD v.s. SGD

• GD 迭代次数少，每个步骤慢，光滑下降

• SGD 迭代次数多，每个步骤快，波动下降

• 整体算出来的性能差不多

• Extending LFM to Include Biases

• 考虑固定效应

• 考虑时间效应

Extended Applications

Social Recommendation

Jointly factorize the rating matrix and social matrix

A missing rating for a user is predicted as a linear combination of ratings of the user and her social relations

写在最后

• 点击 阅读全文可查看思维导图。
• 如果你觉得文章还不错，欢迎点亮 在看，感谢你的支持~
• 关注公众号，收取更多学习笔记。
• 思维导图编辑软件： 幕布，扫描以下二维码获取30天高级会员