KAIST GSDS 대학원 박찬영 교수님의 수업인 추천시스템 및 그래프 기계학습 수업을 필기입니다.
Goal of Recommender system
- To identify things the we might like
- To help people discover new content
- To discover which things go together
- To personalize user experiences in response to user feedback
History of Recommender system
- 1990s - First systems(e.g., GroupLens), basic algorithms
- 2000 ~ 2005 - Research explosion, mainstream applications
- 2006 - The Netflix prize
- 2007 - The first Recommender Systems conference
https://recsys.acm.org/
- Present - very active research, many applications
Problem formulation of Recommender system
- Ranking
- Top-k recommendation problem
- Don't need to predict the 'ratings' to recommend, we need top-k items
- Prediction
- Rating prediction, matrix completion problem
- misssing rating value in incomplete user-item rating matrix
Types of Ratings
Explicit feedback
- Contiuous value
- Interval-based
- Binary
Users are not always willing to rate many items
Number of ratings could be small -> spares rating matrices -> poor recommendation quality
Missing values
Pre-substitution of missing ratings is not recommended
-> 어떤 숫자로든 missing values를 처리하면 Signigicant amount of bias 가능성 존재
Impicit feedback
- Unary: 1 type of feedback ex) click or not
Do not requires additional efforts from user
Easy to collect, but less precise
No mechanism to specify a dislike
-> 책을 구매했다고 해당 책이 꼭 맘에 든다고 할 수 없음, 샀지만 재미없을 수도 있거나 여러 경우의 수 있음
Missing values
Recommended to treat the missing entries as 0s to train an algorithm
-> 0으로 missing values를 처리하지 않으면 Significant overfitting 가능성 존재
Business-centric goals fo recommender system
Primary goal: Increasing product sales
- Relevance: recommend items that are relevant to the user using RMSE, MAE...
- Novelty: User has not seen in the past
ex) 인도음식 즐기던 사람에게 새로운 인도음식점 소개해주기 - Serendipity: truly surprising to the user
user에게 새로운 취향을 제공하면서 장기적으로 benefits이 있을 수 있지만, 관심 없는 items를 추천할 수도 있음
ex) 인도음식 즐기던 사람에게 한국음식점 소개해주기 - Diversity: when all these recommended items are very similar, it increases the risk
'RecSys' 카테고리의 다른 글
| Collaborative Filtering Recommendation (0) | 2022.09.07 |
|---|---|
| Content-Based Recommendation (0) | 2022.09.01 |