Recommender Systems
in 15 minutes
by Hans Fredric Waadeland
Agenda
1. What is recommendation systems?
2. Why do we need them?
3. How do they work?
4. Now what?
Recommendation System
[...is a system] ...that seek to predict the "rating" or "preference" that a user would give to an item.
- Wikipedia (2016)
1. What?
Connect users with data
When the dataset are too big to easily search or browse.
When the datasets are always changing or growing.
- source (2016)
2. Why and when?
Background
- Evolved from the information retrieval research field community in the mid 90s
- Focused on matching documents with search queries
- The Netflix Prize
Google Scholar
YouTube
Audible
Spotify
Netflix
Facebook, LinkedIn, Pandora, Last.fm, SoundCloud, news sites, search engines, dating, collaboration, restaurants, financial services, life insurance, products in general and many more.
Approaches
3. How does it works?
1. Content-based filtering
2. Collaboration filtering
3. Hybrid systems
Techniques
3. How does it works?
1. Heuristic-based - "Common sense", "educated guess", rule of thumb etc.
2. Model-based - Analysis, Machine learning
Content-based filtering
3. How does it works?
1. Finds the commonalities of highly rated items in the past
2. Uses those to find similar or compatible items
3. How does it works?
1. Limited Content Analysis - Lack of attributes, meta or "analysable" content
2. Over Specialization - Selecting too similar items
3. New User Problem (cold start) - lack of ratings/feedback
Challenges
of content-based filtering
Collaborative filtering
3. How does it works?
1. Builds a model of the user's preferences based on previous ratings of items
2. Finds other users that have similar preferences
3. Recommends items that are highly rated by those users
3. How does it works?
1. New User Problem (cold start) - lack of ratings/feedback
2. New Item Problem - Lack of ratings
3. Sparsity - handling unusual tastes or items
Challenges
of collaboration filtering
Hybrid systems
3. How does it works?
1. CB-filtering and C-filtering separately and combining results
2. CB characteristics in C approach
3. C characteristics in CB approach
3. Unifying model that combines both from the ground up
General challenges
3. How does it works?
1. Diversity of recommendations
2. Persistence in recommendations - Re-show recommendations
3. Privacy - what information should be evaluated? Profiling issues
4. User demographics - who likes it?
5. Serpendipity - how surprising is the recommendations?
Direction
4. Now what?
1. General improvements on algorithms
2. Multidimensionality of recommendations
4. Flexibility - Not hardwired or platform spesific
4.5 RQL - Recommendation Query Language
3. Non-intrusiveness