• General overview
• Algorithms
• Evaluation
• Problems
• Existing tools
• Books and Online courses

1. Domain
2. Purpose
3. Recommendation Context
4. Whose Opinions
5. Personalization Level
6. Privacy and Trustworthiness
   
7. Interfaces

Content to Commerce and Beyond

• News, information, “text”
• Products, vendors, bundles
• Matchmaking (other people)
• Sequences (e.g., music playlists)

One particularly interesting property

• New items (e.g., movies, books, …)
  
• Re-recommend old ones (e.g., groceries, music)

• The recommendations themselves (Sales, Information)
• Education of user/customer
• Build a community of users/customers around products or content

What is the User doing at the time of recommendation?

• Shopping
  
• Listening to Music
  
• Hanging out with other people
  

• “Experts”
  
• Ordinary “phoaks”
  
• People like you

• Generic / Non-Personalized: everyone receives same recommendations
• Demographic: matches a target group
  
• Ephemeral: matches current activity
• Persistent: matches long-term interests

Who knows what about me?

• Personal information revealed
  
• Identity
  

Is the recommendation honest?

• Biases built-in by operator
• Vulnerability to external manipulation
• Transparency of “recommenders”

Types of Output

• predictions (Estimates of how much you’ll like an item)
• recommnedations (Recommendations are suggestions for items you might like (or might fit what you’re doing)
• filtering
• Organic vs. explicit presentation (Agent/Discussion Interface)

Types of Input

• explicit (Rating, Review, Vote, Like)
• implicit (Click, Purchase, Follow)

1. Non-Personalized Summary Statistics

2. Content-Based Filtering

• Information Filtering
• Knowledge-Based
  

3. Collaborative Filtering

• User-user
• Item-item
• Dimensionality Reduction

4. Others

• Critique / Interview Based Recommendations
  
• Hybrid Techniques

• Best-seller
• Most popular
• Trending Hot
• Best-liked
• People who X also Y

• User Ratings x Item Attributes => Model
• Model applied to new items via attributes
• Alternative: knowledge-based (Item attributes form model of item space)
• Example: Personalized news feeds

Key ideas:

• Model items according to relevant attributes
• Model or reveal user preferences by attribute

Key concept: Keyword Vector

• The universe of possible keywords defines a content space

• Each keyword is a dimension

• Each item has a position in that space; that position defines a vector

• Each user has a taste profile (or more than one) that is also a vector in that space - - The match between user preference and items is measured by how closely the two vectors align

• May want to limit/collapse keyword space (e.g., stem and stop)

• Apache Lucene (Solr, PyLucene, Lucy, Lucene.Net)
• LensKit
• Machine Learning tools (Mahout, Python (with SciPi + NumPy), Matlab/Ocatve, R .. )

• Use opinions of others to predict/recommend
• User model – set of ratings
• Item model – set of ratings
• Common core: sparse matrix of ratings

Collaborative Filtering Techniques

• User-user (Select neighborhood of similar-taste people and use their opinion)
• Item-item (Pre-compute similarity among items via ratings)
• Dimensionality reduction (Compress and use a taste representation)

Selecting Neighborhoods

• between 20 and 100 is often used

Scoring Items from Neighborhoods

• average
• weighted average
• multiple linear regression

Normalizing Data

• mean centering
• z-score

Computing Similarities

• Pearson correlation
• Cosine of user vectors in rating space

• Explaining Collaborative Filtering Recommendations by Herlocker, Konstan, and Riedl (Proc. CSCW 2000)

Key lessons

• Simplicity is key – users didn’t like “correlation,” statistical terms, or being overwhelmed with data
• Simple visualizations work well – histogram, table
• Supporting information valued – historical success, attribute-linked data, associations

One key mistake:

• We didn’t actually measure explanation effectiveness, but rather persuasiveness

Pre-compute item similarities over all pairs of items

• Usually use cosine similarity between item rating vectors
• Often normalize user ratings first (subtract user mean, subtract item mean)

Look for items similar to those the user likes (purchased, in their basket)

• For each item to score:
• Find similar items the user has rated
• Compute weighted average of user's ratings

Computing Similarities

• Cosine similarity still works
• Can also use conditional probability

Aggregating Scores

• Weighted average works for non-binary
• For binary (0/1), just sum neighbor similarities

Ratings matrix is an overfit representation of user tastes and item descriptions

• Leads to problems of synonymy
• Also leads to computational complexity, potentially poorer results
• Ideal would be to have a more compact representation of user tastes and item descriptions – but how ?

SVD

• The reconstructed matrix R = [U.S.V’] is the closest rank-k matrix to the original matrix R.

Challenge #1: missing values

• ‘Impute’ — assume they are a mean
• Normalize data first — assume they are 0
• ignore them

Challenge #2: computational complexity

• FuncSVD
• SVD++

Challenge #3: lack of transparency / explainability

• Weighting – combine algorithm scores (Can be extended with feature-weighted coefficients)
• Switching – switch algorithms
• Mixed – mix output from diff. algorithms
• Use one algorithm as input to another

Accuracy metrics

• MAE, MSE, RMSE

Decision-support metrics

• ROC AUC, precision/recall

Rank metrics

• Spearman Rank Correlation, Discounted Cumulative Gain, Fraction of Concordant Pairs, Mean Reciprocal Rank

User and Usage-centered metrics

• coverage, user retention, recommendation uptake, satisfaction

More metrics!

• Coverage, Diversity, Serendipity

A Commercial Look

• Lift, cross-sales, up-sales, conversions

• Threats
• 'Cold start' problem

• LensKit (Java)
• Apache Mahout (Java)
• MyMediaLight (C#/.NET)
• GraphLab (C#, Java)
• REST services (EasyRec, PredictionIO)
• crab (Python)
• mrec (Python)
• recommenderlab (R)

Online resources:

• Coursera: Introduction to Recommender Systems https://www.coursera.org/course/recsys

• http://recsys.cs.umn.edu/readings.html

• Recommender System Wiki http://www.recsyswiki.com/wiki/Main_Page

• http://habrahabr.ru/search/?q=Surfingbird

Books:

• Handbook of Recommender Systems (Ricci et al., Eds.) 2011, 2014?
• http://recommenderbook.net/

Articles:

• ACM Conference on Recommender Systems
• ACM Trans. on Information Systems
• User Modeling and User-Adapted Interaction
• FunkSVD http://sifter.org/~simon/journal/20061211.html