Introduction to Collaborative Filtering For Recommender Systems

● Farouq Oyebiyi ● Machine Learning Dude at Konga ● Focused on recommendations and personalizations ● Working on product recommendations at Konga Who am I?

Recommender Systems (RecSys)? ● Information systems that predicts a user’s preference for an item or list of items ● Use cases: ○ Product recommendations on Amazon ○ Music recommendations on Spotify ○ Book recommendations on Goodreads ○ Job recommendations on Jobberman? ○ Hotel recommendations on Hotels.ng? ○ Friend recommendations on Facebook

Why do we need RecSys?

Why do we need RecSys? ● Lots more information than we have time to go through ● Finding the perfect movie/product is a “needle in the haystack” problem 30M songs 200M items

How to build a RecSys? ● Approaches: ○ Content-based filtering ○ Collaborative filtering ○ Hybrid

Collaborative Filtering (CF) ● Predict user preference based on behaviour ● Behaviour includes: ○ Purchase history - Implicit feedback ○ Listening history - implicit feedback ○ Likes - Implicit feedback ○ Shares - Implicit Feedback ○ Review/rating - Explicit feedback

Formulating the CF problem ● Let U be the set of all users ● Let V be the set of all items ● R is a U by V matrix Goal: Predict the value of the empty cells. Super Story Jennifer’s Diaries Papa Ajasco Saworoide Gibran ? 1 ? ? Adichie ? ? 1 ? Fajuyi 1 ? 1 ?

Collaborative Filtering Techniques ● Techniques ○ Memory-based; nearest neighbour, Pearson’s Correlation ○ Model-based; Matrix Factorization

● Given a matrix P, find 2 matrices (X, Y) whose dot product will give you R ● Matrix X and Y will have dimension I which is specified by the user or determined via cross validation ● I represents the number of latent factors in the each matrix Matrix Factorization (MF)

MF Equation C ui - confidence level that user u likes item i P ui - binary value indicating if user u has interacted with item i X u - latent vectors for user u Y i - latent vectors for item i - regularization parameter

Evaluating MF ● Root Mean Squared Error (RMSE) ● Precision/Recall ● Mean Average Precision at k (MAP@k) ● Normalized Discounted Cumulative Gain (NDCG) Ground truth (y) = [1, 0.5, 1, 0.75, 1, 0.2] Predicted rating ( ) = [0.8, 0.45, 0.2, 0.5, 1, 0.3] RMSE =

Advantages of CF ● Does not require side information to make recommendations ● Very good at discovery; serendipity ● Can capture subtle preferences

Disadvantages of CF ● Cold start problem ● Accuracy is low when there’s limited data

Matrix Factorization Code Sample http://spark.apache.org/docs/latest/mllib-collaborative-filtering.html#collaborative-filtering

Value of RecSys

Value of RecSys cont’d ● Engage users ● Increase conversion ● Better UX

Resources ● http://www.slideshare.net/MrChrisJohnson/algorithmic-music-recommendation s-at-spotify ● http://www.slideshare.net/xamat/recommender-systems-machine-learning-sum mer-school-2014-cmu ● http://spark.apache.org/docs/latest/mllib-collaborative-filtering.html

I recommend you ask questions. Thank you