Machine Learning Project Recommender System
Ryan D. Moniz
Clear Goal
• Recommender System for Music.
• e.g. similar to Amazon’s Recommendation System
• similar to iTunes’ Recommendation system
Amazon’s Collaborative Filtering System
iTunes’ Recommender System
Non-Trivial Problem
• Compared with others who bought the same CD or song as you.
Non-Trivial Problem
• Your entire music library compared with everyone who bought the same CD or song as you.
Attributes
• Artist
• Album
• Beats Per Minute
• Genre
• Year
Learning Method
• ART1 - Adaptive Resonance Theory
• Applied to dynamically self-organizing data.
• Unsupervised clustering neural network.
• Incremental but stable clusters.
Learning Method
• Step 1 - Initialization
• Initialize N to the total number of clusters
• Initialize the vigilance parameter p so 0 < p ≤ 1.
• Let p = (0 ∪ 1)* ∀i ∈ [1,N]i
Learning Method
• Step 2 - Apply new input vector
• Let I = [next input vector]
• Let P’ = P be the set of candidate prototype vectors
Learning Method
• Step 3 - Find the closest prototype vector from P’
• Find the i which maximizes
β +
|| p ∪ I ||
||p ||i
i
|| p ∩ I ||
|| I ||< p then P’ = P’ -
Learning Method
• Step 4
• Check if is too far from I
• If
• If P’ is empty goto step 2 otherwise goto step 3.
pi
pi
Learning Method
• Step 5 - Update the matched prototype vector
• Let = ∩ I ; output i; goto step 2.
• β acts a tie-breaker favoring the larger magnitude prototype vectors.
pi
pi
Learning Method
Learning Method
Principle Way to Evaluate
• Recommend relevant songs, albums, artists, genres (better prediction).
• e.g. if you like classical music and have no heavy metal in your collection you probably don’t prefer heavy metal.
≠
Progress
• Algorithm is “almost” implemented.
• Parameter tweaking will need to be done from time to time. (β, p, and d).
• Need to find a way to add lots of data, e.g. Music libraries/databases.
Risk
• Finding enough data for training & evaluation.
• Time to complete project.