Week 1 - Introducing Recommender Systems

Mar 04, 2017 ../recommender-systems-introduction

Intro to Course and Specialization

  • Broken into 4 courses:
    • Non-personalized and content-based.
    • Nearest-neighbor collaborativ filtering.
    • Evaluation and metrics.
    • Matrix factorization and advanced techniques.
  • Capstone project:
    • Case study analysis - design best recommender for a business use case.

Predictions and Recommenders

  • Recommendations have less authority than predictions: "maybe you'll like this", not "I'm sure you'll like this".
  • Predictions:
    • Pro: help quantify item.
    • Con: could provide something falsifiable.
  • Recommendations:
    • Prod: could provide good choice as default.
    • Con: if peeps think it's "top-n", could stop them exploring.
  • Recommenders could be considered "softer sell" eg book stores have some books facing forward, others just spine. TVs in more prominent positions.

Taxonomy of Recommenders

  • Domain of recommendation: "what's being recommended?"
    • News articles
    • Products
    • Matchmaking
    • Sequences (musical playlists)
  • Interesting property:
    • Is it new items (movies, books)?
    • Re-recommend old ones (groceries, music)?
  • Examples of recommenders:
    • Google search results.
  • Purposes of recommdations
    • Sales.
    • Education of user/customer.
      • Tip of the day in Office products.
    • Building community around products or content.
  • Recommendation context
    • What's the user doing when making recommendation?
      • Shopping
      • Listening to music
    • How does the context constrain the recommender?
      • Groups, automatic consumption, level of attention?
  • Whose opinion?
    • Recommenders are usually based on somebodies opinion:
      • Experts, other users, etc.
    • "Phoaks" (People helping one another know stuff)
  • Personalization level
    • Generic / non-personalized
      • Same recs for all.
    • Demographic
      • Women get different recs then men etc.
    • Ephemeral
      • Match what you're currently doing.
    • Persistent
      • Interests over time.
  • Privacy and trustworthiness
    • Who knows what about me?
      • Personal info reveal.
      • Identity.
      • Deniability of pref
    • Is the rec honest?
      • Biases built-in by operator ("business rules")
      • Vulnerability to external manupulation
        • Example: higher scores for new movies. Are movie studies "hacking" the results?
      • Transparency of "recommenders": reputation
  • Interfaces
    • Types of output
      • Predictions
      • Recommendations
      • Filtering
      • Organic vs explicit presentation
    • Types of input
      • Explicit
        • Being asked to review things.
      • Implicity
        • How often have you looked at a certain page?
  • Recommendation algorithms
    • Non-personalized summary stats
    • Content-based filtering
      • Info filtering
      • Knowledge-based
    • Collaborative filtering
      • User-user
      • Item-item
      • Dimensionality reduction
    • Other
      • Critique / interview based recs.
      • Hybrid techniques.

Taxonomy of Recommenders 2

  • Notions every recommender needs:
  • Users
    • Users may have attributes (demographics).
  • Items
  • Ratings
    • Users make rating on items somehow (implicity and explicit).
  • Non-personalized summary stats:
    • External community data:
      • Best selling, most popular, trending stuff.
    • Summary of community ratings:
      • Best liked.
      • Pull out ratings for some item and take average.
    • Examples:
      • Zagat restaurant ratings
      • Billbard music rankings.
      • TripAdvisor hotel ratings.
  • Content-based filtering:
    • User ratings x item attributes => model
    • Model applied to new items via attributes
      • User liked articles about soccer, future articles about soccer may be recommender.
      • Fan of certain genres of movies.
      • Fan of movies with certain actors.
    • Alternative: knowledge-based
      • Item attributes form model of item space.
        • Users navigate/browse that space.
    • Examples:
      • Personalized news feeds.
      • Artist or genre music feeds.
  • Personalized Collaborative Filtering
    • Use opinions of others to predict/recommend.
    • User model - set of ratings
    • Item model - set of ratings
    • Common core: sparse matrix of ratings
      • Fill in missing values (predict)
      • Select promising cells (recommend)
    • Several different techniques.
  • Collaborative Filtering Techniques
    • User-user
      • Get "neighbourhood" of people with similar tasts.
        • Could only select "trustworthy" people.
    • Item-item
      • Compute similarity amongst items using ratings.
      • Use ratings to triangulate for recs.
    • Dimensionality reduction
      • Intuition: taste yields a lower-dim matrix.
      • Compress and use taste representation.
  • Note on evalution:
    • Will spend time on evaulation:
      • Accuracy of predictions.
      • Usefulnes of recommendations: correctness, non-obviousness, diversity.
    • Computational performance.

Tour of Amazon

  • Dimesions of analysis:
    • Recommendations based on implicit purchase data.
    • Personalization level: one product at a time.

Recommender Systems: Past, Present and Future

  • Before recommender systems:
    • Manual personalization
    • Cross-sales and early product associations
    • Product search.
  • Tech bubble:
    • During: recommenders were seen as "key technology"
    • After: recommenders were put in context of the things the business was actually trying to do.
  • Wave 2: The Netflix Prize
    • Netflix $1M prize
    • Recommendation as app area for data mining, machine learning
    • Rapid growth in field
    • New techniques:
      • Algorithm stacking
      • New matrix factorization techniques
  • Mature realizations:
    • Predictions and top-n algos are limited
      • Limitations to how good recommendation engines are with people driving them.
  • State of field today:
    • Lots of well-known algos
    • Effective recs still a "craft"
      • Exploring data.
      • Understanding usage cases and value proposition
    • Still largely focuses on business apps
      • Creativity
      • Dream of consumer-owner not realized.
  • Looking forward:
    • Hard problems unsolved:
      • Temporal recommendations: "what should you consume next?"
      • Recs for education.
      • Low-frequency, high-stakes recs: can we help you find a house or other things you don't have rating for?
    • Recognized speciality that brings ML, business + marketing, human-computer interaction etc.
  • Promising directions:
    • Context.
    • Sequences: music, education etc.
    • Lifetime value.

Introducing the Honors Track

  • Involves implementing algorithms using LensKit toolkit.
  • Software required:
    • Java dev kit.
    • Dev environment (Eclipse, IntelliJ etc)..
    • Data analysis software: Excel, R PyData etc.
  • LensKit handles external stuff like I/O, setting up data for evaluation etc.
  • Written in Java because a) lots of people know it and b) can achieve good performance.