Intro to recommender system

Module 1 – 2

Zhilong Zhu summary note

Everyday decisions

• What to wear?

• What movie to rent?

• What stock to buy?

• What blog post to read?

• What book to read?

can be difficult without prior direct knowledge of the candidates

Recommender system goal is to provides a fine tuned selection experience

How we get recommendations

Historically we rely on

• mentions from your peers

• advise from experts

Limitation on info discovery

• A book a person would enjoy, but no one in the circle heard of it

• A band in another city the experts not knowing

With computer-based system we can

• Expand the set of people/circle

• Mine user’s history and preference

Applications in e-commerce

• Amazon.com – on purchase history

– browsing history

– currently viewing

• Pandora music

• Netflix – Movies

– Netflix Prize: competition on recommendation algorithms to win $1M prize

• Other online retailers

Goals: to increase sale volume

A little vocabulary

• Rating – expression of preference

– Explicit rating (direct from the user)

– Implicit rating (inferred from user activity)

• Prediction – estimate of preference

• Recommendation – selected items for user

• Content – attributes, text, etc.

• Collaborative – using data from other users

Collaborative filtering in MovieLense

rating

ratingPair-wise corr.

request

Find good neighborhood

Collaborative filtering in MovieLense

Dimensions of Analysis

– Domain

– Purpose

– Recommendation Context

– Whose Opinions

– Personalization Level

– Privacy and Trustworthiness

– Interfaces

– Recommendation Algorithms

Domain of recommendation

• 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)

Purpose of recommendation

• The recommendations themselves

– Sales

– Information

• Education of user/customer (e.g. software learning)

• Build a community of users/customers around products or content (e.g. Tripadvisor, referral web as in LinkedIn)

Recommendation context

• What is the User doing at the time of recommendation?

– Shopping

– Listening to Music

– Hanging out with other people

• How does the context constrain the recommender?

– Groups, automatic consumption (vs. suggestion), level of attention, level of interruption?

Personalization Level

• Generic / Non-Personalized

– Everyone receives same recommendations

• Demographic

– Matches a target group

• Ephemeral

– Matches current activity

• Persistent

– Matches long-term interests

Privacy and Trustworthiness

• Who knows what about me?

– Personal information revealed

– Identity

– Deniability of preferences

• Is the recommendation honest?

– Biases built-in by operator• “business rules”

– Vulnerability to external manipulation

– Transparency of “recommenders”; Reputation

Recommendation Algorithmstwo tasks: predict score (how much you’ll like ) and recommend (suggestions for items you might like)

• Non-Personalized Summary Statistics• Content-Based Filtering

– Information Filtering– Knowledge-Based

• Collaborative Filtering– User-User– Item-Item– Dimensionality Reduction

• Others– Critique / Interview Based Recommendations– Hybrid Techniques

Evaluation

To properly understand relative merits of each approach, we will spend significant time on evaluation

– Accuracy of predictions

– Usefulness of recommendations• Correctness

• Non-obviousness

• Diversity

– Computational performance

Non-Personalized recommendation

• Summary statistics – Best-seller; Most popular; Trending Hot

– Zagat, Billboard music rankings, TripAdvisor hotel ratings

– Knows nothing about the person consuming it

– Problem: Averages lack context, diversity issue, self-selection issue

• Product association– Amazon follow up sales.

You brought X, you may be interested in Y.

– Computation: P(X and Y) / P(X) , not symmetric

– Problem: Banana trap or overwhelming effect

– Alternative: P(x and y)

P(x)P(y)

Preference and rating

Implicit: greater volume, hard to interpret

Scoring and ranking

Simple method uses little data

– Predicted score

– Average rating / upvote proportion

– Net upvotes / # of like

– % >= 4 stars (‘positive’)

• Damped mean

– For low confidence with few ratings

• Time decay (Hacker News)

– Old stories aren’t interesting even

• Reddit algorithm