Scribd
Upload
16 views

1 Introduction - Recommender Systems

Uploaded by

whydelete325132
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
16 views

1 Introduction - Recommender Systems

Uploaded by

whydelete325132
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 24

Recommender systems

Alaa BAKHTI
The course Syllabus
- 5 sessions - 21 hours
● Session 1: Introduction to recommender systems
- Each session = 2h theory + 2h
practical work ● Session 2: Content-based filtering
- Grading
- Research paper presentation
● Session 3: Memory-based collaborative filtering

- Implem 2 tower model * ● Session 4: Model-based collaborative filtering


- MCQ
- Bonus ● Session 5: Evaluation of recommender systems and
- Participation
Application of deep learning in this domain.
Research paper presentation

- Subject: choose one paper from this list and post it the course teams
channel (first come, first serve)
- What: 25 min presentation + 5 min Q&A
- Who: 2 students per group
- When: last session
2 tower model implementation

- Subject: implement a 2 tower model (user, movie) on the movielens


dataset using tensorflow
- What: notebook with the implementation
- Who: 2 students per group
- When: last session
Do you know?

TF IDF
Embedding space
word2vec
Cosine similarity
SVD
NMF
EDA
Loss function
L2 regularization
Dropout
Dense layer
AB testing?
Introduction
The long tail problem

Popularity Popular items

The long tail

Items

Picture by Hay Kranen / PD


Information overload
The paradox of choice

Source: Mark Rowland - Your marketing rules

The paradox of choice by Barry Schwartz (book review) - Youtube video


Recommender systems

➔ Help users find compelling content in a


large corpora.
➔ Reduce information overload by
estimating relevance.
➔ Personalise the user experience.
Applications Where Recommender systems are
used?
and business
value - E-commerce websites
- Search engines
Many domains where the - Social networks
recommender systems can be - Movie or music streaming sites
used - mobile app stores
- etc
Netflix: movie recommendation

75%
of the watched content is from
some sort of recommendation

Source: Netflix recommendations: Beyond the 5 stars - X. Amatriain and J. Basilico - Netflix Inc - 2012.
Netflix: movie recommendation

$1B
per year is the estimated business
value of recommendation

Source: The Netflix recommender system: Algorithms, business value, and innovation - C. A. Gomez-Uribe and N. Hunt - Netflix Inc - 2015.
Amazon: product recommendation

35%
of Amazon sales originate from
cross-sales (recommendation)

Source: How retailers can keep up with consumers - Mckinsey & Company - 2013.
Youtube: video recommendation

60%
of the clicks on the home screen
are on the recommendations

Source: The YouTube Video Recommendation System - J. Davidson & al - Google Inc - 2010.
Problem formulation

Listened to

Recommend from
How to determine items that the user may be interested in?

Listened to

Rated

Recommend from
Rating matrix

4/5 2/5 ? ? ? ?

How to determine the user rating of items he didn’t explicitly rate?


User interactions feedback

Explicit Implicit

- Data provided by users intentionally. - Data generated based on the user interaction
with items (easier to collect).
- Example : Press the like button on a
- Example : purchased an item => high rating.
YouTube video.
- Problem : poorly learns low ratings (what the
- Problem : it requires effort from the user
=> doesn’t scale. user doesn’t like).

Recommender systems use the combination of explicit and implicit user


feedbacks.
How to determine the user ratings for items
he interacted with?

How to extrapolate the user ratings for


items he didn’t interact with?
Recommender system
architecture
Candidate generation : generate a
small subset of candidates from a
huge corpus.

Scoring : score and rank the


candidates in order to select the set of
items to display to the user.

Re-ranking : re-rank the items


depending on additional constraints
(boost the score of fresh content for
example). This step help ensure
diversity, freshness, and fairness.

Deep Neural Networks for YouTube Recommendations - Paul Covington et al - 2016


Practical work

Subject
● Exploratory Data Analysis (EDA) on the movielens dataset (ml-latest-small.zip) (another more complex dataset the
Movies dataset)
● Final dataset will be used in the next sessions
● Use only the files ratings.csv and movies.csv

TODO
● Create a Git repository for the recsys class
● Create a virtual environment “recsys”
● Download the data and store it in the “data” folder
● Create a notebook “movielens-eda.ipynb” to prepare and clean up the dataset (remove missing values, duplicates, check
distributions, …)

You might also like