Introduction to Recommender systems

1. Recommender Systems
Vivek Murugesan
10th Jun 2017

2. Agenda
● Recommendations…
● Advantages
● Recommender systems
● Anatomy of recommender systems
● Type of recommender systems or algorithms
● Requirements for scaling
● Learning to rank model
● Where to start…?
● Questions

3. Recommendations...
● It is estimated that close to 30% of Amazon’s revenue comes from the way
they integrated recommendations
● About 75% of Netflix’s business is driven through recommendations
● Posts, groups, people and jobs recommended through Linkedin
● Posts, friends suggested in Facebook
● Google search results, autocomplete suggestions, google news, etc,.
● Stories recommended through Quora
● ...

4. Advantages...
● Increase customer engagement
● Increase customer satisfaction by delivering
relevant contents
● Increase sales with cross sell options
● Drive more traffic
Bought
Similar
Recommend

5. Recommender systems
● Attempts to predict the preference of a given user for an item
● Based on the prediction recommends set of items for the user
● By using,
○ Data: User history, Social profile, etc,.
○ Algorithms: Like Collaborative filtering, content based filtering etc,.
● Delivers the recommendations and let the users interact with them
● Attempts to learn more about the users

6. Anatomy of recommender systems
Item inventory
User details
U
S
E
R
S
I
T
E
M
S
Algorithms and
models
Business
objective
Recommendations
U
S
E
R
S
Deliver
Build
Feedback loop

7. User preference data
● Data that the recommender system rely on to know/learn about the users
● The same being used for personalizing recommendations for the users
● Explicit,
○ Rating
○ Voting
○ Opinion
● Implicit,
○ Click
○ Purchase
○ Follow
● Binary
○ Like/Dislike
○ Click/Ignore
● Rating
○ Scale
● Unary
○ Purchase
○ Views
● Normalize rating
○ Handling scales
○ User biases
● Unary data
○ Special
processing

8. Type of recommender systems or algorithms
● Non-personalized
○ Popular and trending product
○ Based on simple summary
statistics
○ To handle cold start scenario in
some cases
● Semi personalized
○ Association rule mining
○ Market basket recommendation
○ Ephemeral (contextual)
● Personalized
○ Persistent (long term interests)
○ Content based
○ Collaborative filtering
■ User-user similarity
■ Item-item similarity
■ Latent attributes through
SVD
● Advanced
○ Hybrid (contextual + interest and
content + collaborative)
○ Learning to rank

9. Content based filtering
1. Associate each item with certain keywords or attributes
2. Build a vector of keyword preference for user, by inferring based on
their item preferences
3. Use TFIDF or similar mechanism to accumulate preferences on
keywords across items
4. Score each item based on the cosine similarity (i.e. a dot product
between two vectors) of its keyword vector with the user’s keyword
vector

10. Association rule mining
● Attempts to identify rules like people “who bought X also bought Y”
● Based on the items bought together in a transaction or a time window
● Rules of the form X → Y are discovered/mined from the data
● Where X is called antecedent and Y is called the consequent
● Metrics associated with the rules,
○ Support = N(X U Y) / |T| ( P(X AND Y) Ratio of transactions in which X and Y are bought
together)
○ Confidence = Support(X U Y) / Support(X) ( P(Y|X) Percentage of buyers of X, who also
bought Y)
○ Lift = Support(X U Y) / [Support(X) * Support(Y)] ( P(X AND Y) / [P(X) * P(Y)] )
● Some popular product tend to be appearing part of the consequent of all rules
● Lift is the metric that can help to get away with the issue

11. Collaborative filtering
● Unlike content based filtering collaborative filtering doesn’t assume the
presence of attributes or keywords about the items
● Makes use of the User x Item matrix computed based on the history
● Generates recommendations entirely based on this this matrix
● Doesn’t rely on any additional details about the items or users (like
demography)

12. User-user similarity
1. Build neighbourhood for each user based on the User x Item matrix (by using
the correlation or cosine similarity of each user with others)
2. Use the likes/interactions of the top k users to build a potential set of items to
recommend for the user a.
3. Score each item in the potential set based on the preference of the user u and
their similarity with user a.

13. Item-item similarity
1. Build an item neighbourhood based on the preferences expressed by different
users (i.e. based on the rating vectors of two different items)
2. Use set of items that the user has expressed preferences on up front to
generate potential items as recommendations for them
3. Score each item in the potential set based on their similarity with the item that
the user has liked (expressed preference) earlier

14. Requirements for scaling
● Volume of item base:-
Depending on the item
inventory, it can sometimes
turn out to be huge number
● Volume of user base:- Similar
to item base, users can also be
very large in number
● Delivery channel/mode:-
Depending on the mode of
choice delivery, either turn
around time or response time
needs to be focused
● User feedback/interactions:- It is
an opportunity to learn more
about the users, handling the
streaming data of interactions
may pose some challenges
● Incremental update of
models/algorithms:- Interactions
means models need to be
updated in near real time.
Capability of incremental update
is critical to avoid whole bunch of
recomputation.

15. Scaling...
With n items and m users calculating
similarities of a given user with all other
m-1 users take O(mn) time.
When we are talking about performing this
across all users it takes O(m2
n) time, as
there will be np
2 pairs or nc
2
combinations.
Similarly for calculating similarities across
all items it takes O(n2
m) time.
But we need only top k neighbourhood for each
user or item.
When n>>k then it results in lot of unnecessary
computations. Is there a way to avoid these…?
The item similarity API from Spark ML provides
an answer to this.
Clustering:- Let’s say we are trying the
compute the user similarities among m
users, where is really a large number. User
neighbourhood can then be computed only
within the cluster of users.
Dimensionality reduction:- A User x Item matrix
of dimension mxn, can be reduced to mxk by
generating top k principal components from the
matrix. Similarity computation on top of this
reduced matrix will be faster than using the
original matrix.

16. Learning to rank
● Several algorithms and models,
○ Can generate predicted rating items by user
○ Can also generated rank ordered list of
recommendations
○ By consuming user history
● Various indicators of items like trend,
seasonality, etc,.
● User’s preference, context, etc,.
● Business promotional objectives
● Item catalog/inventory coverage
● Eventual rank ordering by combining all of
these
○ To satisfy user’s preference
○ And business objectives

17. Where to start…?
● Open source datasets
○ Movie lens
○ Million song
● Open source framework and API
○ Spark ML
○ Movie lens
● References
○ Recommender systems survey
○ Applying SVD/PCA on recommender systems
○ Learning to rank model
○ Coursera:- Recommender system
● Github
○ https://github.com/vivekmurugesan/recommender-systems

18. Questions…?
Email: vivek.murugesan@gmail.com
Linkedin: https://www.linkedin.com/in/vivek-murugesan/