Recoomendation systems Intro

Recommendation SystemsFrom scratch to a working spark

implementation

Recommendation SystemDictionary definition

Set of tools which help in finding products, information, or services in a specific domain when the user has not enough knowledge or

when there are just too many items

It’s a subclass of Information Retrieval that differ in the fact that it’s active and personal

It’s not the technique, it’s an application.It’s the WHAT not the HOW

Few ExamplesIt’s not only Amazon! It’s everywhere!

Take a few seconds to spot all the recommendations in the coming screen

The RecipeNo one solution to fit them all

Problem DefinitionThe basic elements

Item 1 Item 2 Item 3

User 1 User 2 User 3 User 4

Problem DefinitionThe basic elements

MetadataMight be given, or inferred

(images, text, statistics)

ProportionsItem set is usually much

smaller than User set

IncomingNew users & items

appear on a certain rate

LifetimeItems might have a

lifetime

HierarchyUsers and\or Items could be

aggregated together

Problem DefinitionThe recommendation question



Given User u in a given Context c, and given an existing action history,what Item set out of all Items should be suggested to the user?

Problem DefinitionThe ranking matrix

Item 1 Item 2 Item 3 Item 4

User 1 4 - 5 -

User 2 - 1 - -

User 3 - - 3 3

User 4 - - 3 -

User 5 2 - - -

Ranking matrix is usually very (very) sparse

The values inside matrix represent how much the user “liked” an item.But is it that simple?

Problem DefinitionHow to describe the preference value?

Explicit vs. Implicit- Explicit is not always possible, when do, users are lazy - Explicit might be misleading




User 1 5 4 5 -

User 2 - 1 - -

User 3 3 - 3 3

User 4 - - 3 -

User 5 2 - - -


User 1 1 0 1 0

User 2 0 1 0 0

User 3 0 0 1 1

User 4 0 0 1 0

User 5 1 0 0 0

Binary vs. Scale- Non-binary contain more information- More information could be misleading- Non-binary scale is hard to define and usually biased by the user



Binary vs. Scale- Non-binary contain more information- More information could be misleading- Non-binary scale is hard to define and usually biased by the user



Binary vs. Scale- Non-binary contain more information- Non-binary scale is hard to define and usually biased by the user- More information could be misleading

Missing values- Usually a missing value is indication that the user is not familiar with the item- But it could be a negative indication- Not all suggestion were born equal

How to MeasureTell me where you want to go and I will tell you how

Define success measure in a way that is significant to you

And then

Aim to it

How to MeasureHow to describe the preference value


User 1 4 - 5 -

User 2 - 1 - -

User 3 - - 3 3

User 4 - - 3 -

User 5 2 - - -

Given a recommendation to user 3. How will we judge it?

PredictionHowaccuratecanitestimateUser3

preferencetoItem1&2?

RankingHowaccuratecanitestimateifItem1is

“better”thanItem2forUser3

How to MeasurePrediction vs. Ranking

Item # Actual System A System B

1 9 8.4 10

2 8.8 8.9 7.4

3 8.5 8.5 7.5

4 8 7 6

5 6.5 7.1 5

Which recommender system is better?

How to MeasurePrediction vs. Ranking

Item # Actual System A System B

1 9 8.4 10

2 8.8 8.9 7.4

3 8.5 8.5 7.5

4 8 7 6

5 6.5 7.1 5

Which recommender system is better?

RMSESystem A = 0.588System B = 1.430

nDCGSystem A = 0.990 System B = 0.996

How to MeasureNow let’s go to the soft side

10% Coffee & PastryArcaffe

10% Coffee & PastryAroma

10% Coffee & PastryIlan’s

How to MeasureGive me something different

10% Coffee & PastryArcaffe

10% Coffee & PastryAroma

10% Coffee & PastryIlan’s

Diversity


Diversity


Diversity Novelty


You must join the “See it to believe!” meetup group

Diversity Novelty


You should join the “See it to believe!” meetup group

Diversity TrustNovelty


Diversity TrustNovelty


Diversity TrustNovelty Utility


System A System BVs.

0123456789

Accuracy

Diversity

TrustNovelty

Utility

0123456789

Accuracy

Diversity

TrustNovelty

Utility

How to MeasureHow to conduct an experiment

The common offline Train\Test method


User 1 4 - 5 -

User 2 - 1 - -

User 3 - - 3 3

User 4 - - 3 -

User 5 2 - - -

User Item Preference

1 1 4

1 3 5

2 2 1

3 3 3

3 4 3

4 3 3

5 1 2

Train

Test


User 1 4 - 5 -

User 2 - 1 - -

User 3 - - 3 3

The “Magic”Different modeling approaches

Collaborative-Filtering Content-Based Knowledge-Based

Uses only history preferences to imply

on unknown preferences

Try to find patterns between preferences

and different properties

Ingestion of domain specific knowledge

The “Magic”Collaborative Filtering “similarity” methods

Find similar entities using their previous preferences





User based Item based

The “Magic”Collaborative Filtering matrix factorization methods



Concept 1 Concept 2




Concept 1 Concept 2

U*K




Concept 1 Concept 2

K*I


2 3

1 4

2 2

2 5

3 4

2 5 1

2 4 2

U*K K*I U*I

2 - -

- 4 -

2 - 5

- 5 -

3 - 1


2 3

1 4

2 2

2 5

3 4

2 5 1

2 4 2

U*K K*I U*I

2 - -

10 4 -

2 - 7

- 5 -

3 - 1

Prediction of User 2 preference for Item 1

UserId Feature 1 Feature 2 Feature 3

…

…

ItemId Feature 1 Feature 2 Feature 3 Context 1 Context 2 Context 3 Preference

The “Magic”Content based methods

Similar to any standard supervised learning process

The “Magic”Methods comparison

Simplicity Novelty Cold start

Accuracy Grey sheep

The “Magic”Hybrid is usually the best

The “Magic”Hybrid is usually the best

Mix Average

Flow Rules

Architecture Considerations

Real time

Heavy lifting

Adaptivity

Collect context

Running ML processes

Respond to the user

Adapting to real time behavior

Collect additional sources and maintain user integrity

Serving layer

Architecture considerationsLambda architecture

New data

Batch layer

Speed layer

Analytical storage

Operative storage

QueueStream

processing

Stream processing

App

Architecture considerationsMicro-service architecture

2 - -

- 4 -

2 - 5

- 5 -

3 - 1

Relevant TechnologiesGraph databases vs. General purpose

SparkIntro

Originally developed in

2009 at UC Berkley AMP

Lab

Open sourced in 2010

As of 2014, Spark is a top-level

Apache project

Top open-source community in

2015

SparkIntro

SparkMLlib

SparkALS in MLlib vs.Mahout

MLlib ALSAlternating-Least-Squares

- -

- -

- -

- -

- -

- - -

- - -

U*K K*I U*I

2 - -

10 4 -

2 - 7

- 5 -

3 - 1


- -

- -

- -

- -

- -

- - -

- - -

U*K K*I U*I

2 - -

10 4 -

2 - 7

- 5 -

3 - 1

Fix one matrix solve the second


5 1

2 3

2 4

2 1

3 4

- - -

- - -

U*K K*I U*I

2 - -

10 4 -

2 - 7

- 5 -

3 - 1



5 1

2 3

2 4

2 1

3 4

2 3 1

1 1 2

U*K K*I U*I

2 - -

10 4 -

2 - 7

- 5 -

3 - 1


TakeawayNo one solution to fit them all

Additional MaterialsYour starting point

www.myyellowroad.com

Advanced Analytics

Big Data Architecture

Turnkey Analytics Projects

Big Data Analytics Methodology

From Business Need, With Big Data Analytics, To Business Success

Problem definition

Data modeling

How to measure

The “magic”Tech challenges

Architecture considerations

Relevant technologies Implementation Evaluation

Recoomendation systems Intro

Data & Analytics

Transcript of Recoomendation systems Intro