SlideShare a Scribd company logo

Introduction to recommender systems

Andrea Gigli, profile picture
Andrea Gigli

This document provides an introduction to recommender systems. It discusses how recommender systems can help users filter through large amounts of information and options in an era of information overload. It describes different types of recommender systems, including content-based, collaborative filtering, and context-based recommender systems. The document also discusses challenges like sparsity in data and scaling to large datasets, and how modeling approaches can help address these challenges.

Business
1 of 87
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
Introduction to Recommender Systems Andrea Gigli, Boston Consulting Group & University of Florence @andrgig andrgig@gmail.com
Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge
Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Filtering Information
Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Processing Information
Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Processing Information Information Filtering and Extraction
How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find
How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice
How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice ● Recommender Systems to discover something you didn’t know existed
How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice ● Recommender Systems to discover something you didn’t know existed ○ Items are recommended one the basis of the user (and context) ○ Items means something domain-specific ■ Movies: movie, actor, director, genre ■ Music: track, album, artist, playlist, … ■ E-Commerce: item, bundle of items, … ■ Banking: banking and financial products, investment advice, service bundle, …
“We are leaving the age of information and entering the age of recommendation.” Chris Anderson, 2004, The Long Tail
What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user.
What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user. Recommender Systems are designed to help - Users in doing the best choice for them - Companies in selling more products and increasing customers loyalty
What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user. Recommender Systems are designed to help - Users in doing the best choice for them - Companies in selling more products and increasing customers loyalty Users != “someone to charge” Items != “something to sell to users”
When a Recommender System is good? - It is relevant to the user (personalization) - It represents all the possible interests of one user (diversity) - Expands the user's taste into neighboring areas (serendipity)
Type of Recommender System ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
Type of Recommender System ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based Which is the best one? Depends on the domain and particular problem. Usually combining different approaches works better
Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions.
Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content
Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content - Second, define a user profile by analysing the relevant content of the items the user liked in the past
Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content - Second, define a user profile by analysing the relevant content of the items the user liked in the past - Third, recommended items for a user’s profile built on the second step
Content - Based Recommenders Item profile User profile f1 f2 f3 fN wf1 wf2 wf3 wfN a set of features (characteristics/attributes) characterizing item s. a set of weights, where wfi denotes the importance of feature fi to the user An item is recommended to the user accordingly to the features held by the item which are mostly relevant to the user
Content-Based Recommenders pros ● No need for data on other users. ● Able to recommend to users with unique tastes. ● Able to recommend new and unpopular items. ● Can provide explanations of recommended items by listing content-features that caused an item to be recommended
Content-Based Recommenders cons ● Requires content that can be encoded as meaningful features. ● Users’ tastes must be represented as a learnable function of these content features. ● Hard to exploit quality judgements of other users. ● Easy to overfit
Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
Collaborative Filtering - What is it? The task of predicting (filtering) specific user preferences on new items by collecting taste information from many users (collaborative). Collaborative filtering pros: ● Requires minimal domain knowledge ● Users and Items have no internal structure or characteristics ● Perform well in most cases
Collaborative Filtering - What is it? The task of predicting (filtering) specific user preferences on new items by collecting taste information from many users (collaborative). Collaborative filtering cons: ● Prior behavior determines current behavior ● Many items, few data per user, no data for new user ● Very large datasets
Collaborative Filtering Sparsity Problem In 1,000,000 items, the probability that two users who bought 100 items each, have an item in common is 1% In 10,000,000 items, the probability that two users who bought 50 items each, have an item in common is 0.02%
Collaborative Filtering - How does it work? Let’s represent the information on past users’ choices and choose items on a matrix. Each cell contains the preference of user u for item i Preference can be explicit (eg rating, stars) or implicit (bought/not bought, liked/not liked) U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Collaborative Filtering - How does it work? Let’s represent the information on past users’ choices and choose items on a matrix. Each cell contains the preference of user u for item i Preference can be explicit (eg rating, stars) or implicit (bought/not bought, liked/not lied) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 3. Compute similarities between user and similar users 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 3. Compute similarities between user and similar users 4. Predict user’s preference for each Item weighting for the similarity with other users 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 3. Compute the similarity between the target item and the neighborhood items 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 3. Compute the similarity between the target item and the neighborhood items 4. Predict ratings for the target item for each user weighting for the similarity with other items 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Tackling the Sparsity and Scalability
Tackling the Sparsity and Scalability Sparsity: - Poor relationship among like minded but sparse-rating users. - Accuracy can be poor if one try to make predictions in sparse dataset using nearest neighbor algorithms Scalability: - Nearest neighbor require computation that grows with both the number of users and the number of items. Possible solution: - Using latent factor models to capture similarity between users & items in a reduced dimensional space.
Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 F1 F2 F1 F2
Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 How much latent factor 1 and 2 are contained in Item 2 How much User 3 likes latent factors 1 and 2 F1 F2 F1 F2
Collaborative Filtering via Matrix Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 How much latent factor 1 and 2 are contained in I2 How much U3 likes latent factors 1 and 2 The dot product indicates how much U3 likes I2 F1 F2 F1 F2
Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
Clustering Another way to make recommendations based on past purchases is through Customers clustering → Each cluster will be assigned typical preferences, based on preferences of customers who belong to the cluster → Customers within each cluster will receive recommendations computed at the cluster level PROS CONS Can be used to work on aggregated Data Can be applied as a first step for filtering techniques Can be used to capture latent similarities between users or items Recommendations per customer's cluster are usually less relevant for single user than those from collaborative filtering
Classifiers Classifiers are general computational models trained using positive and negative examples (Logistic Regression, Bayesian Networks, Support Vector Machines, Decision Trees, etc…) They may take in inputs vector of item features (type: action / adventure, actor: Bruce Willis), preferences of customers (like: action / adventure), relations among item
Association Past purchases are transformed into relationships of common purchases 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
Association Past purchases are transformed into relationships of common purchases 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 2 1 1 1 2 1I1 I2 I3 I4 I1 I2 I3 I4 Also bought Whobought
Recommendation as a Ranking Task A recommender system aims at providing the user with a sorted list of suggestions. In Information Retrieval “providing a sorted list of suggestions” is a known as Raking Problem and it can be addressed using - Heuristics, i.e. numeric scores on query/document pairs (cosine similarity, BM25, LMIR probabilities, …) - Machine Learning (Learning to Rank, aka L2R or LETOR) - Ad-hoc metrics: Normalized Discounted Cumulative Gains, Mean Reciprocal Rank, Fraction of Concordant Pairs… - 3 approaches: pointwise, pairwise, listwise
Recommendations and Trust A social RS recommends items that are “popular” with the friends of the user, that is which are popular in a “social” proximity of the user. Trust for Collaborative Filtering ● Use trust to give more weight to some users or combined with similarity Trust for sorting & filtering: ● Prioritize information from trusted sources
Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches - Hybrid methods ● Context-based
+ Context
Context “Context is any information that can be used to characterize the situation of an entity” “Context are those variables which may change when a same activity is performed again and again” “Context is domain specific”
Context Physical context: time, position, and activity of the user, weather, light, and temperature ... Social context: the presence and role of other people around the user Interaction media context: the device used to access the system and the type of media that are browsed and personalized (text, music, images, movies, …) Modal context: the state of mind of the user, the user’s goals, mood, experience, and cognitive capabilities.
“Search should be processed in the context of the information surrounding them, allowing more accurate results that better reflect user’s actual intentions” Finkelstein et al. WWW 2001 Context in Information Retrieval
“Search should be processed in the context of the information surrounding them, allowing more accurate results that better reflect user’s actual intentions” Finkelstein et al. WWW 2001 The same holds for Recommender Systems nowadays Context in Information Retrieval
Context aware system
Spatial-temporal context Social context User Context Personal context Environmental context Task context ❏ Surroundings ❏ Things ❏ Services ❏ People ❏ Information ❏ ... ❏ Time ❏ Location type ❏ Lat/Lon ❏ Movement ❏ ... ❏ Information about friends ❏ User’s role ❏ Information about family ❏ Social Network data ❏ ... ❏ Preferences ❏ Mood ❏ Expertise ❏ Age ❏ Weight ❏ Height ❏ Disabilities ❏ ... ❏ Goal ❏ Task ❏ Activities ❏ ... Context Hierarchy
Adding Context Hierarchy to Domain Ontology Domain ontology describes set of concepts and categories in a specific domain, as well as more generic concept (Task, Goal, Action, Physical Objects…) in a multirelational semantic network. Integrating Domain Ontology with Context Hierarchy allows to build context dependent path in the system.
Spatial-Temporal context Personal context User Context Social context Environmental context Task context Context-aware Systems Context User Context Tag Service Desk Service Provider Time Location Task Action Goal User State Investment Lending Payments Investment Service Lending Service Invest user Savings Make money available to user Payment Service Isa Isa Isa Isa Isa Isa Isa Service Isa Has a Has a Has a Performs Performs Part of Isa Triggers Achieves Results in Has a
Context aware RecSys
Type of Context Aware RecSys Contextual Pre-filtering: Context information used to select relevant portions of data Contextual Post-filtering: Contextual information is used to filter/constrain/re-rank final set of recommendations Contextual Modeling: Context information is used directly as part of learning preference models
Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations c
Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations Data U x I x C 2D Recommender U x I → R Recommendations Contextual Recommendations c c
Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations Data U x I x C 2D Recommender U x I → R Recommendations Contextual Recommendations Data U x I x C1 x C2 x ... MD Recommender U x I x C1 x... → R Contextual Recommendations c c c
Adding Context U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items Context
Adding Context U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items I1 I2 I3 I4 Items | context c U1 U2 U3 U4 U5 U6 Users|contextc
Using Latent Variables U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items Context
Using Latent Variables U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items Context How much latent factors 1 and 2 are contained in C1 How much latent factors 1 and 2 U3 likes How much latent factors are contained in I3 6 X 2 2 X 4 4 X 2
Examples
Let’s represent our input data as two sets of nodes, the first related to assets and the second to customers C = {c1, c2, c3, ...} A = {a1, a2, a3, ….} In our case |C|>>|A| Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 ............ Customer set Asset set
Let’s represent our input data as two sets of nodes, the first related to assets and the second to customers, and draw who bought what C = {c1, c2, c3, ...} A = {a1, a2, a3, ….} In our case |C|>>|A| Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 ............ Customer set Asset set
Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 ............
Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = ............
Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............
Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............
Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............ Let’s compute q(i,j) for each i,j
“Markets are conversations.” The Cluetrain Manifesto, 1999
Word Embedding Methodology for mapping words or phrases from vocabulary to vectors of real numbers. 0.123 ... 5.344 -0.253 ... 2.333 ... 1.296 0.345 -0.453 ... 0.111 9.543 markets are conversations ...
Word2Vec Word2vec model takes as its input a large corpus of text and produces a vector space, with each unique word in the corpus being assigned a corresponding vector in the space. Word vectors are positioned in the vector space such that words that share common contexts in the corpus are located in close proximity to one another in the space.
Why context is relevant Word vectors capture linguistic regularities vec(“Paris”) - vec(“France”) + vec(“Italy”) is close to vec(“Rome”) vec(“walking”) - vec(“swimming”) + vec(“swam”) is close to vec(“walked”)
“You shall know a word by the company it keeps” J.R. Firth, 1957
“You shall know an item by the basket it belongs to”
Item embedding If word embedding can project words in a vector space taking into account of the other words along with they are usually accompanied... … then item embedding can project items in a vector space taking into account of the other items along with they are usually accompanied
Introduction to recommender systems
Questions?

More Related Content

PPT
Impersonal Recommendation system on top of Hadoop
Kostiantyn Kudriavtsev
 
PDF
Recommandation systems -
Yousef Fadila
 
PPTX
Recommender systems: Content-based and collaborative filtering
Viet-Trung TRAN
 
PPT
Social Recommender Systems Tutorial - WWW 2011
idoguy
 
PDF
An introduction to Recommender Systems
David Zibriczky
 
PDF
Recommender Systems
Carlos Castillo (ChaTo)
 
PPTX
Recommender system
Nilotpal Pramanik
 
PPT
Recommendation system
Vikrant Arya
 
Impersonal Recommendation system on top of Hadoop
Kostiantyn Kudriavtsev
 
Recommandation systems -
Yousef Fadila
 
Recommender systems: Content-based and collaborative filtering
Viet-Trung TRAN
 
Social Recommender Systems Tutorial - WWW 2011
idoguy
 
An introduction to Recommender Systems
David Zibriczky
 
Recommender Systems
Carlos Castillo (ChaTo)
 
Recommender system
Nilotpal Pramanik
 
Recommendation system
Vikrant Arya
 

What's hot (20)

PPTX
Recommendation system based on adaptive ontological graphs and weighted ranking
vikramadityajakkula
 
PPTX
Recommendation Systems - Why How and Real Life Applications
Liron Zighelnic
 
PPT
Recommendation techniques
sun9413
 
PPTX
Recommendation Systems Basics
Jarin Tasnim Khan
 
PPTX
Recommendation system
SAIFUR RAHMAN
 
PPTX
Recommendation system by_arpit_sharma
Er. Arpit Sharma
 
PPTX
Recommender system introduction
Liang Xiang
 
PDF
Recommender Systems
Francesco Casalegno
 
PDF
20120140506003
IAEME Publication
 
PDF
Introduction to Recommendation Systems
Zia Babar
 
PDF
Recommender system a-introduction
zh3f
 
PPT
A Hybrid Recommendation system
Pranav Prakash
 
PDF
Introduction to Recommendation Systems
Trieu Nguyen
 
PPTX
Customer to Customer recommendation system
sksaif95
 
PDF
Guide to Recommender Systems
Amancio Bouza
 
PDF
A survey on recommendation system
iosrjce
 
PDF
Recent advances in deep recommender systems
NAVER Engineering
 
PDF
C018211723
IOSR Journals
 
PPTX
Collaborative filtering
Kishor Datta Gupta
 
PPTX
Recommendation system
Akshat Thakar
 
Recommendation system based on adaptive ontological graphs and weighted ranking
vikramadityajakkula
 
Recommendation Systems - Why How and Real Life Applications
Liron Zighelnic
 
Recommendation techniques
sun9413
 
Recommendation Systems Basics
Jarin Tasnim Khan
 
Recommendation system
SAIFUR RAHMAN
 
Recommendation system by_arpit_sharma
Er. Arpit Sharma
 
Recommender system introduction
Liang Xiang
 
Recommender Systems
Francesco Casalegno
 
20120140506003
IAEME Publication
 
Introduction to Recommendation Systems
Zia Babar
 
Recommender system a-introduction
zh3f
 
A Hybrid Recommendation system
Pranav Prakash
 
Introduction to Recommendation Systems
Trieu Nguyen
 
Customer to Customer recommendation system
sksaif95
 
Guide to Recommender Systems
Amancio Bouza
 
A survey on recommendation system
iosrjce
 
Recent advances in deep recommender systems
NAVER Engineering
 
C018211723
IOSR Journals
 
Collaborative filtering
Kishor Datta Gupta
 
Recommendation system
Akshat Thakar
 
Ad

Similar to Introduction to recommender systems (20)

PPTX
Unit 1 Recommender Systems it's most important topic in machine
believemayur
 
PPT
JM Information Retrieval Techniques Unit V
JeyamohanHAsstProfCS
 
PPT
Recommender system in Information Retrieval system.ppt
DrSSriDharan
 
PPT
Culbert (1).pptmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm
MrMohammedZakirBELEC
 
PPT
Culbert recommender systems
Aravindharamanan S
 
PPTX
recommendation system techunique and issue
NutanBhor
 
PPTX
Recommender System _Module 1_Introduction to Recommender System.pptx
Satyam Sharma
 
PDF
Sistema de recomendações de Filmes do Netflix
Gabriel Peixe
 
PDF
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Xavier Amatriain
 
PPTX
Mini-training: Personalization & Recommendation Demystified
Betclic Everest Group Tech Team
 
PDF
Introduction to Recommendation Systems (Vietnam Web Submit)
Trieu Nguyen
 
PPT
recommendation system a topic in marketing analytics
PriyadharshiniG41
 
PDF
recommendationsystem-140410131156-phpapp01 (1).pdf
ssuserff0096
 
PDF
Use of data science in recommendation system
AkashPatil334
 
PDF
I017654651
IOSR Journals
 
DOCX
Recommender system definition, types with examples
vsathyavelusamy
 
PPT
Lec7 collaborative filtering
Aravindharamanan S
 
PPTX
ADM6274 - Final (NEHA)
Neha Gupta
 
PPT
recommender-systems-collaborative-filtering.ppt
aashokanuno
 
PPTX
Recommender Systems Explained
Pratik Bhudke
 
Unit 1 Recommender Systems it's most important topic in machine
believemayur
 
JM Information Retrieval Techniques Unit V
JeyamohanHAsstProfCS
 
Recommender system in Information Retrieval system.ppt
DrSSriDharan
 
Culbert (1).pptmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm
MrMohammedZakirBELEC
 
Culbert recommender systems
Aravindharamanan S
 
recommendation system techunique and issue
NutanBhor
 
Recommender System _Module 1_Introduction to Recommender System.pptx
Satyam Sharma
 
Sistema de recomendações de Filmes do Netflix
Gabriel Peixe
 
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Xavier Amatriain
 
Mini-training: Personalization & Recommendation Demystified
Betclic Everest Group Tech Team
 
Introduction to Recommendation Systems (Vietnam Web Submit)
Trieu Nguyen
 
recommendation system a topic in marketing analytics
PriyadharshiniG41
 
recommendationsystem-140410131156-phpapp01 (1).pdf
ssuserff0096
 
Use of data science in recommendation system
AkashPatil334
 
I017654651
IOSR Journals
 
Recommender system definition, types with examples
vsathyavelusamy
 
Lec7 collaborative filtering
Aravindharamanan S
 
ADM6274 - Final (NEHA)
Neha Gupta
 
recommender-systems-collaborative-filtering.ppt
aashokanuno
 
Recommender Systems Explained
Pratik Bhudke
 
Ad

More from Andrea Gigli (20)

PDF
How organizations can become data-driven: three main rules
Andrea Gigli
 
PDF
Equity Value for Startups.pdf
Andrea Gigli
 
PDF
Data Analytics per Manager
Andrea Gigli
 
PDF
Balance-sheet dynamics impact on FVA, MVA, KVA
Andrea Gigli
 
PDF
Reasons behind XVAs
Andrea Gigli
 
PDF
Recommendation Systems in banking and Financial Services
Andrea Gigli
 
PPTX
Mine the Wine by Andrea Gigli
Andrea Gigli
 
PDF
Fast Feature Selection for Learning to Rank - ACM International Conference on...
Andrea Gigli
 
PPTX
Feature Selection for Document Ranking
Andrea Gigli
 
PPTX
Using R for Building a Simple and Effective Dashboard
Andrea Gigli
 
PDF
Impact of Valuation Adjustments (CVA, DVA, FVA, KVA) on Bank's Processes - An...
Andrea Gigli
 
PDF
Comparing Machine Learning Algorithms in Text Mining
Andrea Gigli
 
PDF
Master in Big Data Analytics and Social Mining 20015
Andrea Gigli
 
PDF
Electricity Derivatives
Andrea Gigli
 
PPTX
Crawling Tripadvisor Attracion Reviews - Italiano
Andrea Gigli
 
PPTX
Search Engine for World Recipes Expo 2015
Andrea Gigli
 
PPTX
A Data Scientist Job Map Visualization Tool using Python, D3.js and MySQL
Andrea Gigli
 
PPTX
Search Engine Query Suggestion Application
Andrea Gigli
 
PPTX
From real to risk neutral probability measure for pricing and managing cva
Andrea Gigli
 
PPTX
Startup Saturday Internet Festival 2014
Andrea Gigli
 
How organizations can become data-driven: three main rules
Andrea Gigli
 
Equity Value for Startups.pdf
Andrea Gigli
 
Data Analytics per Manager
Andrea Gigli
 
Balance-sheet dynamics impact on FVA, MVA, KVA
Andrea Gigli
 
Reasons behind XVAs
Andrea Gigli
 
Recommendation Systems in banking and Financial Services
Andrea Gigli
 
Mine the Wine by Andrea Gigli
Andrea Gigli
 
Fast Feature Selection for Learning to Rank - ACM International Conference on...
Andrea Gigli
 
Feature Selection for Document Ranking
Andrea Gigli
 
Using R for Building a Simple and Effective Dashboard
Andrea Gigli
 
Impact of Valuation Adjustments (CVA, DVA, FVA, KVA) on Bank's Processes - An...
Andrea Gigli
 
Comparing Machine Learning Algorithms in Text Mining
Andrea Gigli
 
Master in Big Data Analytics and Social Mining 20015
Andrea Gigli
 
Electricity Derivatives
Andrea Gigli
 
Crawling Tripadvisor Attracion Reviews - Italiano
Andrea Gigli
 
Search Engine for World Recipes Expo 2015
Andrea Gigli
 
A Data Scientist Job Map Visualization Tool using Python, D3.js and MySQL
Andrea Gigli
 
Search Engine Query Suggestion Application
Andrea Gigli
 
From real to risk neutral probability measure for pricing and managing cva
Andrea Gigli
 
Startup Saturday Internet Festival 2014
Andrea Gigli
 

Recently uploaded (20)

PDF
DOC-20250806-WA0002._20250806_112011_0000.pdf
dhanusriss08
 
PPTX
AI-assistance in Knowledge Collection and Curation supporting Safe and Sustai...
Barry Hardy
 
PDF
Laughter Yoga Basic Learning Workshop Manual
yeechensee
 
PDF
Unit 1 Cost Accounting - Cost sheet
MANJU N
 
PDF
Ôn tập tiếng anh trong kinh doanh nâng cao
thaoonguyenn2311
 
PPT
Chapter four Project-Preparation material
argachewbochena1
 
PPT
Data mining for business intelligence ch04 sharda
salmaalsaeed3
 
PDF
pdfcoffee.com-opt-b1plus-sb-answers.pdfvi
thaoonguyenn2311
 
PDF
Business model innovation report 2022.pdf
pradvapal
 
PDF
Nidhal Samdaie CV - International Business Consultant
Nidhal Samdaie
 
DOCX
Euro SEO Services 1st 3 General Updates.docx
AmirNazir49
 
PDF
Traveri Digital Marketing Seminar 2025 by Corey and Jessica Perlman
Corey Perlman, Social Media Speaker and Consultant
 
PPTX
Lecture (1)-Introduction.pptx business communication
HamzaGhani10
 
PPTX
New Microsoft PowerPoint Presentation - Copy.pptx
itsmesumedh96
 
PDF
Roadmap Map-digital Banking feature MB,IB,AB
Alemayehu
 
PDF
Chapter 5_Foreign Exchange Market in .pdf
sophatphoniu
 
PPTX
HR Introduction Slide (1).pptx on hr intro
suhanidwivedi227
 
DOCX
unit 2 cost accounting- Tender and Quotation & Reconciliation Statement
MANJU N
 
DOCX
unit 1 COST ACCOUNTING AND COST SHEET
MANJU N
 
PDF
Power and position in leadershipDOC-20250808-WA0011..pdf
meghavinikumar2006
 
DOC-20250806-WA0002._20250806_112011_0000.pdf
dhanusriss08
 
AI-assistance in Knowledge Collection and Curation supporting Safe and Sustai...
Barry Hardy
 
Laughter Yoga Basic Learning Workshop Manual
yeechensee
 
Unit 1 Cost Accounting - Cost sheet
MANJU N
 
Ôn tập tiếng anh trong kinh doanh nâng cao
thaoonguyenn2311
 
Chapter four Project-Preparation material
argachewbochena1
 
Data mining for business intelligence ch04 sharda
salmaalsaeed3
 
pdfcoffee.com-opt-b1plus-sb-answers.pdfvi
thaoonguyenn2311
 
Business model innovation report 2022.pdf
pradvapal
 
Nidhal Samdaie CV - International Business Consultant
Nidhal Samdaie
 
Euro SEO Services 1st 3 General Updates.docx
AmirNazir49
 
Traveri Digital Marketing Seminar 2025 by Corey and Jessica Perlman
Corey Perlman, Social Media Speaker and Consultant
 
Lecture (1)-Introduction.pptx business communication
HamzaGhani10
 
New Microsoft PowerPoint Presentation - Copy.pptx
itsmesumedh96
 
Roadmap Map-digital Banking feature MB,IB,AB
Alemayehu
 
Chapter 5_Foreign Exchange Market in .pdf
sophatphoniu
 
HR Introduction Slide (1).pptx on hr intro
suhanidwivedi227
 
unit 2 cost accounting- Tender and Quotation & Reconciliation Statement
MANJU N
 
unit 1 COST ACCOUNTING AND COST SHEET
MANJU N
 
Power and position in leadershipDOC-20250808-WA0011..pdf
meghavinikumar2006
 

Introduction to recommender systems

  • 1. Introduction to Recommender Systems Andrea Gigli, Boston Consulting Group & University of Florence @andrgig andrgig@gmail.com
  • 2. Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge
  • 3. Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Filtering Information
  • 4. Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Processing Information
  • 5. Why a talk on recommender systems? We’re living in the era of ● Information Overwhelming ● Choice deluge Users and Customers react ● Procrastinating the choices ● Ignoring information ● Avoiding the task ● Reshaping the task ● Using shortcuts ● Processing Information Information Filtering and Extraction
  • 6. How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find
  • 7. How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice
  • 8. How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice ● Recommender Systems to discover something you didn’t know existed
  • 9. How to filter information Information Filtering and Extraction: ● Search Engines to look for something you expect to find ○ Given a query, retrieve a list of documents, products, services, ads ○ A query can be a term, sentence, image, audio, voice ● Recommender Systems to discover something you didn’t know existed ○ Items are recommended one the basis of the user (and context) ○ Items means something domain-specific ■ Movies: movie, actor, director, genre ■ Music: track, album, artist, playlist, … ■ E-Commerce: item, bundle of items, … ■ Banking: banking and financial products, investment advice, service bundle, …
  • 10. “We are leaving the age of information and entering the age of recommendation.” Chris Anderson, 2004, The Long Tail
  • 11. What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user.
  • 12. What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user. Recommender Systems are designed to help - Users in doing the best choice for them - Companies in selling more products and increasing customers loyalty
  • 13. What is a recommender system? A Recommender System is an set of principles and procedures that seek to predict which item is preferred by a user. Recommender Systems are designed to help - Users in doing the best choice for them - Companies in selling more products and increasing customers loyalty Users != “someone to charge” Items != “something to sell to users”
  • 14. When a Recommender System is good? - It is relevant to the user (personalization) - It represents all the possible interests of one user (diversity) - Expands the user's taste into neighboring areas (serendipity)
  • 15. Type of Recommender System ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
  • 16. Type of Recommender System ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based Which is the best one? Depends on the domain and particular problem. Usually combining different approaches works better
  • 17. Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
  • 18. Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions.
  • 19. Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content
  • 20. Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content - Second, define a user profile by analysing the relevant content of the items the user liked in the past
  • 21. Content - Based Recommenders Recommendations are based on the information on the content of items rather than on other users’ opinions. - First, define the relevant items content - Second, define a user profile by analysing the relevant content of the items the user liked in the past - Third, recommended items for a user’s profile built on the second step
  • 22. Content - Based Recommenders Item profile User profile f1 f2 f3 fN wf1 wf2 wf3 wfN a set of features (characteristics/attributes) characterizing item s. a set of weights, where wfi denotes the importance of feature fi to the user An item is recommended to the user accordingly to the features held by the item which are mostly relevant to the user
  • 23. Content-Based Recommenders pros ● No need for data on other users. ● Able to recommend to users with unique tastes. ● Able to recommend new and unpopular items. ● Can provide explanations of recommended items by listing content-features that caused an item to be recommended
  • 24. Content-Based Recommenders cons ● Requires content that can be encoded as meaningful features. ● Users’ tastes must be represented as a learnable function of these content features. ● Hard to exploit quality judgements of other users. ● Easy to overfit
  • 25. Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
  • 26. Collaborative Filtering - What is it? The task of predicting (filtering) specific user preferences on new items by collecting taste information from many users (collaborative). Collaborative filtering pros: ● Requires minimal domain knowledge ● Users and Items have no internal structure or characteristics ● Perform well in most cases
  • 27. Collaborative Filtering - What is it? The task of predicting (filtering) specific user preferences on new items by collecting taste information from many users (collaborative). Collaborative filtering cons: ● Prior behavior determines current behavior ● Many items, few data per user, no data for new user ● Very large datasets
  • 28. Collaborative Filtering Sparsity Problem In 1,000,000 items, the probability that two users who bought 100 items each, have an item in common is 1% In 10,000,000 items, the probability that two users who bought 50 items each, have an item in common is 0.02%
  • 29. Collaborative Filtering - How does it work? Let’s represent the information on past users’ choices and choose items on a matrix. Each cell contains the preference of user u for item i Preference can be explicit (eg rating, stars) or implicit (bought/not bought, liked/not liked) U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 30. Collaborative Filtering - How does it work? Let’s represent the information on past users’ choices and choose items on a matrix. Each cell contains the preference of user u for item i Preference can be explicit (eg rating, stars) or implicit (bought/not bought, liked/not lied) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 31. User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 32. User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 33. User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 3. Compute similarities between user and similar users 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 34. User-based Collaborative Filtering 1. Identify the user you want to give a recommendation 2. Identify users who made same choices (similar users) 3. Compute similarities between user and similar users 4. Predict user’s preference for each Item weighting for the similarity with other users 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 35. Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 36. Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 37. Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 3. Compute the similarity between the target item and the neighborhood items 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 38. Item-based Collaborative Filtering 1. Identify the item you want to recommend (target item) 2. Identify items rated by the same users who choose the target item (neighborhood items) 3. Compute the similarity between the target item and the neighborhood items 4. Predict ratings for the target item for each user weighting for the similarity with other items 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 39. Tackling the Sparsity and Scalability
  • 40. Tackling the Sparsity and Scalability Sparsity: - Poor relationship among like minded but sparse-rating users. - Accuracy can be poor if one try to make predictions in sparse dataset using nearest neighbor algorithms Scalability: - Nearest neighbor require computation that grows with both the number of users and the number of items. Possible solution: - Using latent factor models to capture similarity between users & items in a reduced dimensional space.
  • 41. Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 42. Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 F1 F2 F1 F2
  • 43. Model Based Collaborative Filtering: Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 How much latent factor 1 and 2 are contained in Item 2 How much User 3 likes latent factors 1 and 2 F1 F2 F1 F2
  • 44. Collaborative Filtering via Matrix Factorization Idea: project the problem into a lower dimension space 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 1 1 1 1 1 1 1 1 How much latent factor 1 and 2 are contained in I2 How much U3 likes latent factors 1 and 2 The dot product indicates how much U3 likes I2 F1 F2 F1 F2
  • 45. Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches ● Context-Based
  • 46. Clustering Another way to make recommendations based on past purchases is through Customers clustering → Each cluster will be assigned typical preferences, based on preferences of customers who belong to the cluster → Customers within each cluster will receive recommendations computed at the cluster level PROS CONS Can be used to work on aggregated Data Can be applied as a first step for filtering techniques Can be used to capture latent similarities between users or items Recommendations per customer's cluster are usually less relevant for single user than those from collaborative filtering
  • 47. Classifiers Classifiers are general computational models trained using positive and negative examples (Logistic Regression, Bayesian Networks, Support Vector Machines, Decision Trees, etc…) They may take in inputs vector of item features (type: action / adventure, actor: Bruce Willis), preferences of customers (like: action / adventure), relations among item
  • 48. Association Past purchases are transformed into relationships of common purchases 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users
  • 49. Association Past purchases are transformed into relationships of common purchases 1 1 1 1 1 1 1 1 1 1U1 U2 U3 U4 U5 U6 I1 I2 I3 I4 Items Users 2 1 1 1 2 1I1 I2 I3 I4 I1 I2 I3 I4 Also bought Whobought
  • 50. Recommendation as a Ranking Task A recommender system aims at providing the user with a sorted list of suggestions. In Information Retrieval “providing a sorted list of suggestions” is a known as Raking Problem and it can be addressed using - Heuristics, i.e. numeric scores on query/document pairs (cosine similarity, BM25, LMIR probabilities, …) - Machine Learning (Learning to Rank, aka L2R or LETOR) - Ad-hoc metrics: Normalized Discounted Cumulative Gains, Mean Reciprocal Rank, Fraction of Concordant Pairs… - 3 approaches: pointwise, pairwise, listwise
  • 51. Recommendations and Trust A social RS recommends items that are “popular” with the friends of the user, that is which are popular in a “social” proximity of the user. Trust for Collaborative Filtering ● Use trust to give more weight to some users or combined with similarity Trust for sorting & filtering: ● Prioritize information from trusted sources
  • 52. Type of Recommender System? ● Content-Based Recommender Systems ● Collaborative Filtering ● Other Approaches - Hybrid methods ● Context-based
  • 54. Context “Context is any information that can be used to characterize the situation of an entity” “Context are those variables which may change when a same activity is performed again and again” “Context is domain specific”
  • 55. Context Physical context: time, position, and activity of the user, weather, light, and temperature ... Social context: the presence and role of other people around the user Interaction media context: the device used to access the system and the type of media that are browsed and personalized (text, music, images, movies, …) Modal context: the state of mind of the user, the user’s goals, mood, experience, and cognitive capabilities.
  • 56. “Search should be processed in the context of the information surrounding them, allowing more accurate results that better reflect user’s actual intentions” Finkelstein et al. WWW 2001 Context in Information Retrieval
  • 57. “Search should be processed in the context of the information surrounding them, allowing more accurate results that better reflect user’s actual intentions” Finkelstein et al. WWW 2001 The same holds for Recommender Systems nowadays Context in Information Retrieval
  • 59. Spatial-temporal context Social context User Context Personal context Environmental context Task context ❏ Surroundings ❏ Things ❏ Services ❏ People ❏ Information ❏ ... ❏ Time ❏ Location type ❏ Lat/Lon ❏ Movement ❏ ... ❏ Information about friends ❏ User’s role ❏ Information about family ❏ Social Network data ❏ ... ❏ Preferences ❏ Mood ❏ Expertise ❏ Age ❏ Weight ❏ Height ❏ Disabilities ❏ ... ❏ Goal ❏ Task ❏ Activities ❏ ... Context Hierarchy
  • 60. Adding Context Hierarchy to Domain Ontology Domain ontology describes set of concepts and categories in a specific domain, as well as more generic concept (Task, Goal, Action, Physical Objects…) in a multirelational semantic network. Integrating Domain Ontology with Context Hierarchy allows to build context dependent path in the system.
  • 61. Spatial-Temporal context Personal context User Context Social context Environmental context Task context Context-aware Systems Context User Context Tag Service Desk Service Provider Time Location Task Action Goal User State Investment Lending Payments Investment Service Lending Service Invest user Savings Make money available to user Payment Service Isa Isa Isa Isa Isa Isa Isa Service Isa Has a Has a Has a Performs Performs Part of Isa Triggers Achieves Results in Has a
  • 63. Type of Context Aware RecSys Contextual Pre-filtering: Context information used to select relevant portions of data Contextual Post-filtering: Contextual information is used to filter/constrain/re-rank final set of recommendations Contextual Modeling: Context information is used directly as part of learning preference models
  • 64. Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations c
  • 65. Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations Data U x I x C 2D Recommender U x I → R Recommendations Contextual Recommendations c c
  • 66. Type of Context Aware RecSys Data U x I x C Contextualized Data: U x I 2D Recommender U x I → R Contextual Recommendations Data U x I x C 2D Recommender U x I → R Recommendations Contextual Recommendations Data U x I x C1 x C2 x ... MD Recommender U x I x C1 x... → R Contextual Recommendations c c c
  • 68. Adding Context U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items I1 I2 I3 I4 Items | context c U1 U2 U3 U4 U5 U6 Users|contextc
  • 70. Using Latent Variables U1 U2 U3 U4 U5 U6 Users I1 I2 I3 I4 Items Context How much latent factors 1 and 2 are contained in C1 How much latent factors 1 and 2 U3 likes How much latent factors are contained in I3 6 X 2 2 X 4 4 X 2
  • 72. Let’s represent our input data as two sets of nodes, the first related to assets and the second to customers C = {c1, c2, c3, ...} A = {a1, a2, a3, ….} In our case |C|>>|A| Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 ............ Customer set Asset set
  • 73. Let’s represent our input data as two sets of nodes, the first related to assets and the second to customers, and draw who bought what C = {c1, c2, c3, ...} A = {a1, a2, a3, ….} In our case |C|>>|A| Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 ............ Customer set Asset set
  • 75. Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = ............
  • 76. Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............
  • 77. Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............
  • 78. Bipartite Graph a1 a2 a3 a4 a5 c1 c2 c3 c4 c5 a1 a2 a3a4 a5 Each edge can be weighted by a similarity measure, like |C(ai )| + |C(aj )| |C(ai ,aj )| q(i,j) = Example: q(a1,a2) = 1 / (3 + 2) = 0.20 q(a4,a5) = 1 / (1 + 2) = 0.333 ............ Let’s compute q(i,j) for each i,j
  • 80. Word Embedding Methodology for mapping words or phrases from vocabulary to vectors of real numbers. 0.123 ... 5.344 -0.253 ... 2.333 ... 1.296 0.345 -0.453 ... 0.111 9.543 markets are conversations ...
  • 81. Word2Vec Word2vec model takes as its input a large corpus of text and produces a vector space, with each unique word in the corpus being assigned a corresponding vector in the space. Word vectors are positioned in the vector space such that words that share common contexts in the corpus are located in close proximity to one another in the space.
  • 82. Why context is relevant Word vectors capture linguistic regularities vec(“Paris”) - vec(“France”) + vec(“Italy”) is close to vec(“Rome”) vec(“walking”) - vec(“swimming”) + vec(“swam”) is close to vec(“walked”)
  • 83. “You shall know a word by the company it keeps” J.R. Firth, 1957
  • 84. “You shall know an item by the basket it belongs to”
  • 85. Item embedding If word embedding can project words in a vector space taking into account of the other words along with they are usually accompanied... … then item embedding can project items in a vector space taking into account of the other items along with they are usually accompanied