Tutorial: Context In Recommender Systems

Tutorial: Context In
Recommender Systems
Yong Zheng
Center for Web Intelligence
DePaul University, Chicago
Time: 2:30 PM – 6:00 PM, April 4, 2016
Location: Palazzo dei Congressi, Pisa, Italy
The 31st ACM Symposium on Applied Computing, Pisa, Italy, 2016

Introduction
Yong Zheng
DePaul University, Chicago, IL, USA
2010 – 2016, PhD in Computer Science, DePaul University
Research: User Modeling and Recommender Systems
Schedule of this Tutorial:
Time: 2:30 PM – 6:00 PM, April 4, 2016
Coffee Break: 4:00 PM – 4:30 PM, April 4, 2016
2

Topics in this Tutorial
• Traditional Recommendation
e.g., Give me a list of recommended movies to watch
• Context-aware Recommendation
e.g., Give me a list of recommended movies to watch, if
Time & Location: at weekend and in cinema
Companion: with girlfriend v.s. with Kids
• Context Suggestion
The best time/location to watch movie “Life of PI”
3

Outline
• Background: Recommender Systems
Introduction and Applications
Tasks and Evaluations
Traditional Recommendation Algorithms
Context Definition, Acquisition and Selection
Context Incorporation: Algorithms
Other Challenges
CARSKit: A Java-Based Open-source RecSys Library
• Summary and Future Directions
4

Outline
List of Traditional Recommendation Algorithms
Collaborative Filtering
 User/Item Based Collaborative Filtering
 Sparse Linear Method
 Matrix Factorization
6

Outline
7

Recommender System (RS)
• RS: item recommendations tailored to user tastes
8

How it works
12
Binary FeedbackRatings Reviews Behaviors
• User Preferences
Explicit Implicit

Rating-Based Data Sets
13
T1 T2 T3 T4 T5 T6 T7 T8
U1 4 3 4 2 5
U2 3 4 2 5
U3 4 4 2 2 4
U4 3 5 2 4
U5 2 5 2 4 ? 4
User demographic Information: Age, Gender, Race, Country, etc
Item feature information: Movie/Music Genre, Movie director, Music Composer, etc

Outline
14

15
Task and Eval (1): Rating Prediction
User Item Rating
U1 T1 4
U1 T2 3
U1 T3 3
U2 T2 4
U2 T3 5
U2 T4 5
U3 T4 4
U1 T4 3
U2 T1 2
U3 T1 3
U3 T2 3
U3 T3 4
Train
Test
Task: P(U, T) in testing set
Prediction error: e = R(U, T) – P(U, T)
Mean Absolute Error (MAE) =
Other evaluation metrics:
• Root Mean Square Error (RMSE)
• Coverage
• and more …

16
Task and Eval (1): Rating Prediction
User Item Rating
U1 T1 4
U1 T2 3
U1 T3 3
U2 T2 4
U2 T3 5
U2 T4 5
U3 T4 4
U1 T4 3
U2 T1 2
U3 T1 3
U3 T2 3
U3 T3 4
Train
Test
Task: P(U, T) in testing set
1. Build a model, e.g., P(U, T) = Avg (T)
2. Process of Rating Prediction
P(U1, T4) = Avg(T4) = (5+4)/2 = 4.5
P(U2, T1) = Avg(T1) = 4/1 = 4
P(U3, T1) = Avg(T1) = 4/1 = 4
P(U3, T2) = Avg(T2) = (3+4)/2 = 3.5
P(U3, T3) = Avg(T3) = (3+5)/2 = 4
3. Evaluation by Metrics
Mean Absolute Error (MAE) =
ei = R(U, T) – P(U, T)
MAE = (|3 – 4.5| + |2 - 4| + |3 - 4| +
|3 – 3.5| + |4 - 4|) / 5 = 1

17
Task and Eval (2): Top-N Recommendation
User Item Rating
U1 T1 4
U1 T2 3
U1 T3 3
U2 T2 4
U2 T3 5
U2 T4 5
U3 T4 4
U1 T4 3
U2 T1 2
U3 T1 3
U3 T2 3
U3 T3 4
Train
Test
Task: Top-N Items to a user U3
Predicted Rank: T3, T1, T4, T2
Real Rank: T3, T2, T1
Then compare the two lists:
Precision@N = # of hits/N
Other evaluation metrics:
• Recall
• Mean Average Precision (MAP)
• Normalized Discounted Cumulative Gain (NDCG)
• Mean Reciprocal Rank (MRR)
• and more …

18
Task and Eval (2): Top-N Recommendation
User Item Rating
U1 T1 4
U1 T2 3
U1 T3 3
U2 T2 4
U2 T3 5
U2 T4 5
U3 T4 4
U1 T4 3
U2 T1 2
U3 T1 3
U3 T2 3
U3 T3 4
Train
Test
Task: Top-N Items to user U3
1. Build a model, e.g., P(U, T) = Avg (T)
2. Process of Rating Prediction
P(U3, T1) = Avg(T1) = 4/1 = 4
P(U3, T2) = Avg(T2) = (3+4)/2 = 3.5
P(U3, T3) = Avg(T3) = (3+5)/2 = 4
P(U3, T4) = Avg(T4) = (4+5)/2 = 3.5
Predicted Rank: T3, T1, T4, T2
Real Rank: T3, T2, T1
3. Evaluation Based on the two lists
Precision@N = # of hits/N
Precision@1 = 1/1
Precision@2 = 2/2
Precision@3 = 2/3

Outline
19

Traditional Recommendation Algorithms
• Five Types of algorithms by R. Burke, 2002
e.g., Neighborhood-based algorithms
Content-based Recommender
e.g., reusing item features to measure item similarities
Demographic Approaches
e.g., reusing user demographic info for marketing purpose
Knowledge-based Algorithms
e.g., mining knowledge/relations among users, items
Utility-based Recommender
e.g., by maximizing a predefined utility function
20

Outline
21

Preliminary: Collaborative Filtering (CF)
• List of three popular CF-based algorithms
Neighborhood-based Collaborative Filtering
e.g., User/Item based algorithms
Sparse Linear Method (SLIM)
i.e., a learning-based KNN-based CF approach
Matrix Factorization (MF)
i.e., a model based collaborative filtering
22

User-Based Collaborative Filtering
• In User-based K-Nearest Neighbor CF (UserKNN)
 Assumption: U3’s rating on T5 is similar to other users’ ratings on T5,
where these users have similar taste with U3.
 The “K-Nearest Neighbor” (user neighborhood), can be selected from
a list of top-similar users (to U3) identified from the co-ratings by each
pair of the users
23
T1 T2 T3 T4 T5 T6 T7 T8
U1 4 3 4 2 5
U2 3 4 2 5 2 5
U3 4 3 ? 2 2 4
U4 3 5 2 4 3
U5 2 5 2 2 4 2

User-Based Collaborative Filtering
• UserKNN, P. Resnick, et al., 1994
User: a; Item: i; User Neighbor: u
Similarity between user u and a: sim(a, u)
24

Item-Based Collaborative Filtering
• In Item-based K-Nearest Neighbor CF (ItemKNN)
 Assumption: U3’s rating on T5 is similar to U3’s rating on similar items.
 The “K-Nearest Neighbor” (item neighborhood), can be selected from
a list of top-similar items (to T5) identified from the co-ratings by each
pair of the items
25
T1 T2 T3 T4 T5 T6 T7 T8
U1 4 3 4 2 5
U2 3 4 2 5 2 5
U3 4 3 ? 2 2 4
U4 3 5 2 4 3
U5 2 5 2 2 4 2

Item-Based Collaborative Filtering
• ItemKNN, B. Sarwar et al., 2001
User: a; Item: i; Item Neighbor: j
Similarity between item i and j: sim(i, j)
26
1. Neighbor selection
2. Neighbor contribution
3. Item similarity

Sparse Linear Method
• Sparse Linear Method (SLIM), X. Ning, et al., 2011
Item coefficient (W) is the same as item similarity
We learn W directly for top-N recommendation Task
27

Sparse Linear Method
• Sparse Linear Method (SLIM), X. Ning, et al., 2011
28
Squared Error L2 Norm
(Ridge Reg)
L1 Norm
(Lasso Reg)

Matrix Factorization
• Matrix Factorization (MF), Y. Koren, et al., 2009
29
User HarryPotter Batman Spiderman
U1 5 3 4
U2 ? 2 4
U3 4 2 ?
R P Q

30
R P Q
R = Rating Matrix, m users, n movies;
P = User Matrix, m users, f latent factors/features;
Q = Item Matrix, n movies, f latent factors/features;
Interpretation:
pu indicates how much user likes f latent factors;
qi means how much one item obtains f latent factors;
The dot product indicates how much user likes item;

31
minq,p S (u,i) e R ( rui - qt
i pu )2 + l (|qi|2 + |pu|2 )
Goal: Try to learn P and Q by minimizing the squared error
goodness of fit regularization
Goodness of fit: to reduce the prediction errors;
Regularization term: to alleviate the overfitting;

32
minq,p S (u,i) e R ( rui - qt
i pu )2 + l (|qi|2 + |pu|2 )
By using Stochastic Gradient Descent (SGD) or Alternating Least
Squares (ALS), we are able to learn the P and Q iteratively.
goodness of fit regularization

Example of Evaluations on CF Algorithms
• Data set: MovieLens-100K
33
There are 100K ratings given by 943 users on 1,682 movies
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
ItemKNN UserKNN MF SLIM
Precision@10

Summary: Traditional RecSys
• Traditional RecSys: Users × Items Ratings
• Two recommendation Task:
Task 1: Rating Prediction
Task 2: Top-N Recommendation
• There are several types of recsys algorithms
• Three popular collaborative filtering (CF):
User/Item Based K-Nearest Neighbor CF
Sparse Linear Method
Matrix Factorization
34

Outline
Intro: Does context matter?
 Definition: What is Context?
 Acquisition: How to collect context?
 Selection: How to identify the relevant context?
Context Filtering
Context Modeling
Other Challenges and CARSKit
36

Outline
Context Modeling
37

Non-context vs Context
38
Companion
• Decision Making = Rational + Contextual
• Examples:
 Travel destination: in winter vs in summer
 Movie watching: with children vs with partner
 Restaurant: quick lunch vs business dinner
 Music: workout vs study

What is Context?
39
• “Context is any information that can be used to characterize the
situation of an entity” by Anind K. Dey, 2001
• Representative Context: Fully Observable and Static
• Interactive Context: Non-fully observable and Dynamic

Interactive Context Adaptation
40
• Interactive Context: Non-fully observable and Dynamic
List of References:
 M Hosseinzadeh Aghdam, N Hariri, B Mobasher, R Burke. "Adapting
Recommendations to Contextual Changes Using Hierarchical Hidden
Markov Models", ACM RecSys 2015
 N Hariri, B Mobasher, R Burke. "Adapting to user preference changes
in interactive recommendation", IJCAI 2015
 N Hariri, B Mobasher, R Burke. "Context adaptation in interactive
recommender systems", ACM RecSys 2014
 N Hariri, B Mobasher, R Burke. "Context-aware music
recommendation based on latenttopic sequential patterns", ACM
RecSys 2012

CARS With Representative Context
41
• Observed Context:
Contexts are those variables which may change when a same
activity is performed again and again.
• Examples:
Watching a movie: time, location, companion, etc
Listening to a music: time, location, emotions, occasions, etc
Party or Restaurant: time, location, occasion, etc
Travels: time, location, weather, transportation condition, etc

Context-aware RecSys (CARS)
42
• Traditional RS: Users × Items Ratings
• Contextual RS: Users × Items × Contexts Ratings
Example of Multi-dimensional Context-aware Data set
User Item Rating Time Location Companion
U1 T1 3 Weekend Home Kids
U1 T2 5 Weekday Home Partner
U2 T2 2 Weekend Cinema Partner
U2 T3 3 Weekday Cinema Family
U1 T3 ? Weekend Cinema Kids

Terminology in CARS
43
• Example of Multi-dimensional Context-aware Data set
Context Dimension: time, location, companion
Context Condition: Weekend/Weekday, Home/Cinema
Context Situation: {Weekend, Home, Kids}
User Item Rating Time Location Companion
U1 T1 3 Weekend Home Kids
U1 T2 5 Weekday Home Partner
U2 T2 2 Weekend Cinema Partner
U2 T3 3 Weekday Cinema Family
U1 T3 ? Weekend Cinema Kids

Context Acquisition
44
How to Collect the context and user preferences in contexts?
• By User Surveys or Explicitly Asking for User Inputs
Predefine context & ask users to rate items in these situations;
Or directly ask users about their contexts in user interface;
• By Usage data
The log data usually contains time and location (at least);
User behaviors can also infer context signals;

Examples: Context Acquisition (RealTime)
45

Examples: Context Acquisition (Explicit)
46

47

48
Mobile App: South Tyrol Suggests
Personality
Collection
Context
Collection

Examples: Context Acquisition (PreDefined)
49

Examples: Context Acquisition (PreDefined)
50
Google Music: Listen Now

Examples: Context Acquisition (User Behavior)
51

Context Relevance and Context Selection
52
Apparently, not all of the context are relevant or influential
• By User Surveys
e.g., which ones are important for you in this domain
• By Feature Selection
e.g., Principal Component Analysis (PCA)
e.g., Linear Discriminant Analysis (LDA)
• By Statistical Analysis or Detection on Contextual Ratings
Statistical test, e.g., Freeman-Halton Test
Other methods: information gain, mutual information, etc
Reference: Odic, Ante, et al. "Relevant context in a movie
recommender system: Users’ opinion vs. statistical detection."
CARS Workshop@ACM RecSys 2012

Context-aware Data Sets
53
Public Data Set for Research Purpose
• Food: AIST Japan Food, Mexico Tijuana Restaurant Data
• Movies: AdomMovie, DePaulMovie, LDOS-CoMoDa Data
• Music: InCarMusic
• Travel: TripAdvisor, South Tyrol Suggests (STS)
• Mobile: Frappe
Frappe is a large data set, others are either small or sparse
Downloads and References:
https://github.com/irecsys/CARSKit/tree/master/context-
aware_data_sets

Outline
Context Modeling
54

55
• There are three ways to build algorithms for CARS

56
• Next, we focus on the following CARS algorithms:
Contextual Filtering: Use Context as Filter
Contextual Modeling: Independent vs Dependent
Note: We focus on context-aware collaborative filtering

Contextual Filtering
• Differential Context Modeling
• UI Splitting

58
Differential Context Modeling

59
• Data Sparsity Problem in Contextual Rating
User Movie Time Location Companion Rating
U1 Titanic Weekend Home Girlfriend 4
U2 Titanic Weekday Home Girlfriend 5
U3 Titanic Weekday Cinema Sister 4
U1 Titanic Weekday Home Sister ?
Context Matching  Only profiles given in <Weekday, Home, Sister>
Context Relaxation  Use a subset of context dimensions to match
Context Weighting  Use all profiles, but weighted by context similarity

60
• Solutions Applied to Collaborative Filtering
Context Matching  Only profiles given in <Weekday, Home, Sister>
Context Relaxation  Use a subset of context dimensions to match
Context Weighting  Use all profiles, but weighted by context similarity
In short, we want to use a subset of rating profiles in collaborative filtering.
There are some research applied such filters to UserKNN or ItemKNN.
But there are two main drawbacks:
1). They just apply contexts as filters in one component
e.g., the neighborhood selection
2). They just use the same selected contexts as filters
i.e., different context dimensions may be useful for different components

61
• Context Relaxation
Use {Time, Location, Companion}  0 record matched!
Use {Time, Location}  1 record matched!
Use {Time}  2 records matched!
Note: a balance is required for relaxation and accuracy

62
• Context Weighting
c and d are two contexts. (Two red regions in the Table above.)
σ is the weighting vector <w1, w2, w3> for three dimensions.
Assume they are equal weights, w1 = w2 = w3 = 1.
J(c, d, σ) = # of matched dimensions / # of all dimensions = 2/3
Similarity of contexts is measured by
Weighted Jaccard similarity

63
• Notion of “differential”
In short, we apply different context relaxation and context weighting to
each component
1.Neighbor Selection 2.Neighbor contribution
3.User baseline 4.User Similarity

64
• Workflow
Step-1: We decompose an algorithm to different components;
Step-2: We try to find optimal context relaxation/weighting:
 In context relaxation, we select optimal context dimensions
 In context weighting, we find optimal weights for each dimension
• Optimization Problem
Assume there are 4 components and 3 context dimensions
1 2 3 4 5 6 7 8 9 10 11 12
DCR 1 0 0 0 1 1 1 1 0 1 1 1
DCW 0.2 0.3 0 0.1 0.2 0.3 0.5 0.1 0.2 0.1 0.5 0.2
1st
2nd 3rd 4th

65
• Optimization Approach
 Particle Swarm Optimization (PSO)
 Genetic Algorithms
 Other non-linear approaches
Fish Birds Bees

66
• How PSO works?
Swarm = a group of birds
Particle = each bird ≈ search entity in algorithm
Vector = bird’s position in the space ≈ Vectors we need in DCR/DCW
Goal = the distance to location of pizza ≈ prediction error
So, how to find goal by swam intelligence?
1.Looking for the pizza
Assume a machine can tell the distance
2.Each iteration is an attempt or move
3.Cognitive learning from particle itself
Am I closer to the pizza comparing with
my “best ”locations in previous history?
4.Social Learning from the swarm
Hey, my distance is 1 mile. It is the closest!
. Follow me!! Then other birds move towards here.
DCR – Feature selection – Modeled by binary vectors – Binary PSO
DCW – Feature weighting – Modeled by real-number vectors – PSO

67
• Summary
Pros: Alleviate data sparsity problem in CARS
Cons: Computational complexity in optimization
Cons: Local optimum by non-linear optimizer
Our Suggestion:
 We may just run these optimizations offline to find optimal
context relaxation or context weighting solutions; And those
optimal solutions can be obtained periodically;

68
Context-aware Splitting Approaches

69
The underlying idea in item splitting is that the nature of an item, from the user's
point of view, may change in different contextual conditions, hence it may be
useful to consider it as two different items. (L. Baltrunas, F. Ricci, RecSys'09) – In
short, contexts are dependent with items.
Intro: Item Splitting
At Cinema At Home At Swimming Pool

70
User Item Location Rating
U1 M1 Pool 5
U2 M1 Pool 5
U3 M1 Pool 5
U1 M1 Home 2
U4 M1 Home 3
U2 M1 Cinema 2
High Rating
Low Rating
Significant difference?
Let’s split it !!!
M11: being seen at Pool
M12: being seen at Home
M1
Same movie,
different IDs.

71
User Item Loc Rating
U1 M1 Pool 5
U2 M1 Pool 5
U3 M1 Pool 5
U1 M1 Home 2
U4 M1 Home 3
U2 M1 Cinema 2
User Item Rating
U1 M11 5
U2 M11 5
U3 M11 5
U1 M12 2
U4 M12 3
U2 M12 2
Transformation
Question:
How to find such a split? Pool and Non-pool, or Home and Non-home?
Which one is the best or optimal split?

72
• Splitting Criteria (Impurity Criteria)
Impurity criteria and significance test are used to make the selection.
There are 4 impurity criteria for splitting by L. Baltrunas, et al, RecSys'09:
tmean (t-test), tprop (z-test), tchi (chi-square test), tIG (Information gain)
Take tmean for example, tmean, is defined using the two-sample t test and
computes how significantly different are the means of the rating in the two
rating subsets, when the split c (c is a context condition, e.g. location =
Pool) is used. The bigger the t value of the test is, the more likely the
difference of the means in the two partitions is significant (at 95%
confidence value). Choose the largest one!

73
Other Context-aware Splitting Approaches
• User Splitting and UI Splitting
Similarly, the splitting approach can be applied to user too!
• User Splitting: is a similar one. Instead of splitting items, it may be
useful to consider one user as two different users, if user demonstrates
significantly different preferences across contexts. (A. Said et al.,
CARS@RecSys 2011) In short, contexts are dependent with users.
• UI Splitting: simply a combination of item splitting and user splitting –
both approaches are applied to create a new rating matrix – new users
and new items are created in the rating matrix. (Y. Zheng, et al, ACM
SAC 2014). In short, it fuses dependent contexts to users and items
simultaneously at the same time.

74
Splitting and Transformation

75
How Splitting Approaches Work?
• Recommendation Process
Find the best split to perform a splitting approach;
After splitting, we obtain a User-item rating matrix;
And we can further apply any traditional Recommendation algorithms;
Take Matrix Factorization for example:
Rating Prediction:
Objective function:
Parameter updates
Based on SGD

76
Context-aware Splitting Approaches
• Summary
Pros: Contexts are fused into user and/or item dimensions
Cons: We create new users/items, which increases data sparsity
Our Suggestion:
 We may build a hybrid recommender to alleviate the data
sparsity or cold-start problems introduced by UI Splitting

77
Experimental Results
Japan Food Data: 6360 ratings given by 212 users on 20 items within 2 context dimensions
0.6
0.7
0.8
0.9
1
1.1
1.2
PreFiltering DCR DCW uSplitting iSplitting uiSplitting
Food
MAE RMSE
0.2
0.22
0.24
0.26
0.28
0.3
0.32
0.34
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
PreFiltering DCR DCW uSplitting iSplitting uiSplitting
Food
Recall NDCG

78
References
 Y Zheng, R Burke, B Mobasher. Differential context relaxation for
context-aware travel recommendation. EC-Web, 2012
 Y Zheng, R Burke, B Mobashe. Optimal feature selection for
context-aware recommendation using differential relaxation.
CARS@ACM RecSys, 2012
 Y Zheng, R Burke, B Mobasher. Recommendation with differential
context weighting. ACM UMAP, 2013
 Context-based splitting of item ratings in collaborative filtering.
ACM RecSys, 2009
 Alan Said and Ernesto W. De Luca. Inferring Contextual User
Profiles – Improving Recommender Performance. CARS@ACM
RecSys, 2011
 Yong Zheng, Robin Burke, Bamshad Mobasher. Splitting approaches
for context-aware recommendation: An empirical study. ACM SAC,
2014

Contextual Modeling
• Independent Contextual Modeling
e.g., Tensor Factorization
• Dependent Contextual Modeling
1).Deviation-Based Approach
2).Similarity-Based Approach

80
Independent Contextual Modeling
(Tensor Factorization)

81
• Tensor Factorization
Multi-dimensional space: Users × Items × Contexts  Ratings
Each context variable is
modeled as an
individual and
independent dimension
in addition to user &
item dims.
Thus we can create a
multidimensional
space, where rating is
the value in the space.

82
• Tensor Factorization (Optimization)
1).By CANDECOMP/PARAFAC (CP) Decomposition

83
• Tensor Factorization (Optimization)
2).By Tucker Decomposition

84
• Tensor Factorization
Pros: Straightforward, easily to incorporate contexts into the model
Cons: 1). Ignore the dependence between contexts and user/item dims
2). Increased computational cost if more context dimensions
There are some research working on efficiency improvement on TF,
such as reusing GPU computations, and so forth…

85
• Dependence between Users/Items and Contexts
 User and Context, such as user splitting
 Item and Context, such as item splitting
For example, if a user can be splitted by time is weekend or
not. It tells this user is dependent with this context.
• Dependence between Every two Contexts
 Deviation-Based: rating deviation between two contexts
 Similarity-Based: similarity of rating behaviors in two contexts
Dependent Contextual Modeling

86
• Notion: Contextual Rating Deviation (CRD)
CRD how user’s rating is deviated from context c1 to c2?
CRD(D1) = 0.5  Users’ rating in Weekday is generally higher than
users’ rating at Weekend by 0.5
CRD(D2) = -0.1  Users’ rating in Cinema is generally lower than
users’ rating at Home by 0.1
Deviation-Based Contextual Modeling
Context D1: Time D2: Location
c1 Weekend Home
c2 Weekday Cinema
CRD(Di) 0.5 -0.1

87
• Notion: Contextual Rating Deviation (CRD)
CRD how user’s rating is deviated from context c1 to c2?
Assume Rating (U, T, c1) = 4
Predicted Rating (U, T, c2) = Rating (U, T, c1) + CRDs
= 4 + 0.5 -0.1 = 4.4
c1 Weekend Home
c2 Weekday Cinema
CRD(Di) 0.5 -0.1

88
• Build a deviation-based contextual modeling approach
Assume Ø is a special situation: without considering context
Assume Rating (U, T, Ø) = Rating (U, T) = 4
Predicted Rating (U, T, c2) = 4 + 0.5 -0.1 = 4.4
Ø UnKnown UnKnown
c2 Weekday Cinema
CRD(Di) 0.5 -0.1
In other words, F(U, T, C) = P(U, T) + 𝑖=0
𝑁
𝐶𝑅𝐷(𝑖)

89
• Build a deviation-based contextual modeling approach
Note: P(U, T) could be a rating prediction by any traditional
recommender systems, such as matrix factorization
Simplest model: F(U, T, C) = P(U, T) + 𝑖=0
𝑁
𝐶𝑅𝐷(𝑖)
User-personalized model: F(U, T, C) = P(U, T) + 𝑖=0
𝑁
𝐶𝑅𝐷(𝑖, 𝑈)
Item-personalized model: F(U, T, C) = P(U, T) + 𝑖=0
𝑁
𝐶𝑅𝐷(𝑖, 𝑇)

90
• Context-aware Matrix Factorization (CAMF)
By Linas Baltrunas, et al., ACM RecSys 2011
BiasedMF in Traditional RS:
Global Average Rating User bias Item Bias User-Item interaction
CAMF_C Approach:
CAMF_CU Approach:
CAMF_CI Approach:

91
• Contextual Sparse Linear Method (CSLIM)
By Yong Zheng, et al., ACM RecSys 2014
Rating Prediction in ItemKNN:
Score Prediction in SLIM:

92
• Contextual Sparse Linear Method (CSLIM)
By Yong Zheng, et al., ACM RecSys 2014
SLIM:
CSLIM_C:
CSLIM_CU:
CSLIM_CI:

93
• Top-10 Recommendation on the Japan Food Data
0.25
0.26
0.27
0.28
0.29
0.3
0.31
0.32
0.33
0.34
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
CAMF_C CAMF_CI CAMF_CU CSLIM_C CSLIM_CI CSLIM_CU
Recall NDCG

94
• Build a similarity-based contextual modeling approach
Assume Ø is a special situation: without considering context
Assume Rating (U, T, Ø) = Rating (U, T) = 4
Predicted Rating (U, T, c2) = 4 × Sim(Ø, c2)
Similarity-Based Contextual Modeling
Ø UnKnown UnKnown
c2 Weekday Cinema
Sim(Di) 0.5 0.1
In other words, F(U, T, C) = P(U, T) × Sim(Ø, C)

95
• Challenge: how to model context similarity, Sim(c1,c2)
We propose three representations:
• Independent Context Similarity (ICS)
• Latent Context Similarity (LCS)
• Multidimensional Context Similarity (MCS)

96
• Sim(c1, c2): Independent Context Similarity (ICS)
𝑆𝑖𝑚 c1, 𝑐2 = 𝑖=1
𝑁
𝑠𝑖𝑚(𝐷𝑖) = 0.5 × 0.1 = 0.05
c1 Weekend Home
c2 Weekday Cinema
Sim(Di) 0.5 0.1
𝐺𝑒𝑛𝑒𝑟𝑎𝑙𝑙𝑦, 𝐼𝑛 𝐼𝐶𝑆: 𝑆𝑖𝑚 c1, 𝑐2 = 𝑖=1
𝑁
𝑠𝑖𝑚(𝐷𝑖)
Weeend Weekday Home Cinema
Weekend 1 b — —
Weekday a 1 — —
Home — — 1 c
Cinema — — d 1

97
• Sim(c1, c2): Latent Context Similarity (LCS)
In training, we learnt (home, cinema), (work, cinema)
In testing, we need (home, work)
𝐺𝑒𝑛𝑒𝑟𝑎𝑙𝑙𝑦, 𝐼𝑛 𝐿𝐶𝑆: 𝑆𝑖𝑚 c1, 𝑐2 = 𝑖=1
𝑁
𝑆𝑖𝑚 𝐷𝑖 = 𝑑𝑜𝑡𝑃𝑟𝑜𝑑𝑢𝑐𝑡 (𝑉𝑖1, 𝑉𝑖2)
f1 f2 … … … … fN
home 0.1 -0.01 … … … … 0.5
work 0.01 0.2 … … … … 0.01
cinema 0.3 0.25 … … … … 0.05
Vector
Representation

98
• Sim(c1, c2): Multidimensional Context Similarity (MCS)
𝐼𝑛 𝑀𝐶𝑆: 𝐷𝑖𝑠𝑆𝑖𝑚 c1, 𝑐2 = distance between two point

99
Similarity-Based CAMF:
Similarity-Based CSLIM:
General Similarity-Based CSLIM:
• Build algorithms based on traditional recommender
𝐼𝑛 𝐼𝐶𝑆: 𝑆𝑖𝑚 c1, 𝑐2 = 𝑖=1
𝑁
𝐼𝑛 𝐿𝐶𝑆: 𝑆𝑖𝑚 c1, 𝑐2 = 𝑖=1
𝑁
𝑠𝑖𝑚 𝐷𝑖 , 𝑠𝑖𝑚 𝐷𝑖 𝑖𝑠 𝑑𝑜𝑡𝑃𝑟𝑜𝑑𝑢𝑐𝑡
𝐼𝑛 𝑀𝐶𝑆: 𝐷𝑖𝑠𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 𝑖𝑠 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒, 𝑠𝑢𝑐ℎ 𝑎𝑠 𝐸𝑢𝑐𝑙𝑖𝑑𝑒𝑎𝑛 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒

100
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
CAMF_ICS CAMF_LCS CAMF_MCS CSLIM_ICS CSLIM_LCS CSLIM_MCS
Recall NDCG

101
Overall Comparison among Best Performers
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
TF FM Splitting CAMF_Dev CSLIM_Dev CAMF_Sim CSLIM_Sim
Overall Comparison of Best Performers
Recall NDCG

102
References
 A Karatzoglou, X Amatriain, L Baltrunas, N Oliver. Multiverse
Recommendation: N-dimensional Tensor Factorization for
Context-aware Collaborative Filtering. ACM RecSys, 2010
 L Baltrunas, B Ludwig, F Ricci. Matrix factorization techniques for
context aware recommendation. ACM RecSys, 2011
 Y Zheng, B Mobasher, R Burke. CSLIM: Contextual SLIM
Recommendation Algorithms. ACM RecSys, 2014
 Y Zheng, B Mobasher, R Burke. Deviation-Based Contextual SLIM
Recommenders. ACM CIKM, 2014
 Y Zheng, B Mobasher, R Burke. Integrating Context Similarity
with Sparse Linear Recommendation Model. ACM UMAP, 2015
 Y Zheng, B Mobasher, R Burke. Similarity-Based Context-aware
Recommendation. WISE, 2015

Outline
Context Modeling
103

105
Other Challenges
• There could be many other challenges in CARS:
Numeric v.s. Categorical Context Information
Cold-start Problems in CARS
Recommendation Explanations by CARS
New User Interfaces and Interactions
New and More Applications
New Recommendation Opportunities

106
Other Challenges: Numeric Context
• List of Categorical Context
Time: morning, evening, weekend, weekday, etc
Location: home, cinema, work, party, etc
Companion: family, kid, partner, etc
• How about numeric context
Time: 2016, 6:30 PM, 2 PM to 6 PM (time-aware recsys)
Temperature: 12°C, 38°C
Principle component by PCA: numeric values
Other numeric values in context, how to develop CARS?

107
Other Challenges: Cold-Start
• Cold-start Problems
Cold-start user: no rating history by this user
Cold-start item: no rating history on this item
Cold-start context: no rating history within this context
• Solution: Hybrid Method by Matthias Braunhofer, et al.

108
Other Challenges: Explanation
• Recommendation Using social networks (By Netflix)
The improvement is not significant;
Unless we explicitly explain it to the end users;
• Recommendation Using context (Open Research)
Similar thing could happen to context-aware recsys;
How to use contexts to explain recommendations;
How to design new user interface to explain;
How to merge CARS with user-centric evaluations;

109
Other Challenges: User Interface
• Potential Research Problems in User Interface
New UI to collect context;
New UI to interact with users friendly and smoothly;
New UI to explain context-aware recommendation;
New UI to avoid debates on user privacy;
User privacy problems in context collection & usage

110
Other Challenges: New Applications
• More applications are in demand:
Not only e-commerce, movies, music, etc
Tourism: Trip planner, Traffic analyzer and planner
MOOC: online learning via different characteristics
Life Long: Digital health, daily activity tracker
Shared Economy: Uber, Airbnb

111
Other Challenges: New Opportunity
• CARS enable new recommendation opportunities
Context Suggestion
We will introduce later in this tutorial.

112
CARSKit: Recommendation Library

113
Recommendation Library
• Motivations to Build a Recommendation Library
1). Standard Implementations for popular algorithms
2). Standard platform for benchmark or evaluations
3). Helpful for both research purpose and industry practice
4). Helpful as tools in teaching and learning

114
Recommendation Library
There are many recommendation library for traditional recommendation.
Users × Items Ratings

115
CARSKit: A Java-based Open-source
Context-aware Recommendation Library
CARSKit: https://github.com/irecsys/CARSKit
Users × Items × Contexts Ratings
User Guide: http://arxiv.org/abs/1511.03780

116
CARSKit: A Short User Guide
1. Download the JAR library, i.e., CARSKit.jar
2. Prepare your data
3. Setting: setting.conf
4. Run: java –jar CARSKit.jar –c setting.conf

117
Sample of Outputs: Data Statistics

118
Sample of Outputs:
1). Results by Rating Prediction Task
Final Results by CAMF_C, MAE: 0.714544, RMSE: 0.960389, NAME: 0.178636,
rMAE: 0.683435, rRMSE: 1.002392, MPE: 0.000000, numFactors: 10, numIter: 100,
lrate: 2.0E-4, maxlrate: -1.0, regB: 0.001, regU: 0.001, regI: 0.001, regC: 0.001,
isBoldDriver: true, Time: '00:01','00:00‘
2). Results by Top-N Recommendation Task
Final Results by CAMF_C, Pre5: 0.048756,Pre10: 0.050576, Rec5: 0.094997, Rec10:
0.190364, AUC: 0.653558, MAP: 0.054762, NDCG: 0.105859, MRR: 0.107495,
numFactors: 10, numIter: 100, lrate: 2.0E-4, maxlrate: -1.0, regB: 0.001, regU:
0.001, regI: 0.001, regC: 0.001, isBoldDriver: true, Time: '00:01','00:00'

Outline
Other Challenges
119

Context Suggestion
121
• Task: Suggest a list of contexts to users (on items)
Context Rec
Contextual RecTraditional Rec

122
Context Suggestion: Motivations
• Motivation-1: Maximize user experience
User Experience (UX) refers to a person's emotions and
attitudes about using a particular product, system or
service.

123
• To maximize user experience (UX)
Example: Evolution in Retails
Product

124
Product
Service

125
Product
Service
Context

126
• Motivation-1: Maximize user experience
It is not enough to recommend items only

127
• Motivation-2: Contribute to Context Collection
Predefine contexts and suggest them to users

128
• Motivation-3: Connect with Context-aware RecSys
User’s actions on context is a context-query to system

129
Context Suggestion: Applications
• There could be many potential applications:

130

131
Real Examples: Google Music
Input is
a single user

132
Input is
<user, item>

133

134
Input is user;
Output is
kid + movies

135
Input is user;
Output is
day + books
As a gift for Mother’s Day

136

137
As a gift for Mother’s Day

138
A list of item recommendation
associated with context

139
Context Suggestion
• Task: Suggest a list of appropriate contexts to users
For example: Where should I watch movie Life Of PI
• Timeline
In 2008, proposed by Tutorial at ACM RecSys 2008
In 2010, first attempt by Linas et al. ACM RecSys 2010
In 2014, formal discussion by Yong et al., IEEE/ACM WI 2014
In 2015, proposal more applications by Yong, IEEE ICDM 2015
In 2016, working on new solutions for related problems

140
Context Suggestion
• There could be many applications, we focus on two tasks
1).UI-Oriented Context Suggestion
Task: suggest contexts to <user, item>
Example: time & location to watch movie Life of PI
2). User-Oriented Context Suggestion
Task: suggest contexts to each user
Example: Google Music, Pandora, Youtube, etc

141
UI-Oriented Context Suggestion
Solution 1).Multilabel classification (MLC)
 KNN classifiers by Linas et al., ACM RecSys 2010
 Other MLC by Zheng et al., IEEE/ACM WI, 2014
1). Binary Classification
Question: Is this an apple? Yes or No.
2). Multi-class Classification
Question: Is this an apple, banana or orange?
3). Multi-label Classification
Use appropriate words to describe it:
Red, Apple, Fruit, Tech, Mac, iPhone
In our case, we use user and item information
as inputs and features to learn label predictions
Color, Shape, Weight, Origin,
Taste, Price, Vitamin c

142
How MLC works as a solution for context suggestion?
Simply, each context condition is viewed as an individual label.

143
Users (content) + items (content) as features in MLC
User info: ID, gender, age, nationality, language, etc
Item info: genre, director, composer, year, language, etc
Input: Users (content) + items (content) + RatingIsGood
Output: A list of contexts as predicted labels
Note: we can set a rating threshold to determine “Good”
Color, Shape, Weight, Origin,
Taste, Price, Vitamin c

144
MLC Process:
1). Assign a MLC algorithm
1.1). Transformation algorithms
MLC task is a composition of binary/multi-class CL tasks
1.2). Adaptation algorithms, such as MLKNN
2). Assign a classifier
Any traditional classifiers can be used in 1.1), e.g., trees,
KNN, Naïve Bayes, SVM, ensemble classifiers, etc

145
More details: Transformation MLC Algorithms
a).Binary Relevance
Predict the binary value for each label first;
And finally aggregate them together;
Pros: simple and straightforward
Cons: ignore label correlations or dependencies

146
b).Classifier Chains
Predict the binary value for labels in a sequence; former
predicted label will be used as feature to infer next one.
Pros: We take label correlation into consideration
Cons: Next prediction could be wrong if prior is wrong

147
c).Label Powerset
Each subset of labels can be viewed as single label
Pros: We take label correlation into consideration
Cons: More labels, more subset
More computational costs

148
d).RAkEL (Random K-Labelset)
It is an optimizer derived from Label Powerset;
It randomly selects a K-labelset instead of all subsets
Pros: Alleviate costs in Label Powerset
Cons: Local minimum
Only efficient for data with large scale of labels

149
MLC Java-based Open-source Toolkit
a). Mulan
http://mulan.sourceforge.net
b). MEKA
http://meka.sourceforge.net
New: with GUI!! An extension to WEKA

150
Solution 2).Context-aware Recommendation
We can reuse CARS algorithms to recommend contexts.
For example, Tensor Factorization
We put all conditions into a single dimension: context
Then we create 3D space: user, item, context
We recommend contexts for each <user, item>
Other CARS algorithms can also be applied

151
User-Oriented Context Suggestion
It can be viewed as a process of context acquisition
But recommendation task is still involved in it.

152
User-Oriented Context Suggestion
There could be several potential solutions:
1). Most popular or user-popular context suggestion;
2). Most recent or user-recent context suggestion;
3). Collaborative suggestion based on other users’ tastes;
4). Reuse context-aware recommendation algorithms;

153
Context Suggestion: Challenges
It is still a novel and emerging research direction. There are
several challenges to be solved:
1). Evaluations
We do not have user’s taste on context
2). Solutions
Is personalized required? Any personalized solutions?
Popular suggestion is a good solution?
3). User Interface
How to build appropriate UI to interact with users

154
References
 L Baltrunas, M Kaminskas, F Ricci, et al. Best usage context prediction
for music tracks. CARS@ACM RecSys, 2010
 Y Zheng, B Mobasher, R Burke. Context Recommendation Using Multi-
label Classification. IEEE/WIC/ACM WI, 2014
 Y Zheng. Context Suggestion: Solutions and Challenges. ICDM
Workshop, 2015
 Y Zheng. Context-Driven Mobile Apps Management and
Recommendation. ACM SAC, 2016

Outline
Other Challenges
155

Topics in this Tutorial
• Traditional Recommendation
e.g., Give me a list of recommended movies to watch
e.g., Give me a list of recommended movies to watch, if
Time & Location: at weekend and in cinema
Companion: with girlfriend v.s. with Kids
The best time/location to watch movie “Life of PI”
156

Details in this Tutorial
Other Challenges
• Context Suggestion: App, Solution and Challenges
157

Future Research
Treat Numeric Context Information
Cold-start Problems in CARS
Recommendation Explanation by Context
User Interface and More applications by CARS
Data collection for evaluations
Examine different algorithms on real-world data
Design new user interface and applications
158

List of Tutorials/Keynotes about CARS
 Gedas Adomavicius and Alex Tuzhilin, “Context-aware Recommender
Systems”, In ACM RecSys, 2008
 Bamshad Mobasher, “Contextual User Modeling for Recommendation”, In
CARS Workshop@ACM RecSys, 2010
 Francesco Ricci, “Contextualizing Recommendations”, In CARS Workshop@
ACM RecSys, 2012
 Bamshad Mobasher, “Context-aware User Modeling for Recommendation”,
ACM UMAP, 2013
 Bamshad Mobasher, “Context-aware Recommendation”, ACM KDD, 2014
 Bamshad Mobasher, “Context Awareness and Adaptation in
Recommendation”, DMRS Workshop, 2015
 Yong Zheng, “Context In Recommender Systems”, ACM SAC, 2016
159

Tutorial: Context In
Recommender Systems
Yong Zheng
DePaul University, Chicago
The 31st ACM Symposium on Applied Computing, Pisa, Italy, 2016

Tutorial: Context In Recommender Systems

More Related Content

What's hot

Viewers also liked

Similar to Tutorial: Context In Recommender Systems

More from YONG ZHENG

Recently uploaded

Tutorial: Context In Recommender Systems