1 of 10

Agenda

Related Research

A social Recommender System using Item Asymmetric Correlation

Presented By

Data and Evaluation

Introduction

Conclusion

Devised Approach

Arghavan Moradi¹, Hadi Tabatabaei¹, Mehregan Mahdavi²

1 Shahid Beheshti University, Tehran, Iran

2 University of New South Wales, Sydney, Australia

2 of 10

Introduction

Problem Definition

Recommender systems have been one of the most prominent information filtering
The growing number of users and the variety of
items cause problems
they suffer from two major problems: Cold Start, Data Sparsity
Data Sparsity decreases the accuracy of recommendation

Different Datasets fusion
Matrix Factorization is one of the technique to combine different datasets
The secondary dataset can be built by users or extract from latent relationship

Solution Domain

Using user friendship data
Using user membership data
Using explicit relationships
Using implicit relationship

Related Researches

2

Spring 2019

3 of 10

Weaknesses

Weaknesses of related researches

Reduced the effect of the sparsity by combining friendship
the friendship between two users does not mean that their tastes are close
users may add their kin and family members to their list of friends

Friendship

VIEW DETAILS

Membership

VIEW DETAILS

Reduced the effect of the sparsity by combining membership
Membership data is not available in all media
Depending on users interaction
It has sparsity problem

Explicit relationship

VIEW DETAILS

item relationships are more available and trustable than social relationships
Explicit relationship between items is a side information
It is again dependent on the application and environment

3

Spring 2019

4 of 10

Extracting latent relationship between items using an asymmetric similarity

Item Asymmetric Correlation

1

3

Mapping empty cells by using two mapping models: SGD and ALS

Prediction level

2

The fusion of rating matrix (main dataset) and IAC output (built dataset)

Matrix Factorization technique

4

Estimating the accuracy level of predictions by using MAE and RMSE metrics

Accuracy Evaluation

Approach

Devised Approach

4

Spring 2019

5 of 10

IAC

Item Asymmetric Correlation

User/item	Item1	Item2	Item3	Item4	Item5
user1	1	1	-	-	-
user2	-	1	-	-	1
user3	0	0	1	1	0
user4	1	1	-	-	-
user5	-	0	-	-	-
user6	-	-	1	1	-

Cosine Similarity

Item Asymmetric Correlation

5

Spring 2019

6 of 10

MF

Matrix Factorization

m Items

n Users

m Items

MF

i

6

Spring 2019

7 of 10

Dataset

Dataset and Evaluation Metrics

50,000				The Number of users
3000				The Number of Items
	To Item		By User	The Min Number of Rating
	1		1
	To Item	By User		The Max Number of Rating
	34,023	1425
	To Item	By User		The AVG Number of Rating
	1299	77.94

User.getArtistTrack
Group.getMembership
User.getFriendList

% of Data Sparsity	% of Training set	Data
95.71%	60%	Train-60
96.57%	70%	Train-70
97.42%	80%	Train-80
98.88%	90%	Train-90

Training set and Sparsity Distribution

: Mean Absolute Error

: Root Mean Square Error

7

Spring 2019

8 of 10

Evaluation

8

Spring 2019

9 of 10

Conclusion

A Novel MF method to reduce the impact of Sparsity in accuracy of suggestions

Extract implicit relationship between items

Using Asymmetric Correlation to extract latent relationship in items

The presented method is applicable

where the additional datasets are not available

How can we decrease the time complexity of this type of method

Extract the implicit relationship between users

9

Spring 2019

10 of 10

10

Spring 2019