SlideShare a Scribd company logo
1
EXTRA : Integrating External Knowledge into Multimodal
Hashtag Recommendation System
1 Department of Artificial Intelligence, The Catholic University of Korea, Bucheon, South Korea
2 Department of Data Science, The Catholic University of Korea, Bucheon, South Korea
3 Department of Computer Science and Engineering, Korea University, Seoul, South Korea
Hyun-Sik Won1∗, Su-Min Roh1∗, Dohyun Kim3, Min-Ji Kim1, Huiju Kim3, and Kang-Min Kim12
* Equal contribution
Background
2
• Microblogging Platforms
Hashtags allow people to promote their own posts or categorize posts into specific topics.
Introduction
• Hashtag Recommendation Task
: Recommend appropriate hashtags related to the post.
• Previous Works
: Recommend hashtags by utilizing explicit information in the post (i.e., text, image, etc.).
: Recently, each user’s hashtag habits1 from the microblog are also utilized to recommend personalized hashtags.
3
1. Chen, Yu-Chi, et al. "Tagnet: triplet-attention graph networks for hashtag recommendation." IEEE Transactions on Circuits and Systems for Video Technology 32.3 (2021): 1148-1159.
Introduction
• Hashtag Recommendation Task
: Recommend appropriate hashtags related to the post.
• Previous Works
: Recommend hashtags by utilizing explicit information in the post (i.e., text, image, etc.).
: Recently, each user’s hashtag habits1 from the microblog are also utilized to recommend personalized hashtags.
However, these methods only utilize the information found in the microblog.
 This method may have difficulty understanding the contextual information of the post.
4
1. Chen, Yu-Chi, et al. "Tagnet: triplet-attention graph networks for hashtag recommendation." IEEE Transactions on Circuits and Systems for Video Technology 32.3 (2021): 1148-1159.
Introduction
• External Knowledge in Hashtag Recommendation
5
Introduction
6
• External Knowledge in Hashtag Recommendation
Does not appear directly
in the text or image.
Introduction
7
However, #dogtraining can be predicted based on external knowledge that dogs are typically trained.
• External Knowledge in Hashtag Recommendation
Does not appear directly
in the text or image.
Introduction
8
However, #dogtraining can be predicted based on external knowledge that dogs are typically trained.
• External Knowledge in Hashtag Recommendation
Does not appear directly
in the text or image.
How can we integrate the external knowledge to recommend
more relevant hashtags for the post?
Related Work
• Open Directory Project (ODP)
: This is a comprehensive directory of the World Wide Web, constructed and maintained by a community of
volunteer editors.
: It organizes web pages into the most related categories and subcategories.
(e.g., Health/Animal/Mammals/Dogs)
• There have been previous studies utilizing ODP categories as an external knowledge.
: The study1 trained a classifier to predict the category based on the description of the webpage.
9
1. Kim et al., “meChat: In-device Personal Assistant for Conversational Photo Sharing” IEEE Internet Computing 2019
Methodology
10
• Overview of the proposed model
Methodology
11
• Overview of the proposed model
① External Knowledge
Feature Extraction Model
Methodology
12
① External Knowledge
Feature Extraction Model
② Integrating External
Knowledge into
Multimodal Hashtag
Recommendation
System (EXTRA)
• Overview of the proposed model
Methodology
① External Knowledge Feature Extraction Model
13
Methodology
① External Knowledge Feature Extraction Model
14
Caption
Generator
A woman is holding a
certificate with a dog
Methodology
① External Knowledge Feature Extraction Model
15
ODP-based Classifier
(ODP-I)
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Recreation/Pets/Dogs/Breeds/Working_Group
Caption
Generator
A woman is holding a
certificate with a dog
Methodology
① External Knowledge Feature Extraction Model
16
ODP-based Classifier
(ODP-I)
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Recreation/Pets/Dogs/Breeds/Working_Group
Caption
Generator
A woman is holding a
certificate with a dog
Methodology
② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA)
17
Image of the post Text of the post Relevant Category of the Image
Algy is our other graduate. Today, Algy aced the see-saw and
weave poles so impressively with this obstacle awareness..
Recreation/Pets/Dogs/Breeds/Working_Group
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Methodology
② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA)
18
Image of the post Text of the post Relevant Category of the Image
…
Image Encoder
Algy is our other graduate. Today, Algy aced the see-saw and
weave poles so impressively with this obstacle awareness..
Recreation/Pets/Dogs/Breeds/Working_Group
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Methodology
② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA)
19
Image of the post Text of the post Relevant Category of the Image
Dogs Training Working Group
awareness
algy other …
is our obstacle
…
Image Encoder Text Encoder
Algy is our other graduate. Today, Algy aced the see-saw and
weave poles so impressively with this obstacle awareness..
Recreation/Pets/Dogs/Breeds/Working_Group
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Methodology
② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA)
20
Image of the post Text of the post Relevant Category of the Image
Dogs Training Working Group
awareness
algy other …
is our obstacle
…
Image Encoder Text Encoder
Multimodal Encoder
Algy is our other graduate. Today, Algy aced the see-saw and
weave poles so impressively with this obstacle awareness..
Recreation/Pets/Dogs/Breeds/Working_Group
Health/Animal/Mammals/Dogs
Recreation/Pets/Dogs/Training
Dataset
21
• Open Directory Project Dataset (for extracting external knowledge)
- # Web pages : 60,710
- # Categories : 2,531
• MaCon Dataset (for hashtag recommendation)
- # Posts : 624,520
- # Hashtags : 3,896
- # Average of hashtags per post : 9.3
Baselines
22
• Co-Attention (Co-AT)
: the model that sequentially generates text and image feature vectors from a post using a co-attention mechanism.
• MaCon
: the model that applies a parallel co-attention mechanism to combine the text and image features with users’
tagging habits.
• TAGNet
: the model that integrates text, visual, and user habit features through a triplet attention module by constructing
visual similarity graphs.
SOTA
Settings
23
• Implementation Details
- Model : Pre-trained FLAVA1
- Optimizer : AdamW
- Learning Rate : 1e-5 (10% warmup, cosine decay)
- Batch Size : 64
• Evaluation Metric
: Top-k Precision, Recall, F1-score
1. Singh, Amanpreet, et al. "Flava: A foundational language and vision alignment model." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.
Experiments
24
• Comparison with Baselines
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Top-1 Top-5 Top-9
F1-score
Co-AT MaCon TAGNet EXTRA (Ours)
 EXTRA achieved a 39.78% improvement over the state-of-the-art model TAGNet in terms of the Top-5 F1-Score.
Experiments
25
• Comparison Between External Knowledge From Each Modality
 The models that have integrated external knowledge perform better than those that have not.
Conclusions
26
• We propose EXTRA, a novel multimodal hashtag recommendation system that combines text and image with
external knowledge in a Transformer-based architecture.
• We employ the most relevant categories extracted from the image of the post as external knowledge using the
ODP-based classifier.
• Our approach outperformed the existing state-of-the-art methods on the MaCon dataset, demonstrating its
effectiveness in handling multimodal information for the hashtag recommendation task.
27
Experiments
28
• Illustrations of Recommended Hashtags
Predicted
Hashtag
(w/o ODP-I)
#lax #makeportraits #mountainlife
#plasticpollution #pursuitofportraits #spa
#greenery #travel #detail #blackandwhite
Predicted
Hashtag
(w/ ODP-I)
#makeportraits #plasticpollution
#mountainlife #adventure #lax
#pursuitofportraits #travel
#greatnorthcollective #nature
Experiments
29
• Comparison between Each Modality
Experiments
30
• Comparison between The Number of Subcategories

More Related Content

What's hot

Analisis lexico automatas i
Analisis lexico automatas iAnalisis lexico automatas i
Analisis lexico automatas i
rubiyanetvalenciavelazquez
 
Matplotlib Review 2021
Matplotlib Review 2021Matplotlib Review 2021
Matplotlib Review 2021
Bhaskar J.Roy
 
Instruction set-of-8085
Instruction set-of-8085Instruction set-of-8085
Instruction set-of-8085saleForce
 
8051 MICROCONTROLLER
8051 MICROCONTROLLER 8051 MICROCONTROLLER
8051 MICROCONTROLLER
THANDAIAH PRABU
 
00 introducción a Python
00 introducción a Python00 introducción a Python
00 introducción a Python
Juan Rodríguez
 
Sumadores,codificadores, decodificadores,multiplexores y demultiplexores
Sumadores,codificadores, decodificadores,multiplexores y demultiplexoresSumadores,codificadores, decodificadores,multiplexores y demultiplexores
Sumadores,codificadores, decodificadores,multiplexores y demultiplexores
Jose Tomas Rodriguez Henriquez
 
Paridad cruzada
Paridad cruzadaParidad cruzada
Paridad cruzada
Pablo Chumania
 
Los circuitos integrados
Los circuitos integradosLos circuitos integrados
Los circuitos integrados
Jomicast
 
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
JalpaMaheshwari1
 
Chapter 2: Boolean Algebra and Logic Gates
Chapter 2: Boolean Algebra and Logic GatesChapter 2: Boolean Algebra and Logic Gates
Chapter 2: Boolean Algebra and Logic Gates
Er. Nawaraj Bhandari
 
Diferentes tipos de arduino
Diferentes tipos de arduinoDiferentes tipos de arduino
Diferentes tipos de arduino
Santiago Ramirez Castaño
 
Diseño de Algoritmos
Diseño de AlgoritmosDiseño de Algoritmos
Diseño de Algoritmos
LuisDiaz863
 
11 Unit 1 Problem Solving Techniques
11  Unit 1 Problem Solving Techniques11  Unit 1 Problem Solving Techniques
11 Unit 1 Problem Solving Techniques
Praveen M Jigajinni
 
Explain Half Adder and Full Adder with Truth Table
Explain Half Adder and Full Adder with Truth TableExplain Half Adder and Full Adder with Truth Table
Explain Half Adder and Full Adder with Truth Table
elprocus
 
Numpy tutorial
Numpy tutorialNumpy tutorial
Numpy tutorial
HarikaReddy115
 
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
Brayan Galan
 
Ejercicios resueltos
Ejercicios resueltosEjercicios resueltos
Ejercicios resueltosColono
 
The Difference between HTML, XHTML & HTML5 for Beginners
The Difference between HTML, XHTML & HTML5 for BeginnersThe Difference between HTML, XHTML & HTML5 for Beginners
The Difference between HTML, XHTML & HTML5 for Beginners
Rasin Bekkevold
 

What's hot (20)

Analisis lexico automatas i
Analisis lexico automatas iAnalisis lexico automatas i
Analisis lexico automatas i
 
Modelos.diagramabloques
Modelos.diagramabloquesModelos.diagramabloques
Modelos.diagramabloques
 
Matplotlib Review 2021
Matplotlib Review 2021Matplotlib Review 2021
Matplotlib Review 2021
 
Instruction set-of-8085
Instruction set-of-8085Instruction set-of-8085
Instruction set-of-8085
 
8051 MICROCONTROLLER
8051 MICROCONTROLLER 8051 MICROCONTROLLER
8051 MICROCONTROLLER
 
00 introducción a Python
00 introducción a Python00 introducción a Python
00 introducción a Python
 
Sumadores,codificadores, decodificadores,multiplexores y demultiplexores
Sumadores,codificadores, decodificadores,multiplexores y demultiplexoresSumadores,codificadores, decodificadores,multiplexores y demultiplexores
Sumadores,codificadores, decodificadores,multiplexores y demultiplexores
 
Paridad cruzada
Paridad cruzadaParidad cruzada
Paridad cruzada
 
Los circuitos integrados
Los circuitos integradosLos circuitos integrados
Los circuitos integrados
 
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
Introduction to Half and Full Adder Circuit - Part 01 | Digital Logic Design |
 
Chapter 2: Boolean Algebra and Logic Gates
Chapter 2: Boolean Algebra and Logic GatesChapter 2: Boolean Algebra and Logic Gates
Chapter 2: Boolean Algebra and Logic Gates
 
Diferentes tipos de arduino
Diferentes tipos de arduinoDiferentes tipos de arduino
Diferentes tipos de arduino
 
Diseño de Algoritmos
Diseño de AlgoritmosDiseño de Algoritmos
Diseño de Algoritmos
 
11 Unit 1 Problem Solving Techniques
11  Unit 1 Problem Solving Techniques11  Unit 1 Problem Solving Techniques
11 Unit 1 Problem Solving Techniques
 
Pld
PldPld
Pld
 
Explain Half Adder and Full Adder with Truth Table
Explain Half Adder and Full Adder with Truth TableExplain Half Adder and Full Adder with Truth Table
Explain Half Adder and Full Adder with Truth Table
 
Numpy tutorial
Numpy tutorialNumpy tutorial
Numpy tutorial
 
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
FAMILIAS LÓGICAS DE CIRCUITOS INTEGRADOS.
 
Ejercicios resueltos
Ejercicios resueltosEjercicios resueltos
Ejercicios resueltos
 
The Difference between HTML, XHTML & HTML5 for Beginners
The Difference between HTML, XHTML & HTML5 for BeginnersThe Difference between HTML, XHTML & HTML5 for Beginners
The Difference between HTML, XHTML & HTML5 for Beginners
 

Similar to EXTRA: Integrating External Knowledge into Multimodal Hashtag Recommendation System

Teacher training material
Teacher training materialTeacher training material
Teacher training material
Vikram Parmar
 
From Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
From Chalkboards to Chatbots: A Deep Dive into AI for Education WorkshopFrom Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
From Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
Stella Lee
 
ODSC APAC 2022 - Explainable AI
ODSC APAC 2022 - Explainable AIODSC APAC 2022 - Explainable AI
ODSC APAC 2022 - Explainable AI
Aditya Bhattacharya
 
Andrew NG machine learning
Andrew NG machine learningAndrew NG machine learning
Andrew NG machine learning
ShareDocView.com
 
Interactive Machine Learning Appendix
Interactive  Machine Learning AppendixInteractive  Machine Learning Appendix
Interactive Machine Learning Appendix
Zitao Liu
 
2017 10-10 (netflix ml platform meetup) learning item and user representation...
2017 10-10 (netflix ml platform meetup) learning item and user representation...2017 10-10 (netflix ml platform meetup) learning item and user representation...
2017 10-10 (netflix ml platform meetup) learning item and user representation...
Ed Chi
 
Sentiment Analysis of Twitter Data
Sentiment Analysis of Twitter DataSentiment Analysis of Twitter Data
Sentiment Analysis of Twitter Data
Sumit Raj
 
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.comHABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
HABIB FIGA GUYE
 
Leveraging social media for training object detectors
Leveraging social media for training object detectorsLeveraging social media for training object detectors
Leveraging social media for training object detectors
Manish Kumar
 
CodeLess Machine Learning
CodeLess Machine LearningCodeLess Machine Learning
CodeLess Machine Learning
Sharjeel Imtiaz
 
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
Egyptian Engineers Association
 
Machine Learning ass. of tanumalakar.pdf
Machine Learning ass. of tanumalakar.pdfMachine Learning ass. of tanumalakar.pdf
Machine Learning ass. of tanumalakar.pdf
DYDF
 
Essential concepts for machine learning
Essential concepts for machine learning Essential concepts for machine learning
Essential concepts for machine learning
pyingkodi maran
 
H2O World - Intro to Data Science with Erin Ledell
H2O World - Intro to Data Science with Erin LedellH2O World - Intro to Data Science with Erin Ledell
H2O World - Intro to Data Science with Erin Ledell
Sri Ambati
 
AI AND EDUCATIO
AI AND EDUCATIOAI AND EDUCATIO
AI AND EDUCATIO
no0orcom
 
Data-X-Sparse-v2
Data-X-Sparse-v2Data-X-Sparse-v2
Data-X-Sparse-v2
Ikhlaq Sidhu
 
Introduction to Recommender System
Introduction to Recommender SystemIntroduction to Recommender System
Introduction to Recommender System
WQ Fan
 
Active learning
Active learningActive learning
Active learning
Akhilesh Ravi
 
Recsys 2016
Recsys 2016Recsys 2016
Recsys 2016
Mindaugas Zickus
 
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
Alok Singh
 

Similar to EXTRA: Integrating External Knowledge into Multimodal Hashtag Recommendation System (20)

Teacher training material
Teacher training materialTeacher training material
Teacher training material
 
From Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
From Chalkboards to Chatbots: A Deep Dive into AI for Education WorkshopFrom Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
From Chalkboards to Chatbots: A Deep Dive into AI for Education Workshop
 
ODSC APAC 2022 - Explainable AI
ODSC APAC 2022 - Explainable AIODSC APAC 2022 - Explainable AI
ODSC APAC 2022 - Explainable AI
 
Andrew NG machine learning
Andrew NG machine learningAndrew NG machine learning
Andrew NG machine learning
 
Interactive Machine Learning Appendix
Interactive  Machine Learning AppendixInteractive  Machine Learning Appendix
Interactive Machine Learning Appendix
 
2017 10-10 (netflix ml platform meetup) learning item and user representation...
2017 10-10 (netflix ml platform meetup) learning item and user representation...2017 10-10 (netflix ml platform meetup) learning item and user representation...
2017 10-10 (netflix ml platform meetup) learning item and user representation...
 
Sentiment Analysis of Twitter Data
Sentiment Analysis of Twitter DataSentiment Analysis of Twitter Data
Sentiment Analysis of Twitter Data
 
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.comHABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
HABIB FIGA GUYE {BULE HORA UNIVERSITY}(habibifiga@gmail.com
 
Leveraging social media for training object detectors
Leveraging social media for training object detectorsLeveraging social media for training object detectors
Leveraging social media for training object detectors
 
CodeLess Machine Learning
CodeLess Machine LearningCodeLess Machine Learning
CodeLess Machine Learning
 
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
 
Machine Learning ass. of tanumalakar.pdf
Machine Learning ass. of tanumalakar.pdfMachine Learning ass. of tanumalakar.pdf
Machine Learning ass. of tanumalakar.pdf
 
Essential concepts for machine learning
Essential concepts for machine learning Essential concepts for machine learning
Essential concepts for machine learning
 
H2O World - Intro to Data Science with Erin Ledell
H2O World - Intro to Data Science with Erin LedellH2O World - Intro to Data Science with Erin Ledell
H2O World - Intro to Data Science with Erin Ledell
 
AI AND EDUCATIO
AI AND EDUCATIOAI AND EDUCATIO
AI AND EDUCATIO
 
Data-X-Sparse-v2
Data-X-Sparse-v2Data-X-Sparse-v2
Data-X-Sparse-v2
 
Introduction to Recommender System
Introduction to Recommender SystemIntroduction to Recommender System
Introduction to Recommender System
 
Active learning
Active learningActive learning
Active learning
 
Recsys 2016
Recsys 2016Recsys 2016
Recsys 2016
 
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
Big Data Spain 2018: How to build Weighted XGBoost ML model for Imbalance dat...
 

Recently uploaded

María Carolina Martínez - eCommerce Day Colombia 2024
María Carolina Martínez - eCommerce Day Colombia 2024María Carolina Martínez - eCommerce Day Colombia 2024
María Carolina Martínez - eCommerce Day Colombia 2024
eCommerce Institute
 
0x01 - Newton's Third Law: Static vs. Dynamic Abusers
0x01 - Newton's Third Law:  Static vs. Dynamic Abusers0x01 - Newton's Third Law:  Static vs. Dynamic Abusers
0x01 - Newton's Third Law: Static vs. Dynamic Abusers
OWASP Beja
 
Announcement of 18th IEEE International Conference on Software Testing, Verif...
Announcement of 18th IEEE International Conference on Software Testing, Verif...Announcement of 18th IEEE International Conference on Software Testing, Verif...
Announcement of 18th IEEE International Conference on Software Testing, Verif...
Sebastiano Panichella
 
somanykidsbutsofewfathers-140705000023-phpapp02.pptx
somanykidsbutsofewfathers-140705000023-phpapp02.pptxsomanykidsbutsofewfathers-140705000023-phpapp02.pptx
somanykidsbutsofewfathers-140705000023-phpapp02.pptx
Howard Spence
 
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdfSupercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
Access Innovations, Inc.
 
Obesity causes and management and associated medical conditions
Obesity causes and management and associated medical conditionsObesity causes and management and associated medical conditions
Obesity causes and management and associated medical conditions
Faculty of Medicine And Health Sciences
 
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
Orkestra
 
Doctoral Symposium at the 17th IEEE International Conference on Software Test...
Doctoral Symposium at the 17th IEEE International Conference on Software Test...Doctoral Symposium at the 17th IEEE International Conference on Software Test...
Doctoral Symposium at the 17th IEEE International Conference on Software Test...
Sebastiano Panichella
 
Media as a Mind Controlling Strategy In Old and Modern Era
Media as a Mind Controlling Strategy In Old and Modern EraMedia as a Mind Controlling Strategy In Old and Modern Era
Media as a Mind Controlling Strategy In Old and Modern Era
faizulhassanfaiz1670
 
Eureka, I found it! - Special Libraries Association 2021 Presentation
Eureka, I found it! - Special Libraries Association 2021 PresentationEureka, I found it! - Special Libraries Association 2021 Presentation
Eureka, I found it! - Special Libraries Association 2021 Presentation
Access Innovations, Inc.
 
Acorn Recovery: Restore IT infra within minutes
Acorn Recovery: Restore IT infra within minutesAcorn Recovery: Restore IT infra within minutes
Acorn Recovery: Restore IT infra within minutes
IP ServerOne
 
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdfBonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
khadija278284
 
International Workshop on Artificial Intelligence in Software Testing
International Workshop on Artificial Intelligence in Software TestingInternational Workshop on Artificial Intelligence in Software Testing
International Workshop on Artificial Intelligence in Software Testing
Sebastiano Panichella
 
Bitcoin Lightning wallet and tic-tac-toe game XOXO
Bitcoin Lightning wallet and tic-tac-toe game XOXOBitcoin Lightning wallet and tic-tac-toe game XOXO
Bitcoin Lightning wallet and tic-tac-toe game XOXO
Matjaž Lipuš
 
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
OECD Directorate for Financial and Enterprise Affairs
 
Getting started with Amazon Bedrock Studio and Control Tower
Getting started with Amazon Bedrock Studio and Control TowerGetting started with Amazon Bedrock Studio and Control Tower
Getting started with Amazon Bedrock Studio and Control Tower
Vladimir Samoylov
 

Recently uploaded (16)

María Carolina Martínez - eCommerce Day Colombia 2024
María Carolina Martínez - eCommerce Day Colombia 2024María Carolina Martínez - eCommerce Day Colombia 2024
María Carolina Martínez - eCommerce Day Colombia 2024
 
0x01 - Newton's Third Law: Static vs. Dynamic Abusers
0x01 - Newton's Third Law:  Static vs. Dynamic Abusers0x01 - Newton's Third Law:  Static vs. Dynamic Abusers
0x01 - Newton's Third Law: Static vs. Dynamic Abusers
 
Announcement of 18th IEEE International Conference on Software Testing, Verif...
Announcement of 18th IEEE International Conference on Software Testing, Verif...Announcement of 18th IEEE International Conference on Software Testing, Verif...
Announcement of 18th IEEE International Conference on Software Testing, Verif...
 
somanykidsbutsofewfathers-140705000023-phpapp02.pptx
somanykidsbutsofewfathers-140705000023-phpapp02.pptxsomanykidsbutsofewfathers-140705000023-phpapp02.pptx
somanykidsbutsofewfathers-140705000023-phpapp02.pptx
 
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdfSupercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
Supercharge your AI - SSP Industry Breakout Session 2024-v2_1.pdf
 
Obesity causes and management and associated medical conditions
Obesity causes and management and associated medical conditionsObesity causes and management and associated medical conditions
Obesity causes and management and associated medical conditions
 
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
Sharpen existing tools or get a new toolbox? Contemporary cluster initiatives...
 
Doctoral Symposium at the 17th IEEE International Conference on Software Test...
Doctoral Symposium at the 17th IEEE International Conference on Software Test...Doctoral Symposium at the 17th IEEE International Conference on Software Test...
Doctoral Symposium at the 17th IEEE International Conference on Software Test...
 
Media as a Mind Controlling Strategy In Old and Modern Era
Media as a Mind Controlling Strategy In Old and Modern EraMedia as a Mind Controlling Strategy In Old and Modern Era
Media as a Mind Controlling Strategy In Old and Modern Era
 
Eureka, I found it! - Special Libraries Association 2021 Presentation
Eureka, I found it! - Special Libraries Association 2021 PresentationEureka, I found it! - Special Libraries Association 2021 Presentation
Eureka, I found it! - Special Libraries Association 2021 Presentation
 
Acorn Recovery: Restore IT infra within minutes
Acorn Recovery: Restore IT infra within minutesAcorn Recovery: Restore IT infra within minutes
Acorn Recovery: Restore IT infra within minutes
 
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdfBonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
Bonzo subscription_hjjjjjjjj5hhhhhhh_2024.pdf
 
International Workshop on Artificial Intelligence in Software Testing
International Workshop on Artificial Intelligence in Software TestingInternational Workshop on Artificial Intelligence in Software Testing
International Workshop on Artificial Intelligence in Software Testing
 
Bitcoin Lightning wallet and tic-tac-toe game XOXO
Bitcoin Lightning wallet and tic-tac-toe game XOXOBitcoin Lightning wallet and tic-tac-toe game XOXO
Bitcoin Lightning wallet and tic-tac-toe game XOXO
 
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
Competition and Regulation in Professional Services – KLEINER – June 2024 OEC...
 
Getting started with Amazon Bedrock Studio and Control Tower
Getting started with Amazon Bedrock Studio and Control TowerGetting started with Amazon Bedrock Studio and Control Tower
Getting started with Amazon Bedrock Studio and Control Tower
 

EXTRA: Integrating External Knowledge into Multimodal Hashtag Recommendation System

  • 1. 1 EXTRA : Integrating External Knowledge into Multimodal Hashtag Recommendation System 1 Department of Artificial Intelligence, The Catholic University of Korea, Bucheon, South Korea 2 Department of Data Science, The Catholic University of Korea, Bucheon, South Korea 3 Department of Computer Science and Engineering, Korea University, Seoul, South Korea Hyun-Sik Won1∗, Su-Min Roh1∗, Dohyun Kim3, Min-Ji Kim1, Huiju Kim3, and Kang-Min Kim12 * Equal contribution
  • 2. Background 2 • Microblogging Platforms Hashtags allow people to promote their own posts or categorize posts into specific topics.
  • 3. Introduction • Hashtag Recommendation Task : Recommend appropriate hashtags related to the post. • Previous Works : Recommend hashtags by utilizing explicit information in the post (i.e., text, image, etc.). : Recently, each user’s hashtag habits1 from the microblog are also utilized to recommend personalized hashtags. 3 1. Chen, Yu-Chi, et al. "Tagnet: triplet-attention graph networks for hashtag recommendation." IEEE Transactions on Circuits and Systems for Video Technology 32.3 (2021): 1148-1159.
  • 4. Introduction • Hashtag Recommendation Task : Recommend appropriate hashtags related to the post. • Previous Works : Recommend hashtags by utilizing explicit information in the post (i.e., text, image, etc.). : Recently, each user’s hashtag habits1 from the microblog are also utilized to recommend personalized hashtags. However, these methods only utilize the information found in the microblog.  This method may have difficulty understanding the contextual information of the post. 4 1. Chen, Yu-Chi, et al. "Tagnet: triplet-attention graph networks for hashtag recommendation." IEEE Transactions on Circuits and Systems for Video Technology 32.3 (2021): 1148-1159.
  • 5. Introduction • External Knowledge in Hashtag Recommendation 5
  • 6. Introduction 6 • External Knowledge in Hashtag Recommendation Does not appear directly in the text or image.
  • 7. Introduction 7 However, #dogtraining can be predicted based on external knowledge that dogs are typically trained. • External Knowledge in Hashtag Recommendation Does not appear directly in the text or image.
  • 8. Introduction 8 However, #dogtraining can be predicted based on external knowledge that dogs are typically trained. • External Knowledge in Hashtag Recommendation Does not appear directly in the text or image. How can we integrate the external knowledge to recommend more relevant hashtags for the post?
  • 9. Related Work • Open Directory Project (ODP) : This is a comprehensive directory of the World Wide Web, constructed and maintained by a community of volunteer editors. : It organizes web pages into the most related categories and subcategories. (e.g., Health/Animal/Mammals/Dogs) • There have been previous studies utilizing ODP categories as an external knowledge. : The study1 trained a classifier to predict the category based on the description of the webpage. 9 1. Kim et al., “meChat: In-device Personal Assistant for Conversational Photo Sharing” IEEE Internet Computing 2019
  • 10. Methodology 10 • Overview of the proposed model
  • 11. Methodology 11 • Overview of the proposed model ① External Knowledge Feature Extraction Model
  • 12. Methodology 12 ① External Knowledge Feature Extraction Model ② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA) • Overview of the proposed model
  • 13. Methodology ① External Knowledge Feature Extraction Model 13
  • 14. Methodology ① External Knowledge Feature Extraction Model 14 Caption Generator A woman is holding a certificate with a dog
  • 15. Methodology ① External Knowledge Feature Extraction Model 15 ODP-based Classifier (ODP-I) Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training Recreation/Pets/Dogs/Breeds/Working_Group Caption Generator A woman is holding a certificate with a dog
  • 16. Methodology ① External Knowledge Feature Extraction Model 16 ODP-based Classifier (ODP-I) Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training Recreation/Pets/Dogs/Breeds/Working_Group Caption Generator A woman is holding a certificate with a dog
  • 17. Methodology ② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA) 17 Image of the post Text of the post Relevant Category of the Image Algy is our other graduate. Today, Algy aced the see-saw and weave poles so impressively with this obstacle awareness.. Recreation/Pets/Dogs/Breeds/Working_Group Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training
  • 18. Methodology ② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA) 18 Image of the post Text of the post Relevant Category of the Image … Image Encoder Algy is our other graduate. Today, Algy aced the see-saw and weave poles so impressively with this obstacle awareness.. Recreation/Pets/Dogs/Breeds/Working_Group Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training
  • 19. Methodology ② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA) 19 Image of the post Text of the post Relevant Category of the Image Dogs Training Working Group awareness algy other … is our obstacle … Image Encoder Text Encoder Algy is our other graduate. Today, Algy aced the see-saw and weave poles so impressively with this obstacle awareness.. Recreation/Pets/Dogs/Breeds/Working_Group Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training
  • 20. Methodology ② Integrating External Knowledge into Multimodal Hashtag Recommendation System (EXTRA) 20 Image of the post Text of the post Relevant Category of the Image Dogs Training Working Group awareness algy other … is our obstacle … Image Encoder Text Encoder Multimodal Encoder Algy is our other graduate. Today, Algy aced the see-saw and weave poles so impressively with this obstacle awareness.. Recreation/Pets/Dogs/Breeds/Working_Group Health/Animal/Mammals/Dogs Recreation/Pets/Dogs/Training
  • 21. Dataset 21 • Open Directory Project Dataset (for extracting external knowledge) - # Web pages : 60,710 - # Categories : 2,531 • MaCon Dataset (for hashtag recommendation) - # Posts : 624,520 - # Hashtags : 3,896 - # Average of hashtags per post : 9.3
  • 22. Baselines 22 • Co-Attention (Co-AT) : the model that sequentially generates text and image feature vectors from a post using a co-attention mechanism. • MaCon : the model that applies a parallel co-attention mechanism to combine the text and image features with users’ tagging habits. • TAGNet : the model that integrates text, visual, and user habit features through a triplet attention module by constructing visual similarity graphs. SOTA
  • 23. Settings 23 • Implementation Details - Model : Pre-trained FLAVA1 - Optimizer : AdamW - Learning Rate : 1e-5 (10% warmup, cosine decay) - Batch Size : 64 • Evaluation Metric : Top-k Precision, Recall, F1-score 1. Singh, Amanpreet, et al. "Flava: A foundational language and vision alignment model." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.
  • 24. Experiments 24 • Comparison with Baselines 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 Top-1 Top-5 Top-9 F1-score Co-AT MaCon TAGNet EXTRA (Ours)  EXTRA achieved a 39.78% improvement over the state-of-the-art model TAGNet in terms of the Top-5 F1-Score.
  • 25. Experiments 25 • Comparison Between External Knowledge From Each Modality  The models that have integrated external knowledge perform better than those that have not.
  • 26. Conclusions 26 • We propose EXTRA, a novel multimodal hashtag recommendation system that combines text and image with external knowledge in a Transformer-based architecture. • We employ the most relevant categories extracted from the image of the post as external knowledge using the ODP-based classifier. • Our approach outperformed the existing state-of-the-art methods on the MaCon dataset, demonstrating its effectiveness in handling multimodal information for the hashtag recommendation task.
  • 27. 27
  • 28. Experiments 28 • Illustrations of Recommended Hashtags Predicted Hashtag (w/o ODP-I) #lax #makeportraits #mountainlife #plasticpollution #pursuitofportraits #spa #greenery #travel #detail #blackandwhite Predicted Hashtag (w/ ODP-I) #makeportraits #plasticpollution #mountainlife #adventure #lax #pursuitofportraits #travel #greatnorthcollective #nature
  • 30. Experiments 30 • Comparison between The Number of Subcategories