SlideShare a Scribd company logo
EXPLAIN-IT: Towards Explainable
AI for Unsupervised Network
Traffic Analysisº
Andrea Morichetta★, Pedro Casas*, Marco Mellia★
Politecnico di Torino★, Austrian Institute of Technology*
3rd ACM CoNEXT Workshop on Big DAta, Machine Learning and Artificial
Intelligence for Data Communication Networks
The Gap
• Scenario: Rising popularity of ML applications for solving
specific problems in network traffic analysis.
• Ground truth is systematically missing – difficult to obtain
(structural complexity and big data volumes)
• Labeled datasets are frequently simplistic representation of
real-world phenomena, often also outdated.
2
Unsupervised learning to fill the gap
• Unsupervised techniques allow to have a better understanding of the
data, exploring its shape and patterns.
• However, it is difficult to analyze their results
• Typical solutions:
• manual inspection  problem when there are too many or too complex data
• unsupervised quality metrics  the why is missing
• supervised quality metrics  not good if ground truth inherently wrong or
biased
3
Knowledge extraction from the clusters
Goal: have an interpretable representation of the features relevance in
the clusters
• For understanding the clusters content
• For better explanation of the data aggregation
4
Knowledge extraction – a supervised
approach
A possible solution: White box classifiers (white box techniques: e.g.,
linear regression and decision trees)
+Gives us also the opportunity to evaluate the cluster
attribution/assignment (via classification)
+Clear and algorithmically grounded
+Gives an “interpretation” available for the analysis
- It limits the set of applicable techniques
5
How to make this approach more general and extend the
set of algorithms?
Explainable AI - extend the supervised
approach
• EXPLAINABLE AI makes it easier to understand why certain decisions
or predictions have been made.
• Achieved by:
• Restricting the complexity of the machine learning model (intrinsic)
• Or by applying methods that analyze the model after training (post
hoc),
• e.g., LIME (Local Interpretable Model-agnostic Explanations)1 can
explain the predictions of any classifier or regressor, by
approximating it locally with an interpretable model.
6
1Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "" Why should i trust you?" Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD
international conference on knowledge discovery and data mining. 2016.
Approach
7
Data
Exploration Space
Knowledge
Summary Space
Unsupervised techniques
e.g., Clustering
Step 2
Advanced Knowledge Extraction
Splitting model
SVM
Identification of
XAI features with
LIME
Explainable AI
Use case
• 10654 YouTube video sessions, coming from different sources, smartphone
(HTML player and YouTube app) and desktop (HTML player)
• Set of ~500 features:
• at the full video session level (e.g., session downlink throughout)
• as well as at different time resolutions with time slots of ∆t = [1, 5, 10] seconds.
• We focus on the average video quality (AVGQ) metric. We consider video
resolution as follows:
• 0: Low Definition (LD), with AVGQ < 480
• 1: Standard Definition (SD), with 480 ≤ AVGQ < 720
• 2: High Definition (HD), with AVGQ ≥ 720
8
Clustering phase
• Goal: We want to obtain 3 clusters in output:
a. Low Definition, LD
b. Standard definition, SD
c. High Definition, HD
• Algorithms used:
• Agglomerative (1) clustering with Ward Links (Ward minimizes the variance
of the clusters being merged)
• Agglomerative (2) clustering with Single Links (Single single uses the
minimum of the distances between all observations of the two sets)
• K-Means
• BIRCH - Balanced Iterative Reducing and Clustering using Hierarchies
9
Clustering Results – quality metrics
10
Adjusted M
utual Info Score
Adjusted Rand Score
CompletenessScore
FowlkesM
allowsScore
Homogeneity Score
SilhouetteScore
V
M
easureScore
0 0
0 1
0 2
0 3
0 4
0 5
0 6
Algorithm
Agglomerative(1)
Agglomerative(2)
K-Means
Birch
Clustering results – label distribution
11
Label distribution after agglomerative Ward clustering
Clustering results – feature Inspection
12
Example of feature inspection inthe results of agglomerative Ward clustering
Cluster 0 Cluster 1 Cluster 2
Interpret with model – using Support Vector
Machines
• Hyperplane-based classifiers
• The SVM selects the maximum margin separating hyperplane
• Use of kernel function to map points on a high-dimensional space
• However, it is a black-box classifier
• Thus, Explainable AI can aid us
13
Interpret with model – using SVM
14
Agglomerative (1)
Results of SVM applied to Agglomerative with Ward
Results with LIME – an example
Feature Feature Importance
uplink_bytes_second_slot_1s > 10468.5 0.10
dist_packet_length_downlink_p25 > 1379 0.09
dist_slotted_uplink_bytes_p97_1s > 18445.9 0.08
uplink_packets_first_slot_5s > 861.3 0.07
420628.7 < dist_slotted_bytes_p97_1s <= 902383.7 0.07
dist_slotted_downlink_bytes_p97_5s > 2711876.9 0.06
dist_slotted_downlink_bytes_h_1s > 0.7 0.05
335.4 < dist_slotted_uplink_packets_p99_1s <= 502.0 0.04
dist_slotted_uplink_bytes_p90_1s > 7627.6 0.04
dist_slotted_bytes_mean_5s > 845017.6 0.04
15
Instance classified as belonging to cluster 2
Conclusion and future work
• Interesting approach for improving the interpretation of clustering
results by relying on XAI principles
• Is explainable AI an advantage in the YouTube case, where features
are complex?
• Is LIME always good? Look at alternatives, e.g., SHAP
• Is it possible to avoid the classification step?
• Extend it to other scenarios
• Expand the research on different clustering algorithms
• Use different classification techniques
16

More Related Content

Similar to ExplainableAI.pptx

Distilling dark knowledge from neural networks
Distilling dark knowledge from neural networksDistilling dark knowledge from neural networks
Distilling dark knowledge from neural networks
Alexander Korbonits
 
Introduction to image processing and pattern recognition
Introduction to image processing and pattern recognitionIntroduction to image processing and pattern recognition
Introduction to image processing and pattern recognition
Saibee Alam
 
The Data Science Process - Do we need it and how to apply?
The Data Science Process - Do we need it and how to apply?The Data Science Process - Do we need it and how to apply?
The Data Science Process - Do we need it and how to apply?
Ivo Andreev
 
Activity Monitoring Using Wearable Sensors and Smart Phone
Activity Monitoring Using Wearable Sensors and Smart PhoneActivity Monitoring Using Wearable Sensors and Smart Phone
Activity Monitoring Using Wearable Sensors and Smart Phone
DrAhmedZoha
 
Machine learning for IoT - unpacking the blackbox
Machine learning for IoT - unpacking the blackboxMachine learning for IoT - unpacking the blackbox
Machine learning for IoT - unpacking the blackbox
Ivo Andreev
 
Deep Learning Made Easy with Deep Features
Deep Learning Made Easy with Deep FeaturesDeep Learning Made Easy with Deep Features
Deep Learning Made Easy with Deep Features
Turi, Inc.
 
AI & ML in Cyber Security - Why Algorithms Are Dangerous
AI & ML in Cyber Security - Why Algorithms Are DangerousAI & ML in Cyber Security - Why Algorithms Are Dangerous
AI & ML in Cyber Security - Why Algorithms Are Dangerous
Raffael Marty
 
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
ssuser9357dd
 
Deep learning with keras
Deep learning with kerasDeep learning with keras
Deep learning with keras
MOHITKUMAR1379
 
Mastering AIOps with Deep Learning
Mastering AIOps with Deep LearningMastering AIOps with Deep Learning
Mastering AIOps with Deep Learning
Jorge Cardoso
 
230208 MLOps Getting from Good to Great.pptx
230208 MLOps Getting from Good to Great.pptx230208 MLOps Getting from Good to Great.pptx
230208 MLOps Getting from Good to Great.pptx
Arthur240715
 
Computer vision-nit-silchar-hackathon
Computer vision-nit-silchar-hackathonComputer vision-nit-silchar-hackathon
Computer vision-nit-silchar-hackathon
Aditya Bhattacharya
 
Unsupervised Learning: Clustering
Unsupervised Learning: Clustering Unsupervised Learning: Clustering
Unsupervised Learning: Clustering
Experfy
 
background.pptx
background.pptxbackground.pptx
background.pptx
KabileshCm
 
Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]
SubhradeepMaji
 
Object detection with deep learning
Object detection with deep learningObject detection with deep learning
Object detection with deep learning
Sushant Shrivastava
 
Bridging Concepts and Practice in eScience via Simulation-driven Engineering
Bridging Concepts and Practice in eScience via Simulation-driven EngineeringBridging Concepts and Practice in eScience via Simulation-driven Engineering
Bridging Concepts and Practice in eScience via Simulation-driven Engineering
Rafael Ferreira da Silva
 
Image Segmentation Using Deep Learning : A survey
Image Segmentation Using Deep Learning : A surveyImage Segmentation Using Deep Learning : A survey
Image Segmentation Using Deep Learning : A survey
NUPUR YADAV
 
Towards a Comprehensive Machine Learning Benchmark
Towards a Comprehensive Machine Learning BenchmarkTowards a Comprehensive Machine Learning Benchmark
Towards a Comprehensive Machine Learning Benchmark
Turi, Inc.
 
Two strategies for large-scale multi-label classification on the YouTube-8M d...
Two strategies for large-scale multi-label classification on the YouTube-8M d...Two strategies for large-scale multi-label classification on the YouTube-8M d...
Two strategies for large-scale multi-label classification on the YouTube-8M d...
Dalei Li
 

Similar to ExplainableAI.pptx (20)

Distilling dark knowledge from neural networks
Distilling dark knowledge from neural networksDistilling dark knowledge from neural networks
Distilling dark knowledge from neural networks
 
Introduction to image processing and pattern recognition
Introduction to image processing and pattern recognitionIntroduction to image processing and pattern recognition
Introduction to image processing and pattern recognition
 
The Data Science Process - Do we need it and how to apply?
The Data Science Process - Do we need it and how to apply?The Data Science Process - Do we need it and how to apply?
The Data Science Process - Do we need it and how to apply?
 
Activity Monitoring Using Wearable Sensors and Smart Phone
Activity Monitoring Using Wearable Sensors and Smart PhoneActivity Monitoring Using Wearable Sensors and Smart Phone
Activity Monitoring Using Wearable Sensors and Smart Phone
 
Machine learning for IoT - unpacking the blackbox
Machine learning for IoT - unpacking the blackboxMachine learning for IoT - unpacking the blackbox
Machine learning for IoT - unpacking the blackbox
 
Deep Learning Made Easy with Deep Features
Deep Learning Made Easy with Deep FeaturesDeep Learning Made Easy with Deep Features
Deep Learning Made Easy with Deep Features
 
AI & ML in Cyber Security - Why Algorithms Are Dangerous
AI & ML in Cyber Security - Why Algorithms Are DangerousAI & ML in Cyber Security - Why Algorithms Are Dangerous
AI & ML in Cyber Security - Why Algorithms Are Dangerous
 
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
 
Deep learning with keras
Deep learning with kerasDeep learning with keras
Deep learning with keras
 
Mastering AIOps with Deep Learning
Mastering AIOps with Deep LearningMastering AIOps with Deep Learning
Mastering AIOps with Deep Learning
 
230208 MLOps Getting from Good to Great.pptx
230208 MLOps Getting from Good to Great.pptx230208 MLOps Getting from Good to Great.pptx
230208 MLOps Getting from Good to Great.pptx
 
Computer vision-nit-silchar-hackathon
Computer vision-nit-silchar-hackathonComputer vision-nit-silchar-hackathon
Computer vision-nit-silchar-hackathon
 
Unsupervised Learning: Clustering
Unsupervised Learning: Clustering Unsupervised Learning: Clustering
Unsupervised Learning: Clustering
 
background.pptx
background.pptxbackground.pptx
background.pptx
 
Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]
 
Object detection with deep learning
Object detection with deep learningObject detection with deep learning
Object detection with deep learning
 
Bridging Concepts and Practice in eScience via Simulation-driven Engineering
Bridging Concepts and Practice in eScience via Simulation-driven EngineeringBridging Concepts and Practice in eScience via Simulation-driven Engineering
Bridging Concepts and Practice in eScience via Simulation-driven Engineering
 
Image Segmentation Using Deep Learning : A survey
Image Segmentation Using Deep Learning : A surveyImage Segmentation Using Deep Learning : A survey
Image Segmentation Using Deep Learning : A survey
 
Towards a Comprehensive Machine Learning Benchmark
Towards a Comprehensive Machine Learning BenchmarkTowards a Comprehensive Machine Learning Benchmark
Towards a Comprehensive Machine Learning Benchmark
 
Two strategies for large-scale multi-label classification on the YouTube-8M d...
Two strategies for large-scale multi-label classification on the YouTube-8M d...Two strategies for large-scale multi-label classification on the YouTube-8M d...
Two strategies for large-scale multi-label classification on the YouTube-8M d...
 

Recently uploaded

Tutorial for 16S rRNA Gene Analysis with QIIME2.pdf
Tutorial for 16S rRNA Gene Analysis with QIIME2.pdfTutorial for 16S rRNA Gene Analysis with QIIME2.pdf
Tutorial for 16S rRNA Gene Analysis with QIIME2.pdf
aqil azizi
 
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
Mukeshwaran Balu
 
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressionsKuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
Victor Morales
 
Understanding Inductive Bias in Machine Learning
Understanding Inductive Bias in Machine LearningUnderstanding Inductive Bias in Machine Learning
Understanding Inductive Bias in Machine Learning
SUTEJAS
 
Unbalanced Three Phase Systems and circuits.pptx
Unbalanced Three Phase Systems and circuits.pptxUnbalanced Three Phase Systems and circuits.pptx
Unbalanced Three Phase Systems and circuits.pptx
ChristineTorrepenida1
 
ACEP Magazine edition 4th launched on 05.06.2024
ACEP Magazine edition 4th launched on 05.06.2024ACEP Magazine edition 4th launched on 05.06.2024
ACEP Magazine edition 4th launched on 05.06.2024
Rahul
 
14 Template Contractual Notice - EOT Application
14 Template Contractual Notice - EOT Application14 Template Contractual Notice - EOT Application
14 Template Contractual Notice - EOT Application
SyedAbiiAzazi1
 
bank management system in java and mysql report1.pdf
bank management system in java and mysql report1.pdfbank management system in java and mysql report1.pdf
bank management system in java and mysql report1.pdf
Divyam548318
 
Generative AI leverages algorithms to create various forms of content
Generative AI leverages algorithms to create various forms of contentGenerative AI leverages algorithms to create various forms of content
Generative AI leverages algorithms to create various forms of content
Hitesh Mohapatra
 
Hierarchical Digital Twin of a Naval Power System
Hierarchical Digital Twin of a Naval Power SystemHierarchical Digital Twin of a Naval Power System
Hierarchical Digital Twin of a Naval Power System
Kerry Sado
 
digital fundamental by Thomas L.floydl.pdf
digital fundamental by Thomas L.floydl.pdfdigital fundamental by Thomas L.floydl.pdf
digital fundamental by Thomas L.floydl.pdf
drwaing
 
Harnessing WebAssembly for Real-time Stateless Streaming Pipelines
Harnessing WebAssembly for Real-time Stateless Streaming PipelinesHarnessing WebAssembly for Real-time Stateless Streaming Pipelines
Harnessing WebAssembly for Real-time Stateless Streaming Pipelines
Christina Lin
 
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
thanhdowork
 
Technical Drawings introduction to drawing of prisms
Technical Drawings introduction to drawing of prismsTechnical Drawings introduction to drawing of prisms
Technical Drawings introduction to drawing of prisms
heavyhaig
 
Modelagem de um CSTR com reação endotermica.pdf
Modelagem de um CSTR com reação endotermica.pdfModelagem de um CSTR com reação endotermica.pdf
Modelagem de um CSTR com reação endotermica.pdf
camseq
 
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
zwunae
 
Low power architecture of logic gates using adiabatic techniques
Low power architecture of logic gates using adiabatic techniquesLow power architecture of logic gates using adiabatic techniques
Low power architecture of logic gates using adiabatic techniques
nooriasukmaningtyas
 
introduction to solar energy for engineering.pdf
introduction to solar energy for engineering.pdfintroduction to solar energy for engineering.pdf
introduction to solar energy for engineering.pdf
ravindarpurohit26
 
PPT on GRP pipes manufacturing and testing
PPT on GRP pipes manufacturing and testingPPT on GRP pipes manufacturing and testing
PPT on GRP pipes manufacturing and testing
anoopmanoharan2
 
basic-wireline-operations-course-mahmoud-f-radwan.pdf
basic-wireline-operations-course-mahmoud-f-radwan.pdfbasic-wireline-operations-course-mahmoud-f-radwan.pdf
basic-wireline-operations-course-mahmoud-f-radwan.pdf
NidhalKahouli2
 

Recently uploaded (20)

Tutorial for 16S rRNA Gene Analysis with QIIME2.pdf
Tutorial for 16S rRNA Gene Analysis with QIIME2.pdfTutorial for 16S rRNA Gene Analysis with QIIME2.pdf
Tutorial for 16S rRNA Gene Analysis with QIIME2.pdf
 
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
ACRP 4-09 Risk Assessment Method to Support Modification of Airfield Separat...
 
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressionsKuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
KuberTENes Birthday Bash Guadalajara - K8sGPT first impressions
 
Understanding Inductive Bias in Machine Learning
Understanding Inductive Bias in Machine LearningUnderstanding Inductive Bias in Machine Learning
Understanding Inductive Bias in Machine Learning
 
Unbalanced Three Phase Systems and circuits.pptx
Unbalanced Three Phase Systems and circuits.pptxUnbalanced Three Phase Systems and circuits.pptx
Unbalanced Three Phase Systems and circuits.pptx
 
ACEP Magazine edition 4th launched on 05.06.2024
ACEP Magazine edition 4th launched on 05.06.2024ACEP Magazine edition 4th launched on 05.06.2024
ACEP Magazine edition 4th launched on 05.06.2024
 
14 Template Contractual Notice - EOT Application
14 Template Contractual Notice - EOT Application14 Template Contractual Notice - EOT Application
14 Template Contractual Notice - EOT Application
 
bank management system in java and mysql report1.pdf
bank management system in java and mysql report1.pdfbank management system in java and mysql report1.pdf
bank management system in java and mysql report1.pdf
 
Generative AI leverages algorithms to create various forms of content
Generative AI leverages algorithms to create various forms of contentGenerative AI leverages algorithms to create various forms of content
Generative AI leverages algorithms to create various forms of content
 
Hierarchical Digital Twin of a Naval Power System
Hierarchical Digital Twin of a Naval Power SystemHierarchical Digital Twin of a Naval Power System
Hierarchical Digital Twin of a Naval Power System
 
digital fundamental by Thomas L.floydl.pdf
digital fundamental by Thomas L.floydl.pdfdigital fundamental by Thomas L.floydl.pdf
digital fundamental by Thomas L.floydl.pdf
 
Harnessing WebAssembly for Real-time Stateless Streaming Pipelines
Harnessing WebAssembly for Real-time Stateless Streaming PipelinesHarnessing WebAssembly for Real-time Stateless Streaming Pipelines
Harnessing WebAssembly for Real-time Stateless Streaming Pipelines
 
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Hori...
 
Technical Drawings introduction to drawing of prisms
Technical Drawings introduction to drawing of prismsTechnical Drawings introduction to drawing of prisms
Technical Drawings introduction to drawing of prisms
 
Modelagem de um CSTR com reação endotermica.pdf
Modelagem de um CSTR com reação endotermica.pdfModelagem de um CSTR com reação endotermica.pdf
Modelagem de um CSTR com reação endotermica.pdf
 
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单专业办理
 
Low power architecture of logic gates using adiabatic techniques
Low power architecture of logic gates using adiabatic techniquesLow power architecture of logic gates using adiabatic techniques
Low power architecture of logic gates using adiabatic techniques
 
introduction to solar energy for engineering.pdf
introduction to solar energy for engineering.pdfintroduction to solar energy for engineering.pdf
introduction to solar energy for engineering.pdf
 
PPT on GRP pipes manufacturing and testing
PPT on GRP pipes manufacturing and testingPPT on GRP pipes manufacturing and testing
PPT on GRP pipes manufacturing and testing
 
basic-wireline-operations-course-mahmoud-f-radwan.pdf
basic-wireline-operations-course-mahmoud-f-radwan.pdfbasic-wireline-operations-course-mahmoud-f-radwan.pdf
basic-wireline-operations-course-mahmoud-f-radwan.pdf
 

ExplainableAI.pptx

  • 1. EXPLAIN-IT: Towards Explainable AI for Unsupervised Network Traffic Analysisº Andrea Morichetta★, Pedro Casas*, Marco Mellia★ Politecnico di Torino★, Austrian Institute of Technology* 3rd ACM CoNEXT Workshop on Big DAta, Machine Learning and Artificial Intelligence for Data Communication Networks
  • 2. The Gap • Scenario: Rising popularity of ML applications for solving specific problems in network traffic analysis. • Ground truth is systematically missing – difficult to obtain (structural complexity and big data volumes) • Labeled datasets are frequently simplistic representation of real-world phenomena, often also outdated. 2
  • 3. Unsupervised learning to fill the gap • Unsupervised techniques allow to have a better understanding of the data, exploring its shape and patterns. • However, it is difficult to analyze their results • Typical solutions: • manual inspection  problem when there are too many or too complex data • unsupervised quality metrics  the why is missing • supervised quality metrics  not good if ground truth inherently wrong or biased 3
  • 4. Knowledge extraction from the clusters Goal: have an interpretable representation of the features relevance in the clusters • For understanding the clusters content • For better explanation of the data aggregation 4
  • 5. Knowledge extraction – a supervised approach A possible solution: White box classifiers (white box techniques: e.g., linear regression and decision trees) +Gives us also the opportunity to evaluate the cluster attribution/assignment (via classification) +Clear and algorithmically grounded +Gives an “interpretation” available for the analysis - It limits the set of applicable techniques 5 How to make this approach more general and extend the set of algorithms?
  • 6. Explainable AI - extend the supervised approach • EXPLAINABLE AI makes it easier to understand why certain decisions or predictions have been made. • Achieved by: • Restricting the complexity of the machine learning model (intrinsic) • Or by applying methods that analyze the model after training (post hoc), • e.g., LIME (Local Interpretable Model-agnostic Explanations)1 can explain the predictions of any classifier or regressor, by approximating it locally with an interpretable model. 6 1Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "" Why should i trust you?" Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016.
  • 7. Approach 7 Data Exploration Space Knowledge Summary Space Unsupervised techniques e.g., Clustering Step 2 Advanced Knowledge Extraction Splitting model SVM Identification of XAI features with LIME Explainable AI
  • 8. Use case • 10654 YouTube video sessions, coming from different sources, smartphone (HTML player and YouTube app) and desktop (HTML player) • Set of ~500 features: • at the full video session level (e.g., session downlink throughout) • as well as at different time resolutions with time slots of ∆t = [1, 5, 10] seconds. • We focus on the average video quality (AVGQ) metric. We consider video resolution as follows: • 0: Low Definition (LD), with AVGQ < 480 • 1: Standard Definition (SD), with 480 ≤ AVGQ < 720 • 2: High Definition (HD), with AVGQ ≥ 720 8
  • 9. Clustering phase • Goal: We want to obtain 3 clusters in output: a. Low Definition, LD b. Standard definition, SD c. High Definition, HD • Algorithms used: • Agglomerative (1) clustering with Ward Links (Ward minimizes the variance of the clusters being merged) • Agglomerative (2) clustering with Single Links (Single single uses the minimum of the distances between all observations of the two sets) • K-Means • BIRCH - Balanced Iterative Reducing and Clustering using Hierarchies 9
  • 10. Clustering Results – quality metrics 10 Adjusted M utual Info Score Adjusted Rand Score CompletenessScore FowlkesM allowsScore Homogeneity Score SilhouetteScore V M easureScore 0 0 0 1 0 2 0 3 0 4 0 5 0 6 Algorithm Agglomerative(1) Agglomerative(2) K-Means Birch
  • 11. Clustering results – label distribution 11 Label distribution after agglomerative Ward clustering
  • 12. Clustering results – feature Inspection 12 Example of feature inspection inthe results of agglomerative Ward clustering Cluster 0 Cluster 1 Cluster 2
  • 13. Interpret with model – using Support Vector Machines • Hyperplane-based classifiers • The SVM selects the maximum margin separating hyperplane • Use of kernel function to map points on a high-dimensional space • However, it is a black-box classifier • Thus, Explainable AI can aid us 13
  • 14. Interpret with model – using SVM 14 Agglomerative (1) Results of SVM applied to Agglomerative with Ward
  • 15. Results with LIME – an example Feature Feature Importance uplink_bytes_second_slot_1s > 10468.5 0.10 dist_packet_length_downlink_p25 > 1379 0.09 dist_slotted_uplink_bytes_p97_1s > 18445.9 0.08 uplink_packets_first_slot_5s > 861.3 0.07 420628.7 < dist_slotted_bytes_p97_1s <= 902383.7 0.07 dist_slotted_downlink_bytes_p97_5s > 2711876.9 0.06 dist_slotted_downlink_bytes_h_1s > 0.7 0.05 335.4 < dist_slotted_uplink_packets_p99_1s <= 502.0 0.04 dist_slotted_uplink_bytes_p90_1s > 7627.6 0.04 dist_slotted_bytes_mean_5s > 845017.6 0.04 15 Instance classified as belonging to cluster 2
  • 16. Conclusion and future work • Interesting approach for improving the interpretation of clustering results by relying on XAI principles • Is explainable AI an advantage in the YouTube case, where features are complex? • Is LIME always good? Look at alternatives, e.g., SHAP • Is it possible to avoid the classification step? • Extend it to other scenarios • Expand the research on different clustering algorithms • Use different classification techniques 16

Editor's Notes

  1. Why our model predicted a specific label? E.g., if traffic is malicious or not? LIME intuition is to look closer in the area of the predicted decision, and get easier boundaries LIME is only based on inputs and outputs of the model Random generating data points, by perturbation, in the neighborhood of our target data point What we get, is a new dataset in the neighborhood of our target, that we can interpret with a white box model Assign weights to the points closer to the target in order to get these rights when predicting with a local linear model
  2. packet-level video traffic measurements only information extracted from the network traffic for each of the captured packet are packet time and packet size. From these two values, we then derive a full set of 477 different features Overall/full session traffic, downlink traffic and uplink traffic Sampled empirical distributions of overall session traffic, downlink traffic and uplink traffic extracted from the analyzed network video traffic packets into relevant Video Quality Metrics. Six VQMs: initial delay, frequency of stallings, number of stalling events, number of quality switches, average video quality (video vertical resolution, e.g., 480p, 720p, 1080p, etc.) and average video bitrate.