SlideShare a Scribd company logo
1 of 51
Download to read offline
Deep learning in online
commerce
Privileged and Confidential
Stanislav
Stolpovskiy
GD Engineer
Privileged and Confidential
About me
this is my robot:● Grid Dynamics engineer 2012-present
● 10+ years of Java experience
● Transitioned from Search to ML 3 years
ago
● Graduated from St Petersburg IFMO
University
Privileged and Confidential
Agenda
● Case studies
● Foundational capabilities
● Deep dive: How to search by artistic style
Case studies
Privileged and Confidential
How it all started
● Search accuracy hurt from bad catalog data
● Is this a common problem ?
● Initial approach: build an algorithm to recognize
product color from the produce image
○ Works ok in most cases
○ But… not scalable
● Lack of scalability lead us to try a single ML
model that could recognize all product attributes
Privileged and Confidential
Automated attribution for leading department store chain
LENGTH floor
COLOR red
HEM high-low
SLEEVE sleeveless
STYLE ruffled
OCCASION evening
Image
embedding
Text
embedding
Featurefusion
Latent feature
space
AttributeClassifiers
Privileged and Confidential
More like this for leading department store chain
More Like
This
Privileged and Confidential
“Similar Art” feature for leading art retailer
Privileged and Confidential
Pin recommendations for leading home improvements chain
Privileged and Confidential
Part search for leading home improvements chain
Foundational
capabilities
Privileged and Confidential
Deep learning: convolutional neural networks
Privileged and Confidential
Visual recommendations
Privileged and Confidential
Visual recommendations
Image
embedding
Attribute
embedding
Featurefusion
Latent feature
space
More Like This
LENGTH floor
COLOR red
HEM high-low
SLEEVE sleeveless
STYLE ruffled
OCCASION evening
Nearest
neighbors
search
How to search by
artistic style
Privileged and Confidential
Visual search for art
Similar by image contents:
“peaches”
Similar by artistic style:
“old Dutch masters”
Looking for something
like this to decorate my
room...
✅🚫
Privileged and Confidential
“Similar Art” feature
Privileged and Confidential
Deep learning: convolutional neural networks
Privileged and Confidential
Feature extraction (vectoriser)
G G G
fusion
style vector content vector
fusion
image vector
Privileged and Confidential
General approach
KNN - K-Nearest neighbor search
Privileged and Confidential
High level architecture
Privileged and Confidential
Dimensionality reduction
● Why? Image vectors are very large (~60K dimensions), need to
reduce their size without losing much of features
● Randomized PCA is used (fbpca, sklearn)
○ fast, yet hogs memory
● Search space pruning allows to split into batches
Privileged and Confidential
Approximate nearest neighbor search
● Why? Exact algorithm is too slow
○ With 2M vectors in memory, response time ~ 500 msec.
● Approximate KNN
○ Training: ~ 1.5h
○ Accuracy: ~90%
○ Response time: 5-10 ms
Similar image search
on steroids
Privileged and Confidential
Wild image issue
Transfer learning from ImageNet is not working well for wild images
0.5 0.9
Privileged and Confidential
Siamese network
Privileged and Confidential
Triplet loss function
Privileged and Confidential
Image feature extraction
Pin recommendation
Privileged and Confidential
Find a chair in the room
Privileged and Confidential
Find a chair in the room
Privileged and Confidential
Find a objects in the room
Privileged and Confidential
Object detection network
Privileged and Confidential
High level architecture
Part search
Privileged and Confidential
Real-world size problem
● Part search requires high dimensional precision
● Image doesn’t provide any size information
● The same objects can be made in range of different sizes
● We have to ask to use reference object
Privileged and Confidential
The Object Detection approach to localization
Pros:
● Detects multiple objects in an image
● Detects different classes in an image
● There are a lot of pre-trained models for
different frameworks
Cons:
● The bounding box around an object still
contains background
● High recall must be kept. At the same time,
avoiding false positives sometimes failed for
real cases
● The Object Detection model is
computationally heavy.
Privileged and Confidential
UNet localization
U-Net models classify each pixel that belongs to the object of interest.
Privileged and Confidential
High level architecture
Appendix
Privileged and Confidential
AI catalog management for leading department store chain
Merchandiser AI assistant
product attribute
suggestion
product
clustering for
mass edit
category
suggestions
product name
generation
attribute
verification
3. Solution
Solution is based on state-of-the art natural language processing model based on transfer learning of the powerful general
purpose language model. Fine tuning is used to complete the training of the language model for specific domain, in this case -
shopping.
NLP model is able to accurately classify complex phrases which helps to correctly interpret customer intent in a complex dialog.
Same approach can be used for wide range of phrase and text classification applications, such as customer support.
Case study: Intent analysis for conversational platform
Privileged and Confidential 42
1. Situation
An online retailer is developing conversational
commerce solution enabling discovery, selection,
checkout and post-order functionality with smart
voice devices such as Google Home and Alexa.
Online retailer is looking for the solution which is
able to accurately recognize shopping intents of for
utterances such as “Can you suggest me an action
camera?”
2. Scope and Goals
Grid Dynamics is developing a conversational
commerce platform capable of accurately classify
customer intentions, understand the query and
support a seamless dialog with the customer.
can
you
...
...
...
action
camera?
word
embedding
phrase
matrix
LSTM
LSTM
...
LSTM
...
...
pooling
linear
softmax
backward
phrase
embedding
LSTM
...
...
LSTM
...
LSTM
forward
phrase
embedding
CLASS PROB
ACCEPT 0.01
DENY 0.001
NEXT 0.05
… …
DISCOVERY 0.81
SELECTION 0.1
ORDER 0.05
SMALLTALK 0.02
Privileged and Confidential
Deep learning: convolutional neural networks
Privileged and Confidential
UNet localization
Privileged and Confidential
250-dim latent space
Image
features
Feature fusion with CCA
1 x 2048
Text
features 1 x 300
Distances to
minimize
Find a pair of linear transformations Wimage
and Wtext
which minimizes distances in latent space of
a given size for all items in training set
Wimage
Wtext
Wimage
Wtext
Privileged and Confidential
First neural network
LENGTH floor
COLOR red
HEM high-low
SLEEE sleeveless
STYLE ruffled
OCCASION evening
Image
embedding
Featurefusion
Latent feature
space
AttributeClassifiers
We use CNN instead algorithm
Works for all product types
Privileged and Confidential
Results -- adding ML helped!
● Increased merchandiser productivity
● Improved relevancy of search
● Able to cover product attribute checking for an extremely large catalog
● Reduced catalog misattribution using this tool for the last 3 years
Privileged and Confidential
Item identification for online dinnerware retailer
Catalog image Wild image
Privileged and Confidential
Shoppable content
product_type:coat
gender: women
material: fur
prod214380280
women fur coat similar to this
Privileged and Confidential
The next way to improve search with ML
LENGTH floor
COLOR red
HEM high-low
SLEEE sleeveless
STYLE ruffled
OCCASION evening
Image
embedding
Featurefusion
Latent feature
space
AttributeClassifiers
Opportunity to use image
recognition as a search feature
not just a product attribute
checker
Thank you!
www.griddynamics.com

More Related Content

Similar to Deep learning applications in e-commerce search: Dynamic talks Chicago 3/14/2019

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.
 

Similar to Deep learning applications in e-commerce search: Dynamic talks Chicago 3/14/2019 (20)

How to use transfer learning to bootstrap image classification and question a...
How to use transfer learning to bootstrap image classification and question a...How to use transfer learning to bootstrap image classification and question a...
How to use transfer learning to bootstrap image classification and question a...
 
Staying Shallow & Lean in a Deep Learning World
Staying Shallow & Lean in a Deep Learning WorldStaying Shallow & Lean in a Deep Learning World
Staying Shallow & Lean in a Deep Learning World
 
AI hype or reality
AI  hype or realityAI  hype or reality
AI hype or reality
 
Yurii Pashchenko: Tips and tricks for building your own automated visual data...
Yurii Pashchenko: Tips and tricks for building your own automated visual data...Yurii Pashchenko: Tips and tricks for building your own automated visual data...
Yurii Pashchenko: Tips and tricks for building your own automated visual data...
 
[DSC Europe 22] Developing Visual AI Solutions for Online Marketplaces - Mlad...
[DSC Europe 22] Developing Visual AI Solutions for Online Marketplaces - Mlad...[DSC Europe 22] Developing Visual AI Solutions for Online Marketplaces - Mlad...
[DSC Europe 22] Developing Visual AI Solutions for Online Marketplaces - Mlad...
 
Deep Learning and the state of AI / 2016
Deep Learning and the state of AI / 2016Deep Learning and the state of AI / 2016
Deep Learning and the state of AI / 2016
 
Introduction to Machine Learning - WeCloudData
Introduction to Machine Learning - WeCloudDataIntroduction to Machine Learning - WeCloudData
Introduction to Machine Learning - WeCloudData
 
Introduction to Machine Learning - WeCloudData
Introduction to Machine Learning - WeCloudDataIntroduction to Machine Learning - WeCloudData
Introduction to Machine Learning - WeCloudData
 
Semi-Supervised Learning In An Adversarial Environment
Semi-Supervised Learning In An Adversarial EnvironmentSemi-Supervised Learning In An Adversarial Environment
Semi-Supervised Learning In An Adversarial Environment
 
tensorflow.pptx
tensorflow.pptxtensorflow.pptx
tensorflow.pptx
 
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
 
How I became ML Engineer
How I became ML Engineer How I became ML Engineer
How I became ML Engineer
 
Production ML Systems and Computer Vision with Google Cloud
Production ML Systems and Computer Vision with Google CloudProduction ML Systems and Computer Vision with Google Cloud
Production ML Systems and Computer Vision with Google Cloud
 
Pie for ai
Pie for aiPie for ai
Pie for ai
 
Certification Study Group - Professional ML Engineer Session 3 (Machine Learn...
Certification Study Group - Professional ML Engineer Session 3 (Machine Learn...Certification Study Group - Professional ML Engineer Session 3 (Machine Learn...
Certification Study Group - Professional ML Engineer Session 3 (Machine Learn...
 
Build, Train, and Deploy ML Models at Scale
Build, Train, and Deploy ML Models at ScaleBuild, Train, and Deploy ML Models at Scale
Build, Train, and Deploy ML Models at Scale
 
B4UConference_machine learning_deeplearning
B4UConference_machine learning_deeplearningB4UConference_machine learning_deeplearning
B4UConference_machine learning_deeplearning
 
BSSML16 L10. Summary Day 2 Sessions
BSSML16 L10. Summary Day 2 SessionsBSSML16 L10. Summary Day 2 Sessions
BSSML16 L10. Summary Day 2 Sessions
 
Conversational commerce: emerging architectures for smart & useful chatbots -...
Conversational commerce: emerging architectures for smart & useful chatbots -...Conversational commerce: emerging architectures for smart & useful chatbots -...
Conversational commerce: emerging architectures for smart & useful chatbots -...
 
Conversational commerce: emerging architectures for smart & useful chatbots -...
Conversational commerce: emerging architectures for smart & useful chatbots -...Conversational commerce: emerging architectures for smart & useful chatbots -...
Conversational commerce: emerging architectures for smart & useful chatbots -...
 

More from Grid Dynamics

Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
Grid Dynamics
 

More from Grid Dynamics (20)

Are you keeping up with your customer
Are you keeping up with your customer Are you keeping up with your customer
Are you keeping up with your customer
 
"Implementing data quality automation with open source stack" - Max Martynov,...
"Implementing data quality automation with open source stack" - Max Martynov,..."Implementing data quality automation with open source stack" - Max Martynov,...
"Implementing data quality automation with open source stack" - Max Martynov,...
 
"How to build cool & useful voice commerce applications (such as devices like...
"How to build cool & useful voice commerce applications (such as devices like..."How to build cool & useful voice commerce applications (such as devices like...
"How to build cool & useful voice commerce applications (such as devices like...
 
"Challenges for AI in Healthcare" - Peter Graven Ph.D
"Challenges for AI in Healthcare" - Peter Graven Ph.D"Challenges for AI in Healthcare" - Peter Graven Ph.D
"Challenges for AI in Healthcare" - Peter Graven Ph.D
 
Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
Dynamic Talks: "Applications of Big Data, Machine Learning and Artificial Int...
 
Dynamic Talks: "Digital Transformation in Banking & Financial Services… a per...
Dynamic Talks: "Digital Transformation in Banking & Financial Services… a per...Dynamic Talks: "Digital Transformation in Banking & Financial Services… a per...
Dynamic Talks: "Digital Transformation in Banking & Financial Services… a per...
 
Dynamic Talks: "Data Strategy as a Conduit for Data Maturity and Monetization...
Dynamic Talks: "Data Strategy as a Conduit for Data Maturity and Monetization...Dynamic Talks: "Data Strategy as a Conduit for Data Maturity and Monetization...
Dynamic Talks: "Data Strategy as a Conduit for Data Maturity and Monetization...
 
Dynamics Talks: "Writing Spark Pipelines with Less Boilerplate Code" - Egor P...
Dynamics Talks: "Writing Spark Pipelines with Less Boilerplate Code" - Egor P...Dynamics Talks: "Writing Spark Pipelines with Less Boilerplate Code" - Egor P...
Dynamics Talks: "Writing Spark Pipelines with Less Boilerplate Code" - Egor P...
 
"Trends in Building Advanced Analytics Platform for Large Enterprises" - Atul...
"Trends in Building Advanced Analytics Platform for Large Enterprises" - Atul..."Trends in Building Advanced Analytics Platform for Large Enterprises" - Atul...
"Trends in Building Advanced Analytics Platform for Large Enterprises" - Atul...
 
The New Era of Public Safety Records Management: Dynamic talks Chicago 9/24/2019
The New Era of Public Safety Records Management: Dynamic talks Chicago 9/24/2019The New Era of Public Safety Records Management: Dynamic talks Chicago 9/24/2019
The New Era of Public Safety Records Management: Dynamic talks Chicago 9/24/2019
 
Dynamic Talks: "Implementing data quality automation with open source stack" ...
Dynamic Talks: "Implementing data quality automation with open source stack" ...Dynamic Talks: "Implementing data quality automation with open source stack" ...
Dynamic Talks: "Implementing data quality automation with open source stack" ...
 
"Implementing AI for New Business Models and Efficiencies" - Parag Shrivastav...
"Implementing AI for New Business Models and Efficiencies" - Parag Shrivastav..."Implementing AI for New Business Models and Efficiencies" - Parag Shrivastav...
"Implementing AI for New Business Models and Efficiencies" - Parag Shrivastav...
 
Reducing No-shows and Late Cancelations in Healthcare Enterprise" - Shervin M...
Reducing No-shows and Late Cancelations in Healthcare Enterprise" - Shervin M...Reducing No-shows and Late Cancelations in Healthcare Enterprise" - Shervin M...
Reducing No-shows and Late Cancelations in Healthcare Enterprise" - Shervin M...
 
Customer intelligence: a Machine Learning Approach: Dynamic talks Atlanta 8/2...
Customer intelligence: a Machine Learning Approach: Dynamic talks Atlanta 8/2...Customer intelligence: a Machine Learning Approach: Dynamic talks Atlanta 8/2...
Customer intelligence: a Machine Learning Approach: Dynamic talks Atlanta 8/2...
 
"ML Services - How do you begin and when do you start scaling?" - Madhura Dud...
"ML Services - How do you begin and when do you start scaling?" - Madhura Dud..."ML Services - How do you begin and when do you start scaling?" - Madhura Dud...
"ML Services - How do you begin and when do you start scaling?" - Madhura Dud...
 
Realtime Contextual Product Recommendations…that scale and generate revenue -...
Realtime Contextual Product Recommendations…that scale and generate revenue -...Realtime Contextual Product Recommendations…that scale and generate revenue -...
Realtime Contextual Product Recommendations…that scale and generate revenue -...
 
Decision Automation in Marketing Systems using Reinforcement Learning: Dynami...
Decision Automation in Marketing Systems using Reinforcement Learning: Dynami...Decision Automation in Marketing Systems using Reinforcement Learning: Dynami...
Decision Automation in Marketing Systems using Reinforcement Learning: Dynami...
 
Best practices for enterprise-grade microservices implementations with Google...
Best practices for enterprise-grade microservices implementations with Google...Best practices for enterprise-grade microservices implementations with Google...
Best practices for enterprise-grade microservices implementations with Google...
 
Attribution Modelling 101: Credit Where Credit is Due!: Dynamic talks Seattle...
Attribution Modelling 101: Credit Where Credit is Due!: Dynamic talks Seattle...Attribution Modelling 101: Credit Where Credit is Due!: Dynamic talks Seattle...
Attribution Modelling 101: Credit Where Credit is Due!: Dynamic talks Seattle...
 
Building an algorithmic price management system using ML: Dynamic talks Seatt...
Building an algorithmic price management system using ML: Dynamic talks Seatt...Building an algorithmic price management system using ML: Dynamic talks Seatt...
Building an algorithmic price management system using ML: Dynamic talks Seatt...
 

Recently uploaded

Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native Applications
WSO2
 

Recently uploaded (20)

[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdf[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdf
 
Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native Applications
 
CNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In PakistanCNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In Pakistan
 
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
 
FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024
 
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with Milvus
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor Presentation
 
WSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering DevelopersWSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering Developers
 
Vector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptxVector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptx
 
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
 
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
 
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemkeProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
 
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
 
Mcleodganj Call Girls 🥰 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls 🥰 8617370543 Service Offer VIP Hot ModelMcleodganj Call Girls 🥰 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls 🥰 8617370543 Service Offer VIP Hot Model
 
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ..."I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of Terraform
 
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
 
Platformless Horizons for Digital Adaptability
Platformless Horizons for Digital AdaptabilityPlatformless Horizons for Digital Adaptability
Platformless Horizons for Digital Adaptability
 

Deep learning applications in e-commerce search: Dynamic talks Chicago 3/14/2019