4. DM
Why should we care about deep learning based RS?
many successes in various application areas in both academia and industry.
Deep Learning techniques enables to exploit complex data
image, tags, text,
5. DM
이 논문이 너무 많은 모델을 다루고 있어서
모델의 성능이나, 자세한 알고리즘은 제외하고
Deep Learning 모델을 사용한 이유
이런 점을 Deep Learning을 이용해 해결하려 했다
같은 Motivation과 Insight만 note하겠습니다.
7. DM
Frequently used Deep Learning Techniques
• Multi-Layer Perceptron (MLP)
• Autoencoder (AE)
• Convolutional Neural Network (CNN)
• Recurrent Neural Network (RNN)
• Deep Semantic Similarity Model (DSSM)
• Restricted Boltzmann Machine (RBM)
• Neural Autoregressive Distribution Estimation (NADE)
• Generative Adversarial Network (GAN)
8. DM
Recommender systems as neural network models
1) Models Using Single Deep-Learning Techniques
a) exploit one of mentioned eight techniques
i) Extracting item features from images
ii) Recommendation with explanations
2) Composite Models
a) exploit more than one deep learning techniques
i) GAN using AE as generator, MLP as discriminator.
ii) applying extracted item features using CNN to RNN
9. DM
Integration of Deep learning with Traditional RS
1) Solely Deep-Learning based Models
a) Predicting scores using MLP
1) Coupled Models (Recommender model + Neural network model)
a) AutoEncoder + extracted image features with CNN
i) two model objectives are updated simultaneously (Tightly Coupled models)
ii) two step learning (Loosely Coupled models )
12. DMNeural Collaborative Filtering (NCF)
Motivation
Previous Matrix Factorization only consi
ders linear relationship between users a
nd items
𝑦 𝑢𝑖 = 𝑑𝑜𝑡 𝑢, 𝑖
Solution
Non-linear Relationship between users
and items
𝑦 𝑢𝑖 = 𝑓 𝑢, 𝑖 Θ
Where 𝑓 Θ is multi-layer perceptron.
User-Item interaction prediction
13. DMWide & Deep Learning
Motivation
Linear regression is good at
memorization
Deep MLP is good at generalization,
catching deep features.
Solution
Integrate linear regression and MLP
구글 플레이스토어에서, 쿼리에 맞
는 application하는 데 사용 중인
알고리즘
14. DMYoutube Video Recommendation (Covington et al.)
Motivation
There are too many features to tuned
Solution
MLP works quite well without
exhaustive feature engineering
Youtube에서 personalized video
recommendation을 할 때 사용중
인 알고리즘
16. DM추천 시스템에서 Autoencoder가 주로 사용되는 용
도
Predict interaction between user
and items
Motivation
Denosing, Reconstruction effects of
Autoencoder will be helpful in
revealing un-observed preferences of
users.
Feature Extraction
Motivation
Use Autoencoder to learn low-
dimension representation in the
bottleneck layer
Autoencoder는 주로 이 두 가지 방법으로 사용되고, IR과 Recsys
community에서는 상당히 기본적인 방법이 되었다.
21. DMConvolutional Neural Network is
Capable to extract low-dimensional feature from complex data
(image, fixed-size sequences…)
These side information(image) can be used to alleviate cold-start
problem.
Motivation
CNN-based recommender systems mainly aims to help recommendation quality
using side information
22. DMGeneral CNN based Recommender systems
Motivation
Images of items contains much
information
Solution
Use CNN to extract these information
into latent codes.
Image of item i
CNN
Latent code
Item vector of
item i
Output : Item representation
Addition
23. DM
Personalized Image Tag Recommendation
(Nguyen et al.)
Motivation
CNN can code image into vectors
User can be embed into vectors
using current ML techniques.
Image에 어울리는 tag를 generate
하기 위해 제안된 알고리즘
Image input
Convolution
Max-pooling User info
Tag Food Holid
ay
Love … Selfie
Score 0.3 0.1 0.5 … 0.8
Image에서 feature를 미리
추출해서 pretrained vector로
사용하는 경우도 많다.
24. DMCNN for Audio Feature Extraction
Motivation
Allivate cold-start problem using
Content information
Solution
Learn latent vector representation
using Audio features
Audio Features for item i
Convolution
𝑦𝑖′ 𝑦𝑖
𝑦𝑖 는 WMF model이 만든 item i의 latent
representation
26. DMRecurrent Neural Network
Mainly concerned with sequential behaviors
User/Item evolves with time (temporal dynamics)
Session based recommendations
rare user information is known
Sequential behavior become more important than normal cases.
Text understanding
Like CNN, RNN enables to exploit rich side information(text)
27. DMGRU-Rec (Hidasi et al.)
Motivation
user’s each click in web-page is
closely related with each others
Solution
Bidirectional GRU fits to model this
sequential behavior well
Input :
One-
hot
Enco
ded
Vect
or
Emb
eddin
g
Layer
GRU
Layer
GRU
Layer
GRU
Layer
…
Feed
forwa
rd
Layer
Outp
ut :
score
s on
items
28. DMRecurrent Recommender Network (Wu et al.)
Motivation
Users and items behavior change
over time
Solution
Models user and item behaviors
using LSTM
𝒓 𝒖𝒊|𝒕 = 𝒇 𝒖 𝒖, 𝒖 𝒖𝒕, 𝒗𝒊, 𝒗𝒊𝒕
29. DMGRU Multitask Learning (Bansal et al.)
Motivation
Rich text information will be helpful
Solution
Use GRU to encode text sequences
(review, description of items)
Text description
of item i
GRU
Latent code
Item vector of
item i
Output : Item representation
Addition
여러 representation을 합칠 때,
의외로 dot이나 elementwise
multiplication보다 단순 addition이
성능이 더 좋은 경우가 많은 것 같다.
34. DMDeep Structured Semantic Model
Motivation
User,item 둘 다 side-information을 갖
고 있다.
Solution
DSSM을 그대로 적용하자.
DSSM 자체는 similarity measure를
제외하면 Neural Collaborative
filtering, 혹은 MF with side
information의 generalization으로 볼
수 있다.
Query가 아닌 user의
information을 input으로
35. DMFuture research directions
Better understanding of users and items
딥러닝 모델은 이미지나 텍스트를 잘 모델링한다!(적어도 기존의 방법보다는) 하지만
이러한 모델링은 아직 미숙한 수준이며(Deep learning에서 이용되는 아이디어를 그대
로 가져온 수준), 더 잘 하는 방법이 있을 것이다.
Temporal Dynamics
Session based Recommendation systems, and recommender with temporal
dynamics are not novel research topics, but they are largely underinvestiagted.
36. DMFuture research directions
Cross-domain / Multi-task learning
User의 purchase behavior 등을 은행 등에서 활용하려는 시도가 있다.
한 도메인의 knowledge을, 다른 도메인에 적용하거나,
여러 도메인의 지식을 동시에 활용했을 시의 성능 향상을 기대할 수 있다.
Novel Evaluation Metrics
Accuracy, Relevance measures are not good for evaluating recommender.
Hinweis der Redaktion
토픽이 너무 많아서, 일단 1-4까지 먼저 Cover하는 것으로 하고, 5,6,7,8은 다음 번에 제가 발표할 때 하도록 할게요…;;;;