An informative talk about deep learning and its potential uses in recommender systems. Presented at the Budapest Startup Safary, 21 April, 2016. The breakthroughs of the last decade in neural network research and the quick increasing of computational power resulted in the revival of deep neural networks and the field focusing on their training: deep learning. Deep learning methods have succeeded in complex tasks where other machine learning methods have failed, such as computer vision and natural language processing. Recently deep learning has began to gain ground in recommender systems as well. This talk introduces deep learning and its applications, with emphasis on how deep learning methods can solve long standing recommendation problems.

1. Deep learning:
the future of recommendations
Balázs Hidasi
Head of Data Mining and Research
Gravity meetup @ Startup Safary
April 21, 2016

2. Deep learning in the headlines

3. Deep learning in the background
• Life improving services
 Speech recognition
 Personal assistants (e.g. Siri,
Cortana)
 Computer vision, object
recognition
 Machine translation
 Chatbot technology
 Natural Language Processing
 Face recognition
 Self driving cars
• For fun
 Text generation
 Composing music
 Painting pictures
 Etc.

4. What is deep learning?
• A class of machine learning algorithms
 that use a cascade of multiple non-linear processing layers
 and complex model structures
 to learn different representations of the data in each layer
 where higher level features are derived from lower level
features
 to form a hierarchical representation.

5. Deep learning is not a new topic
• First deep network proposed in the 1970s
• More papers in the 80s and 90s
• Why now?
 Older research was not used widely in practice
 Applications were much more simplistic that today’s

6. Neural networks: a brief overview

7. Neurons, neural networks
• Neuron: rough abstraction of the human neuron
 Receives inputs (signals)
 Sum weighted inputs is big enough  signal
 Amplifiers and inhibitors
 Basic pattern recognition
• Neural network: neurons connected to one another
• Feedforward networks: neurons are organized into
layers
 Connections only between subsequent layers
𝑦
𝑥1
𝑥2
𝑥3
𝑥4
𝑓(. )
𝑖=1
𝑁
𝑤𝑖 𝑥𝑖 + 𝑏
𝑥1
𝑥2
𝑥3
ℎ1
1
ℎ2
1
ℎ3
1
ℎ1
2
ℎ2
2

8. Networks that big enough: go deep not wide
• Feedforward neural networks are universal
approximators
 Can imitate any function if they are big enough
 (Also needs enough in-out pairs to learn)
• What is big enough?
 Number of layers / neurons
 Theoretical „big enough” conditions massively overshoot
• Go deep, not wide
 The number of neurons required for good approximation is
polynomial in the input if the network is deep enough
 Otherwise it is exponential

9. Training neural networks
• Forward pass: get the current estimate of the target
o 𝑠𝑗
1
= 𝑖 𝑤𝑖,𝑗
1
𝑥𝑖 + 𝑏𝑗
1
; ℎ𝑗
1
= 𝑓 𝑠𝑗
1
o 𝑠 𝑘
2
= 𝑗 𝑤𝑗,𝑘
2
ℎ𝑗
1
+ 𝑏 𝑘
2
; ℎ 𝑘
2
= 𝑓 𝑠 𝑘
2
o …
o 𝑠𝑙
𝑂
= 𝑘 𝑤 𝑘,𝑙
𝑁+1
ℎ 𝑘
𝑁
+ 𝑏𝑙
𝑂
; 𝑦𝑙 = 𝑓 𝑠𝑙
𝑂
• Backward pass: correct weights to reduce error
 Gradient descentLayer Error Gradient
(w.r.t. weights between current and prev. layer)
Output Defined loss
(e.g. 𝐿 = 𝑖=1
𝑁 𝑜
𝑦𝑖 − 𝑦𝑖
2
)
𝜕𝐿
𝜕𝑤𝑗,𝑖
(𝑁+1)
=
𝜕𝐿
𝜕𝑦𝑖
∗
𝜕𝑦𝑖
𝜕𝑠𝑖
𝑂 ∗
𝜕𝑠𝑖
𝑂
𝜕𝑤𝑗,𝑖
𝑁+1 =
𝜕𝐿
𝜕𝑦𝑖
𝑓′
𝑠𝑖
𝑂
ℎ𝑗
𝑁
𝑁 𝑡ℎ
hidden
𝛿𝑖
𝑁
=
𝜕𝐿
𝜕𝑦𝑖
∗
𝜕𝑦𝑖
𝜕𝑠𝑖
𝑂
𝜕𝐿
𝜕𝑤 𝑘,𝑗
𝑁 =
𝑖
𝜕𝐿
𝜕𝑦𝑖
∗
𝜕𝑦𝑖
𝜕𝑠𝑖
𝑂 ∗
𝜕𝑠𝑖
𝑂
𝜕ℎ𝑗
𝑁 ∗
𝜕ℎ𝑗
𝑁
𝜕𝑠𝑗
𝑁 ∗
𝜕𝑠𝑗
𝑁
𝜕𝑤𝑗,𝑖
𝑁 =
𝑖
𝛿𝑖
𝑁
𝑤𝑖,𝑗
𝑁+1
𝑓′ 𝑠𝑗
𝑁
ℎ 𝑘
𝑁−1
(𝑁 −
𝛿𝑗
𝑁−1
=
𝑖
𝛿𝑖
𝑁
𝑤𝑖,𝑗
𝑁+1
𝑓′ 𝑠𝑗
𝑁
𝜕𝐿
𝜕𝑤𝑙,𝑘
𝑁−1 =
𝑗
𝛿𝑗
𝑁−1
𝑤𝑗,𝑘
𝑁
𝑓′ 𝑠 𝑘
𝑁−1
ℎ𝑙
𝑁−2
…
1 𝑠𝑡
hidden
𝛿 𝑘
1 𝜕𝐿
𝜕𝑤𝑖,𝑗
1 =
𝑘
𝛿 𝑘
1
𝑤 𝑘,𝑙
2
𝑓′
𝑠𝑗
1
𝑥𝑖

10. Challenges of training deep networks
• Saturation
• Vanishing gradients
• Overfitting
• Slowness of second order methods
• Slow convergence, stucks in local optima with first
order methods
• (Exploding gradients)

11. Why now?

12. Breakthroughs in research
• Saturation & vanishing gradients
 Layer-by-layer training (2006)
 Non-saturating activation functions, e.g. ReLU (2013)
• Overfitting
 Dropout (2014)
• Convergence problems
 Adagrad, Adadelta, Adam, RMSProp, etc.

13. Computational power
• Natural increase in computational power
• GP GPU technology

14. Intermission

15. Don’t give in to the HYPE
• Deep learning is impressive but
 deep learning is not true AI
o it may be a component of it when
and if AI is created
 deep learning is not how the human
brain works
 95% of machine learning tasks don’t
require deep learning
 deep learning requires a lot of
computational power
• Deep learning is a tool
 which is successful in certain,
previously very challenging domains
(speech recognition, computer
vision, NLP, etc.)
 that excels in pattern recognition
You are here

16. Deep learning for RecSys

17. From the Netflix prize...
• Netflix prize (2006-2009)
 Gave a huge push to recommender systems research
 Determined the direction of research for years
 Task:
o Some (User, Item, Rating) known triplets
o (User, Item) pairs with unknown rating
o Predict the missing ratings (1-5)

18. ... to recommenders in practice
• Ratings  events [implicit feedback]
 Lots of services don’t allow for rating
 Majority of users don’t rate
 Monitored passively  preferences have to be infered
• Rating prediction  ranking [top N recommendations]
 All that matters is the relevancy of the top N items
 Rating prediction is biased
• User  session / situation [session-based / context-driven
recommendation]
 Users are not logged in, identification is unreliable
 Accounts used by multiple users
 Aim of the session (e.g. buy a good laptop)
 Similar behavior of different users in a situation, different behavior of the same
user in different situations

19. Challenges in RecSys
• Session modeling
 Most of the algorithms are personalized
 A few are item-to-item
o Recommends similar items
o Also used for session-based recommendations (industry de facto standard)
 There are no good session based solutions
• Incorporating factors that influence user clicks
 Users click based on what they see
o Title
o Product image
o Description
 and on their knowledge of the product
o Usually harder to model
o Except when the product is content (e.g. music)

20. Deep learning to the rescue – Session modeling
• Recurrent Neural Networks (RNN)
 Sequence modeling
 Hidden state: next state is based on the previous hidden state and the current input
 „Infinite” depth
 More sophisticated versions: GRU, LSTM
• Needs to be adapted to the recommendation task
• GRU4Rec:
 Session-parallel minibatch training for handling the large variance in session lengths
 Sampling the output for reasonable training times, without losing much accuracy
 Ranking loss for better item ranking
• Results: 15-30% improvement over item-to-item recommendations
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
RSC15 VIDEO
Recall@20
Item-kNN
GRU4Rec
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
RSC15 VIDEO
MRR@20
Item-kNN
GRU4Rec

21. Other uses of deep learning for recsys
• Incorporating content directly
 Music, images, video, text
 User influencing aspects of the items
 Direct content representation
• Context-state modeling from sensory data
 IoT devices
 Lot of sensory data
 Some missing and noise
 Infer context state and recommend accordingly
• Interactive recommenders using chatbots
• Personalized content generation
 Today’s news
 Images in personalized style with personalized content
• Etc...

22. There is work to be done
• DL + RecSys research: just started
 Last year:
o 0 long papers, 1 short paper and 1 poster that is loosely connected
 This year:
o 10+ submissions to RecSys in this topic
o DLRS 2016 workshop @ RecSys
• Open questions
 (More) Application areas
 Adaptations required for the recsys problem
 Scalability
 Best practices
 ...

23. Thanks for your attention!