This document summarizes a research paper on personalized top-N sequential recommendation using convolutional sequence embedding. The paper proposes a model called Caser that uses horizontal and vertical convolutional filters to capture sequential patterns at different levels from user behavior data. Caser outperforms previous methods by modeling both general user preferences and sequential patterns in a unified framework. The document provides details on Caser's network architecture, training approach, and evaluation on real-world datasets showing it achieves better performance than baseline methods.

1. January 8, 2020 Page 1/27
Personalized Top-N Sequential Recommendation
via Convolutional Sequence Embedding (WSDM’18)
Jihoo Kim
datartist@hanyang.ac.kr
Dept. of Computer and Software, Hanyang University
Jiaxi Tang, Ke Wang
Simon Fraser University

2. January 8, 2020 Page 2/27
Jiaxi Tang
PhD Student
School of Computing Science
Simon Fraser University
Intern at Google AI
Research & Machine Intelligence Team
Ke Wang
Professor
School of Computing Science
Simon Fraser University
PhD, Georgia Institute of Technology
MS, Georgia Institute of Technology
Recent papers
Towards Neural Mixture Recommender for Long Range Dependent User Sequences (WWW’19)
Jiaxi Tang*, Francois Belletti*, Sagar Jain, Minmin Chen, Alex Beutel, Can Xu and Ed H. Chi
Ranking Distillation: Learning Compact Ranking Models With High Performance for Recommender System (KDD’18)
Jiaxi Tang, Ke Wang
Author

3. January 8, 2020 Page 3/27
Minimum qualifications:
• Currently enrolled in a Master’s or PhD degree in Computer Science or a related technical field.
• Experience (classroom/work) in Natural Language Understanding, Neural Networks, Computer Vision, Machine
Learning, Deep Learning, Algorithmic Foundations of Optimization, Data Science, Data Mining and/or Machine
Intelligence/Artificial Intelligence.
• Experience with one or more general purpose programming languages: Java, C++ or Python.
• Experience with research communities and/or efforts, including having published papers (being listed as author)
at conferences (e.g. NIPS, ICML, ACL, CVPR, etc).
About the job
Research and Machine Intelligence is a high impact team that’s building the next generation of intelligence and
language understanding for all Google products. To achieve this, we’re working on projects that utilize the latest
techniques in Artificial Intelligence, Machine Learning (including Deep Learning approaches like Google AI) and
Natural Language Understanding. We impact products across Google including Search, Maps and Google Now.
https://careers.google.com/jobs/results/136271419680924358-research-intern-2020/
Google AI Research Intern

4. January 8, 2020 Page 4/27
Contents
1. Introduction
1.1 Top-N Sequential Recommendation
1.2 Limitations of Previous Work
1.3 Contributions
2. Related Work
3. Proposed Methodology
3.1 Embedding Look-up
3.2 Convolutional Layers
3.3 Fully-connected Layers
3.4 Network Training
3.5 Recommendation
4. Experiments
4.1 Experimental Setup
4.2 Performance Comparison
4.3 Network Visualization

5. January 8, 2020 Page 5/27
User’s long term
and static behaviors
User’s short term
and dynamic behaviors
General
preferences
Sequential
patterns
<
always
After buying an iPhone, buy phone accessories
“I love Apple’s products”
vs
recent next
<Motivation>
1. Introduction

6. January 8, 2020 Page 6/27
1.1 Top-N Sequential Recommendation
Users
Items
Sequence
order
General preferences
Sequential patterns
Input Output
A list of items
for user u
<Top-N Sequential Recommendation>
<Notations>
1. Introduction

7. January 8, 2020 Page 7/27
1.2 Limitations of Previous Work
<Markov chain based model>
1) FPMC (Factorized Personalized Markov Chains) WWW’10
2) Fossil (Factorized Sequential Prediction with Item Similarity Model) ICDM’16
<Two major limitations>
1) Fail to model union-level* sequential patterns.
2) Fail to allow skip behaviors**.
milk flour
*Union-Level?
butter… …
**Skip behaviors?
… airport hotel
rest-
aurant
bar
attr-
action
not necessary
…
Figure 1
1. Introduction

8. January 8, 2020 Page 8/27
1.2 Limitations of Previous Work
To provide evidences of union-level influences and skip behaviors
minimum support count = 5
minimum confidence = 50%
X Y
sequence
Figure 2
Sequential Association
Rules
→
1. Introduction

9. January 8, 2020 Page 9/27
1.3 Contributions
Caser (ConvolutionAl Sequence Embedding Recommendation Model)
• Caser uses horizontal and vertical convolutional filters to capture sequential patterns
at point-level, union-level, and of skip behaviors.
• Caser models both users’ general preferences and sequential patterns, and
generalizes several existing state-of-the-art methods in a single unified framework.
• Caser outperforms state-of-the-art methods for top-N sequential recommendation on
real life data sets.
1. Introduction

10. January 8, 2020 Page 10/27
• Sequential pattern mining depends on the explicit representation of patterns, thus, could
miss patterns in unobserved states. (= could miss implicit patterns)
• CNN has been used to extract users’ preferences from their reviews. None of these works
is for sequential recommendation.
• RNN was used for session-based recommendation. It may not work well in sequential
recommendation, because not all adjacent actions have dependency relationships.
• Temporal recommendation is related but different problem. (Session-based is also different)
(ex. Recommend coffee in the morning, instead of evening.)
2. Related Work

11. January 8, 2020 Page 11/27
Figure 3
<Network Architecture of Caser>
3. Proposed Methodology

12. January 8, 2020 Page 12/27
The user 𝒖’s sequence
every 𝑳 successive
items
as input
their next 𝑻 items
as the targets
window of
size 𝑳 + 𝑻
The embedding for item 𝒊
d is the number of latent dimensions
𝑺 𝟏
𝒖
𝑺 𝟐
𝒖
𝑺 𝟑
𝒖
𝑺 𝟒
𝒖
𝑺 𝟓
𝒖
𝑬(𝒖,𝟑)
=
𝑸 𝑺 𝟏
𝒖
𝑸 𝑺 𝟐
𝒖
𝑬(𝒖,𝟒) =
𝑸 𝑺 𝟐
𝒖
𝑸 𝑺 𝟑
𝒖
𝑬(𝒖,𝟓) =
𝑸 𝑺 𝟑
𝒖
𝑸 𝑺 𝟒
𝒖
3.1 Embedding Look-up
3. Proposed Methodology

13. January 8, 2020 Page 13/27
image
local features
= 𝑳 × 𝒅 matrix 𝑬
= sequential pattern
Figure 4
Unlike image recognition,
“image” 𝑬 is not given…
and must be learnt
3.2 Convolutional Layers
3. Proposed Methodology

14. January 8, 2020 Page 14/27
𝑳 = 𝟒
𝒉 = 𝟐
𝒅 = 𝟑
𝑭 𝒌
∈ ℝ 𝟐×𝟑
𝒊 = 𝟏
𝒊 = 𝑳 − 𝒉 + 𝟏
= 𝟒 − 𝟐 + 𝟏
= 𝟑
𝑬 𝟏:𝟐
𝑬 𝟐:𝟑
𝑬 𝟑:𝟒
inner
product
activation
function
𝑖-th convolution value
<Max Pooling><Horizontal Filter>
𝑳 = 𝟒
𝒅 = 𝟑
෩𝑭 𝒌 ∈ ℝ 𝟒×𝟏
<Vertical Filter>
→ weighted sum
→ no max pooling
3. Proposed Methodology
𝑘-th filter
# of filter
height of filter
Convolution value (by 𝑭 𝒌
)

15. January 8, 2020 Page 15/27
activation function
convolutional
sequence embedding
3.3 Fully-connected Layers
the probability of
how likely user 𝒖 will interact
with item 𝒊
at time step 𝒕
3. Proposed Methodology

16. January 8, 2020 Page 16/27
union-level
sequential patterns
point-level
sequential patterns
short-term
sequential patterns
long-term
general preferences
3. Proposed Methodology

17. January 8, 2020 Page 17/27
3.4 Network Training
To train the network, we transform the values of the output layers to probabilities
sigmoid function
the collection of the time steps
for which we would like to make
predictions for user 𝒖
the likelihood of all sequences in the dataset
3. Proposed Methodology

18. January 8, 2020 Page 18/27
3.4 Network Training
To further capture skip behaviors, we could consider the next 𝑻 target items
Taking the negative logarithm of likelihood, we get the objective function “binary cross-entropy loss”
model parameters
hyper-parameters
are learned by minimizing the loss function (13)
are tuned on the validation set via grid search
3. Proposed Methodology

19. January 8, 2020 Page 19/27
3.5 Recommendation
After obtaining the trained neural network, to make recommendations for a user 𝒖 at time step 𝒕
We recommend 𝑵 items
that have the highest values
in the output layer 𝒖
𝒖’s last 𝑳 items’
embedding 𝑬(𝒖,𝒕)
𝒖’s latent
embedding 𝑷 𝒖
Input Output
3. Proposed Methodology

20. January 8, 2020 Page 20/27
4.1 Experimental Setup
<Datasets>
Amazon data was not used, due to its SI
0.0026 for ‘Office Products’
0.0019 for ‘Clothing’ / ‘Shoes’ / ‘Jewelry’ / ‘Video Games’
70% 10% 20%
validation testtraining
sequence
4. Experiments

21. January 8, 2020 Page 21/27
<Evaluation Metrics>
4.1 Experimental Setup
MAP(Mean Average Precision): the average of AP for all users
Precision, Recall
top 𝑵 predicted items
for a user
the last 20% of actions
in user’s sequence (= test set)
4. Experiments

22. January 8, 2020 Page 22/27
4.2 Performance Comparison
4. Experiments

23. January 8, 2020 Page 23/27
4.2 Performance Comparison
<Influence of hyper-parameter 𝒅, 𝑳, 𝑻,>
4. Experiments

24. January 8, 2020 Page 24/27
4.2 Performance Comparison
<Analysis of Caser Components>
𝒉 denotes horizontal convolutional layer
𝒗 denotes vertical convolutional layer
𝒑 denotes personalization
Any missing component is represented
by setting its corresponding 𝒐, ෥𝒐, 𝑷 𝒖 to zero.
4. Experiments

25. January 8, 2020 Page 25/27
4.3 Network Visualization
Caser puts more emphasis on recent actions,
demonstrating a major difference from the conventional top-N recommendation.
<Vertical convolutional filters>
4. Experiments

26. January 8, 2020 Page 26/27
4.3 Network Visualization
<Horizontal convolutional filters>
<Previous Sequence>
𝑺 𝟏 (13th Warrior) History
𝑺 𝟐 (American Beauty), Romance
𝑺 𝟑 (Star Trek), Action & SF
𝑺 𝟒 (Star Trek III)
𝑺 𝟓 (Star Trek IV)
<Predictions>
𝑹 𝟏 (Mad Max)
𝑹 𝟐 (Star War)
𝑹 𝟑 (Star Trek) >> Ground Truth
4. Experiments

27. January 8, 2020 Page 27/27
Thank you!
Q & A