1. Factorization Machines for Hybrid
Recommendation Systems Based
on Behavioral, Product, and
Customer Data
Stijn Geuens
2. Agenda
• PhD Trajectory
• Goals
• Research Questions
• Progress
• Future Work
RecSys 2015 s.geuens@ieseg.fr
3. PhD Trajectory
RecSys 2015 s.geuens@ieseg.fr
Computer
Science
Machine
Learning Math &
Statistics
Business
Expertise
Data
Engineering
Business
Analytics
Data
Science
5. Research Questions
RecSys 2015 s.geuens@ieseg.fr
Machine Learning What is the added value of combining different
data sources?
• More data beats better models (Halevy, Norveg, Pereira, 2009)
• Rich database
– Explicit Ratings
– Implicit Ratings
– Customer Data
– Product Data
– Context Data
• Different combination methods
6. Research Questions
RecSys 2015 s.geuens@ieseg.fr
How can we evaluate recommender systems in
online settings using business metrics?
• Collaboration with company
• Witch metric to optimize?
– Click rates
– conversion
– Turnover
– Loyalty
– Etc.
• Does a RecSys affect these business performance?
Business Analytics
7. Current Study
RecSys 2015 s.geuens@ieseg.fr
Factorization Machines for Hybrid
Recommendation Systems Based
on Behavioral, Product, and
Customer Data
8. Motivation
• Typologies of systems using different input data:
– Collaborative filtering, content-based, and hybrid (Adomavicius, Tuzhilin, 2005)
– Collaborative filtering, content-based, demographic, knowledge-based,
hybrid (Burke, 2000; Bobadilla et al. 2013)
• Each systems has its advantages and disadvantages
• Hybridization resolves these issues and leads to better performance
• More data trumps better models (Halevy, Norveg, Pereira, 2009)
• This study: Hybridization by combining different data sources
(customer, product, behavioral data) by feature combination using a
single state-of-the-art algorithm, factorization machines (FM)
Combining all different data sources in one algorithm is never
done before, especially not in factorization machines research
RecSys 2015 s.geuens@ieseg.fr
9. Factorization Machines (FM)
RecSys 2015 s.geuens@ieseg.fr
• Introduced by Rendle (2010)
• Based on Support Vector Machines (SVM) and factorization
models and combines the advantages of both.
• SVM: Works with any real valued feature vector, allowing to
integrated different data sources
• Factorization Models: Variable interaction is calculated based
on factorized parameters, allowing to estimate interaction
under huge sparsity, where SVM’s fail.
• General FM model equation of degree 2:
10. Algorithms
RecSys 2015 s.geuens@ieseg.fr
• 4 factorization machines
– 3 single data source FMs
• Behavioral data (FMBD)
• Customer data (FMCD)
• Product data (FMPD)
– 1 Hybrid FM based on the 3 distinct data sources (FMBD/CD/PD)
• 1 company used hybrid CF benchmark model
– Input user-item matrix (M), where each element is defined as follows:
11. Data
RecSys 2015 s.geuens@ieseg.fr
• 2 distinct data sets:
– Furniture: 5,368 users and 2,601 items
– Children’s clothing: 5,999 users and 4,372 items
14. Future Work: This study
RecSys 2015 s.geuens@ieseg.fr
• Preform grid search to identify witch data sources
are the most important (on data type level and
individual variable level)
• Creating a benchmark hybrid algorithm combining
results of different systems created based on each
of the data sources
• Evaluation based on other theoretical metrics
(precision, F1, AUC, diversity, novelty, etc.)
15. Future Work: PhD
RecSys 2015 s.geuens@ieseg.fr
• Implement model at the company and perform a real-life
A/B tests
– Email system
– Webshop
• Evaluation of the implemented algorithm in terms of
business metrics (click rates, conversion rates, turnover, loyalty,
etc.)
• Investigate which (combination of) business metrics
optimize(s) economic value of the RecSys in both short and
long term
• Investigate the impact of a RecSys on economic performance
of a company
16. RecSys 2015 s.geuens@ieseg.fr
Thank you for
your Attention
Contact:
Stijn Geuens (0)3.20.545.892
IESEG School of Management s.geuens@ieseg.fr
3 Rue de la Digue fr.linkedin.com/pub/stijn-geuens/
F-59000 Lille stijn.geuens
17. Advantages and disadvantages of
different systems
Pros Cons
Collaborative Filtering • No metadata
engineering needed
• Serendipity in results
• Adaptive
• Scalability
• Cold Start for new users
and items
• Long tail problem
• Stability
Content-based • Comparision between
items possible
• No metadata
engineering needed
• Adaptive
• Overspecialization
• Cold start for new users
• Collection of product
information
RecSys 2015 s.geuens@ieseg.fr
18. Advantages and disadvantages of
different systems
Pros Cons
Knowlegde-based • Deterministic
• No cold-start
• Knowledge engineering
requered
• Subjective
• Static
Demographic • No metadata
engineering needed
• Serendipity in results
• Long tail
• Cold start for new users
• Static
RecSys 2015 s.geuens@ieseg.fr
Editor's Notes
Let me start by showing you a well-know diagram about different domains of knowledge. In this picture, three distinct expertise domains are identified, being Computer science, Math and Statistics and Business Expertise.
As some of you probably already noticed I am affiliated with IESEG School of Management, an institution mainly focusing on Subject Matter Expertise, the yellow part of this diagram. How is it than that I am here presenting at RecSys, where the participants are consisting of mainly computer scientists and mathematicians and statisticians? Even tough creating algorithms is at the heart of Recommendation Systems, interests of the business people and mainly marketers is growing as well. Changes in retail models and expectations of clients are changing and so companies have to adapt to this given. Rather than living in an age of mass marketing, we are living in an age of mass personalization now. Nowadays people want the advantages of mass marketing, like low prices and availability, and the advantages of personalization. For some time now, marketers are facing this challenge and are reaching out to new techniques to mass personalize and of course, Recommendation systems are perfectly suited to fill this cavet.
So where does that possition me in this diagram. I would say, stuck in the middle just like you. Why? Well first altough the attention of the business world, machine learning is still at heart of recommendation systems. No algorithms without machine learning. Therefore the first part of my PhD mainly focusses on first: Understanding the most popular algorithms used, and second develloping new algorithms.
So far so good. But what makes my PhD Trajectory differ from the a pure machine learning path is the fact that I work in collaboration with a company, LaRedoute.fr.
The outcome of my PhD is not purely academicly defined, like writing papers and participating in conferences. Next to the academic aspects, it is also expected I (help) delivering a fully functioning recommendation systems with a good performance. Which can be situated more in the data engineering intersection of the graph.
So far, I mentioned the machine learning part and the traditional sofware parts of my PhD project, but what makes me really stuck in the middle a third traditional research part. Marketing managers want a fully functioning recommendation system and take this for granted, what they really interested in the performance and return of these systems. As data scientists mainly focus on traditional performance metrics like recall, F1, AUC, etc., managers are interested in EUROS. They want to know what the improvement in click rates and conversion rates are using these systems and so in the end how sales, revenue and customer loyalty are positively affected by recommendation systems. By execting A/B tests, analysing sales figures and conducting customer surveys, these aspects can be shown. During these few days, I got the impression that indeed a shift in the comunity is going on. Although algorithm developpement remains very important, more and more attention is drawn to the final goal of a recommender system, being helping the company to increase profit and so optimize business metrics.
In sum: My PhD project exists of 3 main parts being:
Developping new algorithms or a machine learning part;
Implementing sytems at La Redoute and so a traditional sofware part;
Analysing the impact of a recommenadation system on the comapany’s results and so a traditional research part.
The combination of these aspects makes this project a real data science project.