Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
A Recommender System Based on
Omni-Channel Customer DataMatthias Carnein, Leschek Homann, Heike Trautmann, Gottfried Vossen
University of Münster, Germany
17th July 2019
A Recommender System Based on Omni-Channel Customer Data
Recommender Systems
▶ Recommender systems suggest products to users that might be of interest to them
▶ They help customers sift through the product catalogue and discover new products
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 1
A Recommender System Based on Omni-Channel Customer Data
Case Study
▶ Spanish retailer in the home décor business
▶ > 100 stores and an online shop
▶ In total 2.2 million transactions and 800, 000members of a loyalty programme
female
90%
male
10%
Age
0 204060801000
50,000
100,000
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 2
A Recommender System Based on Omni-Channel Customer Data
User-Item Matrix
▶ State-of-the-art: analyse what ‘like-minded’ users enjoy
▶ Often based on explicit feedback (e.g. movie ratings)
▶ In retailing, interest needs to be inferred from implicit feedback, e.g. purchase history
▶ Preferences are stored in a large but sparse User-Item matrix
i1 i2 i3 i4 i5u1 3 4 4
u2 1 2 3
u3 5
u4 1 3 2 3
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 3
1. Omni-Channel
A Recommender System Based on Omni-Channel Customer Data
Omni-Channel
▶ Using the purchase history works well as an
indication of interest
▶ But there is much more information available to
infer preference
▶ Online Search and Surf Behaviour, Newsletters
and Social Media
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 5
Website – Searches
* for confidentiallity reasons, the example shows an unrelated company
A Recommender System Based on Omni-Channel Customer Data
Weblogs – Searches
▶ Use the search behaviour on the website as an indicator for interest
▶ Analyse the stored HTML-GET requests for search queries (‘q=’)▶ Match the query to product name, description or product family
▶ Use cookies to identify users after they logged in once
red toweluser search: GET/ www.url.com/q=red towelserver log:
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 7
Website – Visits
* for confidentiallity reasons, the example shows an unrelated company
A Recommender System Based on Omni-Channel Customer Data
Weblogs – Visits
▶ Use the website visits as an indication of interest
▶ Analyse the stored HTML-GET requests for product pages
▶ Use cookies to identify users after they logged in once
bagsvisited category: GET/ www.url.com/bags.htmlserver log:
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 9
Newsletter
▶ Company sends out
newsletters for new
products and discounts
▶ Record newsletter opens
and click throughs as an
indicaton of preference
▶ Identify users based on
personalised link / cookie
* for confidentiallity reasons, the example shows an unrelated company
Social Media
▶ Company posts
promotions to facebook
▶ Use facebook comments
and likes as an indication
of interest
▶ Identify users based on
name
* for confidentiallity reasons, the example shows an unrelated company
A Recommender System Based on Omni-Channel Customer Data
Omni-Channel
1
21 21 1
Purchase History
purchases
2s1v2s
1v
Weblog
search s and visit v
1l
1l1l
1p1l
Social Media
post p and like l
1c 2o1c
1o1c
Newsletter
open o and click c
i1 i2 i3 i4 i5u1 5 1u2 2 4 2u3 1 6u4 1 2 1
Preference Matrix
(weighted)
sum
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 12
A Recommender System Based on Omni-Channel Customer Data
Information per Channel
Purcha
ses
Websi
te Sear
ches
Websi
te Visit
s
News
letter
Open
News
letter
Click
Facebo
ok
1 Mio.
2 Mio.
3 Mio.
643,409
48,127
2,410,111
28,359
544,075 396,662Item
Ratings
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 13
A Recommender System Based on Omni-Channel Customer Data
Store Location
▶ Detect WiFi probe requests from
phones
▶ Identify users based on purchases
with MAC address present
▶ Possible Scenario: Trigger a
recommendation when customer
visited a store but left without a
purchase
▶ Because of privacy concerns, we
did not make use of this data
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 14
2. Recommendations
A Recommender System Based on Omni-Channel Customer Data
k-Nearest Neighbour (kNN)
▶ Weighted average of ‘like-minded’ users [Liu 2011]
▶ Improvement: Weighted average over items
i1 i2 i3 i4 i5u1 3 4 4
u2 1 ? 2 3
u3 5
u4 1 3 2 3
r2,3 = 3
User-based CF
i1 i2 i3 i4 i5u1 3 4 4
u2 1 ? 2 3
u3 5
u4 1 3 2 3
r2,3 = 3
Item-based CF
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 16
A Recommender System Based on Omni-Channel Customer Data
Singular Value Decomposition (SVD)
▶ Split the large User-Item matrix in two smaller matrices [Funk 2018]
▶ By detecting hidden characteristics that approximate the purchase behaviour
▶ Special case for implicit feedback using Alternating Least Squares [Hu et al. 2008]
i1 i2 i3 i4 i5u1
u2u3
u4
fu1 fu2u1
u2u3
u4
i1 i2 i3 i4 i5fi1fi2
≈ ×
User-Item Matrix User Factors
Item Factors
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 17
A Recommender System Based on Omni-Channel Customer Data
Co-Clustering
▶ Cluster users and items such that prediction error is minimized [George et al. 2005]
▶ Implicitly forms co-clusters of users and items
▶ Use the average rating of user, item, and User-Item cluster for prediction
Co-Cluster
User-Cluster Item-Cluster
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 18
A Recommender System Based on Omni-Channel Customer Data
Baseline
▶ For comparison we include a baseline algorithm
▶ Randomly recommends by sampling from the User-Item matrix
i1 i2 i3 i4 i5u1 3 0 4 0 4
u2 1 0 ? 2 3
u3 0 5 0 0 0
u4 1 0 3 2 3r2,3 = random(ru,i)
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 19
3. Evaluation
A Recommender System Based on Omni-Channel Customer Data
Masking
4 1
2 3 2
1 6
1 2 1
1 0 0 1 0
1 0 1 0 1
0 1 0 1 0
1 0 1 0 1
1 0 0 1 0
1 0 0 0 1
0 0 0 1 0
1 0 1 0 0
1 0 0 1 0
1 0 1 0 1
0 1 0 1 0
1 0 1 0 1
User-Item Matrix Fill Zero & Binarize
Train set
Test set
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 21
A Recommender System Based on Omni-Channel Customer Data
Rank Measure
▶ Ideal: Relevant products should be ranked higher
▶ Measure: Average percentile-rank of items where the user has interest
Product Percentile Rank Purchase Rank Measure
Chandellier 0% 3 0%
Blackout Curtain 3% 7
Wall Clock 6% 3 6%
⋮Towel 97% 7
Canvas Art 100% 7
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 22
A Recommender System Based on Omni-Channel Customer Data
Recommendation Quality
SVD (ALS) SVD Co-Clustering Item-based CF Baseline
0
50
100
7.76
34.10
16.9711.11
49.99
rank
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 23
4. Conclusion
A Recommender System Based on Omni-Channel Customer Data
Conclusion
▶ Recommender System based on Omni-Channel Customer Data
▶ Using information from purchases, search and browsing behaviour as well as
newsletter and Facebook interactions
▶ Implemented using a real-life case
▶ Very high recommendation accuracy
Outlook:
▶ Improve customer recognition across the different channels
▶ Observe recommendation performance in real-life
Matthias Carnein Omni-Channel Recommendations Evaluation Conclusion 25
A Recommender System Based on Omni-Channel Customer Data
References
[Liu 2011] Liu, Bing (2011).Web Data Mining. Exploring Hyperlinks, Contents, and Usage Data. Springer Berlin Heidelberg. isbn: 978-3-642-19459-7.doi: 10.1007/978-3-642-19460-3.
[Funk 2018] Funk, Simon (23rd Jan. 2018). Netflix Update: Try This at Home. url: http://sifter.org/simon/journal/20061211.html.[Hu et al. 2008] Hu, Yifan, Yehuda Koren and Chris Volinsky (Dec. 2008). ‘Collaborative Filtering for Implicit Feedback Datasets’.
In: 8th IEEE International Conference on Data Mining, pp. 263–272. doi: 10.1109/ICDM.2008.22.[George et al. 2005] George, Thomas and Srujana Merugu (2005). ‘A Scalable Collaborative Filtering Framework Based on Co-Clustering’.
In: Proceedings of the Fifth IEEE International Conference on Data Mining. ICDM ’05. Washington, DC, USA: IEEE Computer Society,pp. 625–628. isbn: 0-7695-2278-5. doi: 10.1109/ICDM.2005.14.
Matthias Carnein 26