A Recommender System Based on Omni-Channel Customer Data · Omni-Channel 1 2 1 2 1 1 PurchaseHistory purchases 2s 1v 2s 1v Weblog searchsandvisitv 1l 1l 1l 1p 1l SocialMedia postpandlikel

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


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



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


1. Omni-Channel


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


Website – Searches

* for confidentiallity reasons, the example shows an unrelated company


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:


Website – Visits



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:


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


Social Media

▶ Company posts

promotions to facebook

▶ Use facebook comments

and likes as an indication

of interest

▶ Identify users based on

name



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



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



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


2. Recommendations


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



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



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



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)


3. Evaluation


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



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



Recommendation Quality

SVD (ALS) SVD Co-Clustering Item-based CF Baseline

0

50

100

7.76

34.10

16.9711.11

49.99

rank


4. Conclusion


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



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

https://doi.org/10.1007/978-3-642-19460-3

http://sifter.org/simon/journal/20061211.html

https://doi.org/10.1109/ICDM.2008.22

https://doi.org/10.1109/ICDM.2005.14

Documents

A Recommender System Based on Omni-Channel Customer Data · Omni-Channel 1 2 1 2 1 1 PurchaseHistory purchases 2s 1v 2s 1v Weblog searchsandvisitv 1l 1l 1l 1p 1l SocialMedia postpandlikel