Personalization in multilevel e-commerce system
Lukas Grolig
1
Personalization
Is about providing customer his own user experience. Display features that he might
need and promote products might buy.
2
Multi-level marketing
Companies having multiple sales streams. Company is recruiting consultants - self
employed individuals offering products. In multi-level there are also people responsible
for recruitment of consultants as their main income. But company can also be selling
directly to the customers.
3
Problems
When designing personalization in multi-level there are many types of users with
completely different needs. Whole website can change so drastically that it is not about
sales anymore.
4
Target persons
«person»
V i s i t o r
Anonymous site visitor
«person»
Builder
Recruiter. Finds people
willing to sell the products.
Provides education about
p r o d u c t s .
«person»
C o n s u l t a n t
Shops on behalf of his
c u s to mers .
«person»
S h o p O w n e r
Buys products in batch to
resell then in his/hers
s t o r e .
«person»
D i r e c t C u s t o m e r
Browses website and
orders products for himself
or his family.
5
System Design (Context)
«person»
U s e r
Visiting site
«container»
E - C o m m e r c e We b s i t e
[Some classical technology
stack]
Provides user functionality
such as product catalogue,
detail, basket, checkout
process.
«container»
E-Commerce API
[Some microservice technology
stack]
Responsible for business
logic, providing and storing
data.
«container»
P e r s o n a l i z a t i o n
Engine
[ ]
Multiple systems to
provide user specific data.
«container»
Feature Flag Engine
[ ]
Turns on/off functionality
for specific segment.
U s e s
[HTTPS]
U s e s
[HTTPS]
U s e s
[HTTPS]
U s e s
[HTTPS]
U s e s
[HTTPS]
U s e s
[HTTPS]
6
Personalization Engine (Container)
« c o n t a i n e r »
P r o d u c t
P e r s o n a l i z a t i o n
[ C o n t e x t u a l B a n d i t ]
P r o v i d e s p r o d u c t s p e r
s e g m e n t .
« c o n t a i n e r »
F e a t u r e
P e r s o n a l i z a t i o n
[ C o n t e x t u a l B a n d i t ]
P r o v i d e s r i g h t f e a t u r e s f o r
t a r g e t p e r s o n a .
« c o n t a i n e r »
C u s t o m e r
S e g m e n t a t i o n
[ ]
C l a s s i f i c a t i o n m o d e l t o
p r o v i d e i n f o r m a t i o n a b o u t
c u s t o m e r s e g m e n t .
7
Segmentation
offline experimentational training
number of target groups selected by data scientist
manual analysis of each segment (based on sales, features used)
preparation for classification engine (defining features)
8
Classification
part of ML pipeline
based on segmentation analysis output, dataset is created after closings
classification is based on Catboost (gradient boosting)
models are trained, evaluated and deployed automatically
9
Contextual Bandit
« c o n t a i n e r »
S t a t e
[ ]
« c o n t a i n e r »
A c t i o n
[ ]
« c o n t a i n e r »
R e w a r d
[ ]
* * * * * * * *
* * * *
10
Modeling the Problem
Context: product category, season, region

Decision: products to display

Feasible actions: { category1, category2, bestsellers, trending }

Outcome: user adds product to basket or displays product detail page withing 20s

Reward: 3 when added to basket, 1 when detail displayed, 0 otherwise
11
Contextual Bandit (Context)
«container»
A p p
[ ]
«container»
P e r s o n a l i z a t i o n
S e r v i c e
[ ]
«container»
Join Service
[ ]
«container»
O n l i n e L e a r n e r
[ ]
«container»
O f f l i n e L e a r n e r
[ ]
«container»
D a t a s e t
[ ]
c o n t e x t d e c i s i o nr e w a r d
R e w a r d s
* * * *
T r a i n i n g d a t a
B e s t m o d e l
P a r a m e t e r s , M o d e l s
* * * *
12
Service Design
each service must use telemetry to trace request through the system
API to retrieve plain data is not accepting context
personalization is separate API on top of plain
13
Personalized data source
on top of data sources there can be personalizes data source accepting context
Problems:
personalized sources are much slower
caching can be challenging (memory consumption)
14
Personalized feature flags
on top of regular feature flag service there is another providing features based on
context
enables complete customization of browsing experience
affects frontend development and backend-for-frontend
nearly every UI component has it's own flag
components can be versioned (eg. suffix to feature folder)
15
Automated A/B Testing
it is crucial to specify right metrics (display time, time to action, selling targets)
each service must provide analytic outputs
multiple metric sources must be aggregated (traces, google tags...)
data are stored in a lake
using ETL raw data are processed and stored in the lake again
data are consumed by systems when needed
16
Join service
2 possible designs - offline based on ETL, online based on message hub (eg. Kafka)
in microservice world it is the most complex part of the system
heavy utilization of OpenTelemetry
join service takes nearly all events same as logging service does
17
Questions?
18
Thanks a lot.
19