This talk shall focus on making real-time pipelines using cutting edge Big Data technologies and applying ML on gathered data. The first part of the presentation shall cover importance and necessity for streaming data processing. In addition, tools that could be used in order to build a streaming pipeline shall be proposed. The second part of this talk shall focus on making machine learning models in customer support. There shall be introduced success stories covering the need for more efficient customer support, problem resolution and gained benefits.
4. Our products
SOLUTIONS IPI – Ibis
Performance
Insights
FOI -
Field
Operation
s
Insights
Description
Ibis
Risk Analytics
Technologi
es
iCEM Mobile Network
Advanced
Analytics (5G)
Machine
Learning use cases
Ibis
Data Lake
E2E Network
monitoring and
analytics solution.
Deep monitoring for:
Docsis/HFC, IPDR, PNM,
WiFi optimization, MPL
S,
UPS, DSL, GPON, TR069,
Mobile Network.
Centralized
monitoring and
control of Field
Operations.
Analytics over
measurement data.
Risk analysis for
banks:
- Calculation of
IFRS9 in one
system.
-Scoring/Rating
tool.
- EWS tool.
Ibis Data Lake - a
reference
architecture for
implementing
Data Lake by Ibis
best practice,
allows data-based
innovation.
A modern solution
for managing
customer
experience based on
analytics for
Telecommunication
providers and other
industries.
An advanced
mobile network
analysis, with a
focus on today's
and future
challenges.
Realization of
different use cases
based on machine
learning such as:
- Churn prevention,
- Smart pricing
- Next best offer.
Work in progress
12. REAL TIME PROCESSING
• Process continuous data streams
• Reduce time increase
information value
• Filter only useful bits
• Streaming is a much more natural
model
REALTIMEPROCESSING
13. APACHE FLINK
• Framework and distributed processing
engine
• Process Unbounded and Bounded Data
• Leverage In-Memory Performance
• Building Blocks for Streaming Applications
• How does Flink support data pipelines?
APACHEFLINK
19. EKWB
• EKWB is a premium liquid cooling manufacturer
• Founded in 1999 by Edvard König
• Their products are available in more than 30 countries worldwide
• EKWB offers a full range of products for end users and enterprise as well
• Some of the products are cooling systems (CPU and GPU), fittings, radiators,
water blocks, reservoirs, pumps
EKWB
20. • EKWB uses Zendesk customer
support software
• It offers interactions between
customers and EKWB support team
through so called tickets
• Customers create tickets where they
describe their questions with EKWB
products/orders
Zendesk
ZENDESK
21. Tickets
• Tickets are resolved by EKWB support groups
• There are two EKWB support groups:
• Technical
• Shipping
• There are three different categories of tickets:
• Technical
• Shipping
• AND GENERAL!!!
TICKETS
22. Challenges
Manual assignment of tickets to
support groups
Faster reply time to more critical
tickets
Which agent is going to reply
unassigned tickets
CHALLENGES
23. Solution
• Automatic ticket classification to Support or Technical
team
• Calculation of ticket priority
• Algorithm for automatic ticket assignment
SOLUTION
IS IT REALLY ?!?
84%
57%
25%
91%
55%
30%
25. Support Team Classification
• Challenges:
• Preparing training and test set
• Validation of results
• Implementation of model in working process
• It‘s a NLP problem!
• We used BERT embeddings for text
representations
• Model classifies tickets with SVM algorithm
• We achieved balanced accuracy of 89,33%!
SUPPORTTEAMCLASSIFICATION
AND WE ARE IMPROVING
IT!!!
26. Automatic Priority Calculation
• Goal was to prepare ticket priority
evaluation
• EKWB team had to mark tickets on
scale 1-5 for few months!
• Tickets need to be ranked by their
priority so that tickets with higher
priority are solved sooner
• QUESTION: classification || regression?
• Model evaluates ticket on scale 1-100%
PRIORITYCALCULATION
4 5
6,4% 40,2
%
5
34,7%
27. Automatic Agent Assignment
• Who is going to solve next ticket?
• Person with the least tickets?
• Person with least high priority tickets?
• How to know which agent is active?
• Our system smartly assigns tickets to one of
active agents
• Trying to find balance
AGENTASSIGNMENT
28. Future Work
• Dockerize everything!!!
• Automatic cancelation of orders based on tickets
• Connection with ERP system
• Automatic responses
(or at least semi automatic)
• DEEP THOUGHT
*Hitchhikers guide to the galaxy
FUTUREWORK