2. Problem description, trends and challenges
Today’s challenge for Communication Service Providers
(CSPs) is to deliver high quality service with low operating
costs
With services not being limited to delivering basic
connectivity services, i.e. voice and data, the number of
service quality parameters to be measured has also
increased making analysis a complex and time consuming
task
This gets compounded with the fact that service quality
parameters have multi-dimensional sources
Network
IT infrastructure
Applications
Subscribers
3. Service Quality Analysis
Bayesian Network Approach
Mathematically proven Bayesian network algorithm can be used to
analyze service quality where
There is a lot of data (Big data)
As well as, missing data
Bayesian networks provide a well defined structure (as Directed Acyclic
Graphs) to represent the problem domain
The nodes represent the variables and the arcs represent the relationships
Information flow is omnidirectional
From Service quality perspective
Nodes represent the parameters
Arcs represent their relationships
4. Bayesian Networks for Service Quality Analysis
Example Use Cases
Service specific network route selection
In this use case, service quality parameter data
is available
Service specific international roaming list
prioritization
In this use case, there are several service
quality parameters but there is a lack of data
Here we need to take service quality indicators
as parameters.
5. Service Specific Network Route Selection
Traditional Approach
Real time services such as Video on Demand (VoD) require dedicated bandwidth to
the subscriber for the defined period
When subscriber requests VoD service, the service manager application requests the
network management layer to assign the bandwidth to provide the service
Traffic Engineering protocols like Resource Reservation Protocol (RSVP), select the
network routes which has less congestion and the required bandwidth required to
deliver service
However, this doesn’t take into account if the route selected is actually suited (based
on past history) for the required service (in our example Video on Demand)
This may lead to a low service quality experience if the link selected is not suited for
real time services leading to an unhappy subscriber
6. Service Specific Network Route Selection
Using Bayesian Network Prediction Models
We can deploy a Bayesian model to study the characteristics
of links and when required propose the suited resource path
based on the target service to be delivered
The parameters that define a network line characteristics are
Latency
Jitter
Reliability (packet drops)
Network line with high reliability (less packet drops) is more
suited for transactional applications e.g., online bank
transactions (even if the line faces latency problems)
Network line with low latency (and jitter) will be more suited for
real time applications like voice and video services (even if the
line reliability is not good)
7. Service Specific Network Route Selection
Using Bayesian Network Prediction Models (2)
In the figure below, we need to deliver Video on Demand from source to destination
with 2 routes connecting source to destination with equal bandwidth
Let’s represent the Bayesian network for Line A as example with parameters
Latency, Jitter and Packet Loss
Line A
Line A -
10 Gbps Jitter Latency Packet Drops Real Time Transactional
High High High 50% 50%
Low 10 90
Source Destination
Low High 60 40
Line A Low 30 70
Line B - Low High High 90 10
10 Gbps Low 60 40
Low High 90 10
Jitter
Jitter Low 50 50
Latency High Low
Packet Drops
High 70% 30%
Packet Drops
Low 40% 60%
High Low
50% 50%
Latency
High Low
Latency
Marginal probability
50% 50%
distribution
8. Service Specific Network Route Selection
Using Bayesian Network Prediction Models (3)
There is a 90% chance of Line A
being suited for transactional
services when there is high
latency and jitter and low packet
loss
Line A
Jitter Latency Packet Drops Real Time Transactional
Conditional probability High High High 50 50
There is a 70% distribution Low 10 90
chance of
experiencing Low High 60 40
high jitter when Line A Low 30 70
there is high
latency Low High High 90 10
Low 60 40
Jitter Low High 90 10
Jitter
Low 50 50
Latency High Low
Packet Drops
High 70 30
Low 40 60 Packet Drops
High Low
Latency 50 50
High Low
Latency
Marginal probability
50 50
distribution
9. Service Specific Network Route Selection
Using Bayesian Network Prediction Models (4)
When such Bayesian network models are deployed for each line, the models learn through
evidences from the network monitoring applications; the probabilities for the parameters
change based on usage experience
Thus, when a service is requested from end users, the network is better informed to make
the right resource selection thereby providing a predictable Quality of Service
Line A -
10 Gbps
Source Destination
Line B -
10 Gbps
10. Service specific international roaming list
prioritization
Mobile operators are facing a continuous decline in Average
Revenue Per User (ARPU)
With deregulations, competition is increasing and so is subscriber
churn
Operators look to focus on protecting high value subscribers and
look to offer high service quality for their premium base
International roaming being a high revenue and a key
service, roaming steering optimization is one of the challenges
operators face due to lack of quality data
Operators cannot tap network data from foreign networks their customers
have visited and connected
Operators apply business rules to prioritize international roaming lists
Bayesian Belief Network models provide a good platform where we
can work with lack of data to predict the most preferred roaming list
11. Service specific international roaming list
prioritization (2)
In this example, we build a Bayesian network for Operator A’s voice quality
Due to lack of roaming network quality data, we use the following indicators
Frequent Call Attempts (FCA) – by gauging the Call Detail Records (CDRs), this can be used as an
indication of multiple attempts to make a call due to network problems (coverage, handovers,…)
Manual Network Selection (MNS) – if users select a network which is not as per the prioritized
roaming list, it can be an indication that users prefer the selected network quality over the suggested
network while roaming
Average Call Duration (ACD) – the average call duration can be a good indicator when you compare
the subscriber’s home network average call duration to the roaming call duration
Operator A
Voice Quality Conditional probability
distribution
Frequent Call
Frequent Call Attempts
Attempts
Avg. Call Duration
High Low
Average Call Duration
70% 30%
High Low
50% 50%
Manual Network Selections
High Low Manual Network
Selections Marginal probability
50% 50%
distribution
12. Service specific international roaming list
prioritization (3)
There is a 90% chance of
We can compute the conditional probability of Operator Operator A’s voice quality being
A’s voice quality by taking evidence of the marginal good when there is high ACD
and FCA
probabilities of the voice quality indicative parameters
from home network databases e.g. HLR
Operator A Voice Quality
FCA MNS ACD Good Average Bad
High High High 70 20 10
Low 30 40 30
Operator A Low High 10 50 40
Voice Quality
Low 10 20 70
Low High High 90 10 0
Low 60 30 10
Low High 70 20 10
Frequent Call Low 33 34 33
Frequent Call Attempts
Attempts
High Low Avg. Call Duration
70% 30% Average Call Duration
High Low
50% 50%
Manual Network Selections
High Low Manual Network
Selections
50% 50%
13. Service specific international roaming list
prioritization (4)
Applying such models to the operator list, we can derive a dynamic roaming
steering list based on probabilities learnt from the Bayesian network models about
operator’s service quality indicators.
Voice Quality
Operator
Good Average Poor
B 70 20 10
C 55 25 20
.
. A 10 30 60
. … … …
.
14. Conclusion
The examples in this concept presentation illustrate
the generic nature of Bayesian network algorithm
and it’s applications to various data driven analysis
Both examples show how using Bayesian network
models can help predict service quality in cases
where there is a lot of evidence data and where
there is missing data
Deploying such prediction models with existing
applications, both datacom and telecom operators
can leverage the data analysis to improve service
quality (rather predict service quality)