AD ALTA
JOURNAL OF INTERDISCIPLINARY RESEARCH
Tab.1 Business continuity data for safety-related functions in a
selected public university (Source: own work based on data from
(Columbus Technical College, 2018)
Business
Function
Number
of
Human
Actors
Number
Of
Involved
Processes
Number
Of
Technical
Actors
UBFRP
(points)
Emergency
Communication
12
2.5
5.5
20
Public Information
10.5
4
5.5
20
Risk Management
3
2.5
3
8.5
Police and Security
9
2.5
1
12.5
Mail Services
12
3.5
1.5
17
Core IT Systems
9
2.5
2.5
14
Emergency Services
13.5
2.5
0
16
Based on the inferred UBFRP computations and according to the
BCPTs classifier (Podaras, 2018) specific impact value levels
can be assigned for every individual function which can then be
compared to the proposed by the university BCM experts impact
value levels. These levels do not appear in the utilized university
BCM guide but they have been inferred via mapping the
proposed by the university BCM recovery team resumption
timeframes for these functions. According to this value a
corresponding IVL has been defined and compared to the
predicted IVL via the BCPTs approach and the currently
proposed business intelligence tool. In this way, the BCPTs
accuracy can be verified (Tab.2)
Tab.2 The comparison between the predicted IVL with the
proposed by university BCM team members IVL
Business
Function
UBFRP
(points)
Impact
Value Level
(IVL)
(BCPTs
prediction)
Proposed
Recovery
Time (by
the BCM
Team)
(hours)
Proposed
IVL
(by the
BCM
team)
(hours)
Emergency
Communication
20
L2
24
L2
Public Information
20
L2
24
L2
Risk
Management
8.5
L4
24
L2
Police and
Security
12.5
L3
48
L3
Mail Services
17
L2
24
L2
Core Technology
Infrastructure
14
L3
72
L3
Emergency
Services
16
L2
24
L2
From the above recorded predictions it can be concluded that the
BCPTs classifier and the incoproration of the currrent business
intelligence web application may infer highly accurate business
continuity management predictions for safety-related business
functions in public universities. Based on the selected case study
the predictive accuracy of the BCPTs classifier is 85.71% (6 out
of 7 criticality ranking predictions have been proved correct for
safety-related operations).
4.3 Data Mining Tasks
The exported spreadsheet data can be further used for machine
learning activities such as classification and regression
techniques. So far, the BCPTs speedy classifier has been based
exclusively on UBFRP input variable for the prediction of the
Impact Value Level. Based on the data exported by the currently
proposed BI solution (Tab.1), more input (explanatory) variabes
can be utilized for predicting criticality ranking for individual
business function. Moreover more assciation rules among the
incorporated variables can be explored. The machine learning
path for BCM knowledge discovery is illustrated (Fig.10)
Fig. 10.
The machine learning path for business continuity
predictive knowledge discovery via the proposed BCPTs BI tool.
Source: (Source: own work).
Regression analysis tasks can be also implemented in order to
predict recovery time values. However, this is not feasible when
relying exclusively on the speedy BCPTs classifier. Data mining
regression analysis tasks can be implemented after considering
the appropriate recovery scenarios, a set of technical,
environmenatal and unexpected recovery factors (TRF, ERF,
URF) and after estimating the Adjusted Points variable (ABFRP)
for conducting precise recovery time computations.
Another issue which requires further clarification is the
possibility to boost the predictive accuracy of the speedy
classifier. Several robust ensemble classification techniques,
such as random forests, k-folds cross validation, k-NN (nearest
neighbor), logistic regression and support vector machine (SVM)
may be incorporated. However, importing the conducted .csv
reports into a sophisticated machine learning software package is
also demanded. One of the most commonly utilized machine
learning free software tool is the R package (Rahlf, 2017).
Currently the possibility to connect the proposed business
intelligence tool with the R package for faster machine learning
activities is under consideration.
In order to test the predictive accuracy of the speedy classifier
(as explained in rule 1), the full dataset (42 critical business
functions) has been imported into the R-Package. We used the
CART decision tree algorithm (Breiman et al, 1984) along with
the 10-folds cross validation (Machine Learning Mastery, 2018)
and the random forests (Breiman, 2001) classification algorithms
in order to test the BCPTs accuracy in predicting Impact Value
Levels (else criticality ranking or recovery priorities) for the
entire data set. The ensemble machine learnning techniques have
been used to avoid overfitting. The evaluation metrics used for
measuring the accuracy of the three tested machine learning
techniques has been the confusion matrix. The advantage of the
data mining is the possibility to conduct additional classifiers
based on different inputa variables and to investigate diverse
association rules among the variables included in the data set.
For example, we could focus on exploring the accuracy in IVL
predictions based only on human actors or considering only the
number of involved processes as input. We may also use
association rules to explore relationships among the included
variables (Fig. 11).
Fig.11. The data mining classification procedure for critical
business functions based on a public university data set.
For practical demonstration, the entire dataset has been used to
predict IVL based on UBFRP input. The data set is composed of
42 records and 1 input (UBFRP) and 1 output variable (IVL).
The data preprocessing procedure includes importing data and
splitting data in a logical ratio beteween training and testing data.
In our example, the splitting ratio has been 0.7 (70% training
- 362 -