Page 37 - Fister jr., Iztok, and Andrej Brodnik (eds.). StuCoSReC. Proceedings of the 2018 5th Student Computer Science Research Conference. Koper: University of Primorska Press, 2018
P. 37
as soon as an internal node appears which has all its children Table 1. Discretization of numeric attributes
expanded into leaves, pruning begins: the algorithm checks
whether that node is better replaced by a single leaf. This is just Inches Weight Price_euros
the standard “subtree replacement” operation of decision-tree
pruning, and the PART implementation makes the decision in Bins Name Bins Name Bins Name
exactly the same way as C4.5. If replacement is performed the
algorithm backtracks in the standard way, exploring siblings of [10.1 – 14] Small [0.69 – 1.75] Light [174 – 800] Low
the newly-replaced node. However, if during backtracking a node
is encountered whose every child is not a leaf (and this will (14 – 16] Standard (1.75 – 2.21] Medium (800 – 1300] Mid-range
happen as soon as a potential subtree replacement is not
performed) then the remaining subsets are left unexplored and the (16 – 18.4] Gaming (2.21 – 4.7] Heavy (1300 – 6099] High
corresponding subtrees are left undefined. Due to the recursive
structure of the algorithm this event automatically terminates tree In Table 2, the best class association rules which have at least
generation. Additional details about the PART algorithm can be 90% confidence are shown according to the Apriori algorithm.
found in [5]. Table 3 shows the classification rules which are found by the
PART algorithm (Conf. and Cov. In these tables stand for
4. EXPERIMENTAL EVALUATION AND confidence or accuracy and coverage, respectively).
RESULTS
Results show individual class association rules are more accurate
In this paper, we used a real-life dataset [13] called “Laptop (confident) and have a higher coverage compared to individual
prices” to illustrate the class association and classification rules classification rules. On the other hand, classification rules cover
learning process. This dataset consists of 1.304 examples which the search space “completely” and do not overlap while class
have 11 “independent” attributes each and one class attribute. association rules may (heavily) overlap and are not guaranteed to
Three numeric attributes (Inches, Weight, and Price_euros) are cover all the examples.
discretized into nominal attributes (as illustrated in Table 1).
Table 2. Class Association Rules found by the Apriori algorithm Conf. Cov.
Class association rules 100% 114
100% 110
{ Inches=Standard, Cpu=Intel Core i3} ==> { Price_euros=Low } 100% 89
{ ProductTypeName=Notebook, Inches=Standard, Cpu=Intel Core i3} ==> { Price_euros=Low } 100% 85
{ Inches=Standard, Cpu=Intel Core i3, Ram=4GB } ==> { Price_euros=Low } 100% 84
{ ProductTypeName=Notebook, Inches=Standard, Cpu=Intel Core i3, Ram=4GB } ==> { Price_euros=Low } 100% 82
{ Inches=Standard, Cpu=Intel Core i3, OpSys=Windows 10} ==> {Price_euros=Low} 100% 76
{ Inches=Standard, Cpu=Intel Core i3, Gpu=Intel } ==> {Price_euros=Low} 100% 67
{ Inches=Standard, Cpu=Intel Core i3, Memory=HDD } ==> { Price_euros=Low } 100% 67
{ ScreenResolution=1366x768, Cpu=Intel Celeron, Gpu=Intel } ==> { Price_euros=Low } 94% 72
{ Inches=Standard, Cpu=Intel Core i3, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia } 94% 72
==> { Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia, 94% 143
OpSys=Windows 10} ==> { Price_euros=High } 94% 142
{ Inches=Standard, Ram=4GB, Memory=HDD, Weight=Medium } ==> { Price_euros=Low } 94% 71
{ ProductTypeName=Notebook, Ram=4GB, Memory=HDD, Weight=Medium }==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD} 94% 71
==> { Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD, 93% 120
OpSys=Windows 10} ==> { Price_euros=High } 93% 112
{ ScreenResolution=1366x768, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Notebook, ScreenResolution=1366x768, Ram=4GB, Weight=Medium } 93% 120
==> { Price_euros=Low } 93% 120
{ Inches=Standard, ScreenResolution=1366x768, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Notebook, Inches=Standard, ScreenResolution=1366x768, Ram=4GB, Weight=Medium} 93% 75
==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD } 93% 75
==> { Price_euros=High } 93% 75
{ ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia }==> {Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, OpSys=Windows 10} 93% 74
==> {Price_euros=High}
{ ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia } 93% 72
==> {Price_euros=High}
{ ProductTypeName=Gaming, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia, OpSys=Windows 10, Weight=Heavy }
==> {Price_euros=High}
StuCoSReC Proceedings of the 2018 5th Student Computer Science Research Conference 37
Ljubljana, Slovenia, 9 October
expanded into leaves, pruning begins: the algorithm checks
whether that node is better replaced by a single leaf. This is just Inches Weight Price_euros
the standard “subtree replacement” operation of decision-tree
pruning, and the PART implementation makes the decision in Bins Name Bins Name Bins Name
exactly the same way as C4.5. If replacement is performed the
algorithm backtracks in the standard way, exploring siblings of [10.1 – 14] Small [0.69 – 1.75] Light [174 – 800] Low
the newly-replaced node. However, if during backtracking a node
is encountered whose every child is not a leaf (and this will (14 – 16] Standard (1.75 – 2.21] Medium (800 – 1300] Mid-range
happen as soon as a potential subtree replacement is not
performed) then the remaining subsets are left unexplored and the (16 – 18.4] Gaming (2.21 – 4.7] Heavy (1300 – 6099] High
corresponding subtrees are left undefined. Due to the recursive
structure of the algorithm this event automatically terminates tree In Table 2, the best class association rules which have at least
generation. Additional details about the PART algorithm can be 90% confidence are shown according to the Apriori algorithm.
found in [5]. Table 3 shows the classification rules which are found by the
PART algorithm (Conf. and Cov. In these tables stand for
4. EXPERIMENTAL EVALUATION AND confidence or accuracy and coverage, respectively).
RESULTS
Results show individual class association rules are more accurate
In this paper, we used a real-life dataset [13] called “Laptop (confident) and have a higher coverage compared to individual
prices” to illustrate the class association and classification rules classification rules. On the other hand, classification rules cover
learning process. This dataset consists of 1.304 examples which the search space “completely” and do not overlap while class
have 11 “independent” attributes each and one class attribute. association rules may (heavily) overlap and are not guaranteed to
Three numeric attributes (Inches, Weight, and Price_euros) are cover all the examples.
discretized into nominal attributes (as illustrated in Table 1).
Table 2. Class Association Rules found by the Apriori algorithm Conf. Cov.
Class association rules 100% 114
100% 110
{ Inches=Standard, Cpu=Intel Core i3} ==> { Price_euros=Low } 100% 89
{ ProductTypeName=Notebook, Inches=Standard, Cpu=Intel Core i3} ==> { Price_euros=Low } 100% 85
{ Inches=Standard, Cpu=Intel Core i3, Ram=4GB } ==> { Price_euros=Low } 100% 84
{ ProductTypeName=Notebook, Inches=Standard, Cpu=Intel Core i3, Ram=4GB } ==> { Price_euros=Low } 100% 82
{ Inches=Standard, Cpu=Intel Core i3, OpSys=Windows 10} ==> {Price_euros=Low} 100% 76
{ Inches=Standard, Cpu=Intel Core i3, Gpu=Intel } ==> {Price_euros=Low} 100% 67
{ Inches=Standard, Cpu=Intel Core i3, Memory=HDD } ==> { Price_euros=Low } 100% 67
{ ScreenResolution=1366x768, Cpu=Intel Celeron, Gpu=Intel } ==> { Price_euros=Low } 94% 72
{ Inches=Standard, Cpu=Intel Core i3, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia } 94% 72
==> { Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia, 94% 143
OpSys=Windows 10} ==> { Price_euros=High } 94% 142
{ Inches=Standard, Ram=4GB, Memory=HDD, Weight=Medium } ==> { Price_euros=Low } 94% 71
{ ProductTypeName=Notebook, Ram=4GB, Memory=HDD, Weight=Medium }==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD} 94% 71
==> { Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD, 93% 120
OpSys=Windows 10} ==> { Price_euros=High } 93% 112
{ ScreenResolution=1366x768, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Notebook, ScreenResolution=1366x768, Ram=4GB, Weight=Medium } 93% 120
==> { Price_euros=Low } 93% 120
{ Inches=Standard, ScreenResolution=1366x768, Ram=4GB, Weight=Medium } ==> { Price_euros=Low }
{ ProductTypeName=Notebook, Inches=Standard, ScreenResolution=1366x768, Ram=4GB, Weight=Medium} 93% 75
==> { Price_euros=Low }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD } 93% 75
==> { Price_euros=High } 93% 75
{ ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia }==> {Price_euros=High }
{ ProductTypeName=Gaming, ScreenResolution=Full HD, Ram=16GB, Memory=SSD + HDD, OpSys=Windows 10} 93% 74
==> {Price_euros=High}
{ ScreenResolution=Full HD, Cpu=Intel Core i7, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia } 93% 72
==> {Price_euros=High}
{ ProductTypeName=Gaming, Ram=16GB, Memory=SSD + HDD, Gpu=Nvidia, OpSys=Windows 10, Weight=Heavy }
==> {Price_euros=High}
StuCoSReC Proceedings of the 2018 5th Student Computer Science Research Conference 37
Ljubljana, Slovenia, 9 October
