and Biological Sciences www.textroad.com
*Corresponding Author: Asadollah Shahbahrami, Department of Computer Engineering, Faculty of Engineering,
Evaluation of Different Data Mining Algorithms to Predict Earthquakes Using Seismic Hazard Data
Asadollah Shahbahrami and Zinat Mehdidoust Jalali Department of Computer Engineering, Faculty of Engineering,
University of Guilan, Rasht, Iran
Received: September 11, 2016 Accepted: November 22, 2016
ABSTRACT
Several methods such as mathematical modeling, ionosphere analysis, and expert systems have been proposing to predict an earthquake. The accuracy of expert systems is usually higher than other techniques. Different algorithms, such as neural networks, decision tree, Naive Bayes and Support Vector Machine (SVM) have been using in expert systems. In order to evaluate these data mining algorithms in terms of accuracy and computational time, a practical example is simulated on the Rapid Miner platform using seismic hazard data.
Simulation results show that the SVM is the fastest algorithm and it has the lowest accuracy, while multilayer perceptron neural network has the highest accuracy. In order to increase the accuracy of SVM algorithm, we optimize this algorithm using Particle Swarm Optimization (PSO). After optimization, SVM-PSO has the highest accuracy compared to other algorithms.
KEYWORDS: Data Mining Algorithm, Earthquake, Expert System, Prediction, Support Vector Machine.e
1. INTRODUCTION
Humans face different natural crises such as earthquake, flood, fire and volcano in their life [1, 2]. These crises affect their life and impose irreparable damages. If we can predict the occurrence of crises such as earthquakes, it certainly reduces damages. Several methods such as mathematical modeling, ionosphere analysis and studying the animal behavioral changes have been performed to predict earthquakes. Some of these methods usually use a single feature and cannot use different features for earthquake prediction. Therefore, whether an earthquake is going to occur in a region or time interval cannot be identified through focusing on a single feature. An acceptable technique shall be achieved for earthquake prediction through simultaneous studying of some features, which could be realized using expert systems [3-5].
Recently, some expert systems have been proposed which have the ability to use more data and features.
For example, in hydrology analysis usually considered changes in water ions to predict earthquakes and other features such as fault behavior, changes in sea level are not overseen to predict. Some expert systems predict the future earthquakes using time, location and depth data of the previous earthquakes. Expert systems usually use different algorithms such as artificial neural network, decision tree, Support Vector Machine (SVM) [6, 7] and Naive Bayes [8]. The accuracy of these algorithms’ prediction is different due to dependence on the type of data and features which have been used in them such as seismic, energy, pulse and number of bumps [9].
The purpose of this paper is to study and evaluate some of these data mining algorithms in terms of accuracy and computational time. A practical example in the field of earthquake prediction was simulated on the Rapid Miner platform using real data. Simulation results show that SVM algorithm has the lowest accuracy while Multilayer Perceptron (MLP) neural network algorithm has the highest accuracy. In addition, SVM is the fastest algorithm. In order to increase the accuracy of SVM algorithm, we have optimized this algorithm using Particle Swarm Optimization (PSO) [10]. After optimizing this algorithm, simulation results show that SVM- PSO algorithm has the highest accuracy compared to other algorithms.
The remainder of the paper is organized as follows. Section 2 gives an insight into related works concerning the usage of machine learning methods in earthquake prediction and Section 3 expresses process of earthquake prediction, simulation platform, data set information and the simulation results. Section 4 illustrates discussion. Finally, Section 5 summarizes the work which has been performed.
2. RELATED WORKS
The comparison of some expert systems in terms of the magnitude of the earthquake and the used data that are applied to predict earthquakes is shown in Table 1. As can be seen in the table, eight expert systems are introduced. The simulation platforms of some of these systems have been Java, Rapid Miner and Matlab. Some of these systems predict earthquake by about a high probability because if more amount of features are used, the obtained accuracy will be higher. As indicated in Table 1, in these expert systems, different algorithms such as neural networks and decision tree have been used.
Table 1. The introduction of expert systems to predict earthquakes
References Earthquake magnitude Country Used data The probability of occurring
(percentage)
Simulation platform [11, 12] M≥5 USA and Taiwan Pacific Earthquake Engineering Research
Center database (contains 1082 records)
- -
[13] M≥3.6 Global, the number of
datasets from different parts of the world.
Data for a period of 43 years (1970 to 2012) The top 95 percent of earthquakes before they occur within a maximum of 15 hours.
Java and Rapid Miner
[14] M≥3.6 Global United States Geological Survey, advanced
financial system and seismic data for a period of time (from 1972 to 2013
The top 95 percent for up to 12 hours prior to the earthquake.
Java and Rapid Miner
[15, 16, 17] 4.53≤M≤7.9 USA and Europe The data set contains 1482 records from 1068 stations and 94 major earthquakes events.
- -
[18] All types of earthquakes Greece Seismic Institute, National Observatory of Athens
80.55 -
M≥5.2 Greece Seismic Institute, National Observatory of Athens
58.02 and 52.81 -
All types of earthquakes Greece Seismic Institute, National Observatory of Athens, the main signals and random signals
Up 60 -
M≥5.2 Greece Seismic Institute, National Observatory of Athens, the main signals and random signals
84.01 -
[19] - Iran The earthquake research institutes of
Washington and the information of Iran's earthquakes.
MLP neural network 71.4286 and classification neuro-fuzzy 82.8571
Matlab
[20] M≤1.5 , M≥1.5 - 155 samples and 15 characteristics. A total of
147 samples for (M≤1.5) and eight number for (M≥1.5).
Rough set 88.39 and decision tree 93.55
Java
[21] M≥5 Japan - 73 -
All types of earthquakes Japan - 67 -
3. Expert System Simulation to Predict Earthquake
In this section, we first describe process of earthquake prediction. Then, data set information and learning algorithm are discussed. Last, performance evaluation parameters are estimated.
3.1. Process of Earthquake Prediction using Seismic Hazard Data
An expert system was designed to predict the earthquake crises in Poland coal mines on the Rapid Miner platform [22-24]. First, the input values are extracted from a scientific source and then logged into the software as input. Based on the input, the percent of earthquake occurrence probability is calculated as the output. Block diagram of using data mining platform such as Rapid Miner to predict earthquake using seismic hazard data is depicted in Fig. 1. First, the training data and defined features are logged into Rapid Miner platform. Then, some preprocessing of operation such as normalization, type conversion and data cleaning are performed. Finally, data mining algorithms are trained using processed data and features. In testing stage, testing data and features are given to the system and the algorithm based on their training predict earthquake.
Fig. 1. Block diagram of using data mining platform such as Rapid Miner to predict earthquake.
3.2. Data Set Information
It should be noted that a total of 2584 data was used. Table 2 shows the features of data derived from scientific sources. In order to predict earthquake, features according to the geological conditions of mines in Poland were used, as each earthquake depending on the location and geological features had some unique characteristics.
Table 2. Data Set Information [24]
Data Set Information Data description
Data Set Characteristics Multivariate
Number of Instances (Row) 2584
Area Polish
Attribute Characteristics Real
Number of Attributes 19
In the collected data, each sample and instance is a summary statement of seismic activity related to one shift that is 8 hours [23]. A sample of data has been presented in appendix A. Attribute information of data is depicted in Table 3 [24].
Table 3. Attribute description [24].
Attribute Description of the attribute
Seismic (a - lack of hazard, b - low hazard, c - high hazard, d - danger state);
Seismoacoustic Result of shift seismic hazard assessment;
Shift The type of a shift (W - coal-getting, N -preparation shift);
G energy Seismic energy that is recorded by geophones
G puls A number of pulses that is recorded within the previous shift by Geophone Maximum (G Max);
Gd energy A deviation of energy that is recorded by geophones Gd puls A deviation of a number of pulses recorded by geophones G hazard Result of shift seismic hazard assessment coming for G Max only;
N bumps The number of seismic bumps recorded within previous shift;
Nbumps2 The number of seismic bumps (in the energy range [10^2, 10^3]) Nbumps3 The number of seismic bumps (in the energy range [10^3, 10^4]) Nbumps4 The number of seismic bumps (in the energy range [10^4, 10^5]) Nbumps5 The number of seismic bumps (in the energy range [10^5, 10^6]) Nbumps6 The number of seismic bumps (in the energy range [10^6, 10^7]) Nbumps7 The number of seismic bumps (in the energy range [10^7, 10^8]) Nbumps89 The number of seismic bumps (in the energy range [10^8, 10^10]) Energy Total energy of seismic bumps
Max energy The maximum energy of seismic bumps Class '1'- ('hazardous state'), '0' - ('non-hazardous state').
In addition to the features listed in Table 3, there are other features such as changes in groundwater levels [25, 26], changes in the critical frequency deviation of ionosphere layer [27] and changes in the behavior of animals [28], as any earthquake depending on the location and geological characteristics has some unique features [13, 29]. For example, rule-based regression algorithm and rule-based classification algorithm which utilized features such as seismic, energy and the number of pulses were used [23].
3.3. Learning Algorithm
As already explained, different data mining algorithms were used in the proposed system which their conditions are as follows. Neural network consists of the optimal values of the momentum term and the learning rate, which are 0.077 and 1, respectively.
The efficiency of SVM is monitored by parameter C and parameter γ. Kernel type parameter specifies the type of kernel function. Different types of kernels such as dot, radial, polynomial, neural, epachnenikov, gaussian_combination and multiquadric can be selected. Neural and gaussian_combination types with set of examples were unsuccessful in assessments and the conclusion time was more than the expected time. In the neural kernel, the process was not resulted in a final answer after 25 minutes yet. Two kernel types of dot and polynomial had an output for a hazardous state. The rest of the cases did not have any prediction or record return for a hazardous state.
Polynomial kernels are very good for the problems which all training data have been normalized. The constant C defines the complexity of SVM. It is important to determine the correct constant. If the constant is too large, it can lead to a lot of inappropriate connection points, while very small amounts may lead to spread the points more. Max iteration parameter is an optimization parameter which does the stopping procedure after a specified number of assessments.
So, we can see how small changes have a significant impact on results. Therefore, having a good understanding of used kernel parameters is essential. As well as having a good understanding of the different types of kernel, selecting the most appropriate type of kernel for sample set is also equally important.
In this paper, a polynomial kernel function was used instead of some other kernel functions such as Gaussian kernel for SVM algorithm, because Gaussian kernel function had almost low accuracy. Optimized algorithms such as PSO can be used to accelerate the performance of SVM algorithm and enhance its accuracy. The PSO algorithm used to optimize SVM parameters such as C, penalty factor and g or γ, kernel function parameter.
In Naive Bayes, the results obtained classes listed in the situation "hazardous" and "nonhazardous" have been reported. Decision tree consists of 95 nodes with 96 terminal leaves.
The benefits of particle swarm optimization algorithm include memory utilization, sharing information, better flexibility against the local optimum problem, high-speed convergence and easy implementation. Block diagrams of SVM-PSO algorithm [30] is depicted in Fig. 2. In this study, we used 5-fold cross-validation method.
Fig. 2. Block diagram of SVM-PSO algorithm
3.4. Performance Evaluation Parameters
Several parameters have been used to evaluate the performance of different algorithms in the expert system that are as follows [31].
1. True positive (TP). The number of times that algorithm predicted hazardous seismic correctly.
2. True negative (TN). The number of times that algorithm predicted non-hazardous seismic correctly.
3. False positive (FP). The number of times that algorithm predicted hazardous seismic wrongly.
4. False negative (FN). The number of times that algorithm predicted non-hazardous seismic wrongly.
The overall accuracy is obtained using Eq. (1).
Accuracy = (TP + TN) / (TP + FP + TN + FN) (1)
3.5. Simulation Results
As already mentioned, 5-fold cross-validation has been used in the training and testing phases in data mining algorithms. In other words, whole dataset, 2584 has been divided into five sets. In each step, four sets have been used for training and one set has been applied for testing. The accuracy of each step was computed and final accuracy was obtained using averaging of all accuracy as can be seen in Table 4. This causes all datasets have participated in training and testing phases. In addition to the accuracy, the computational time of each algorithm was measured. Although the SVM algorithm is the fastest one, it has the smallest accuracy.
While SVM-PSO algorithm has the highest accuracy and its computational time is smaller than MLP and NN algorithms. In order to clearly show the accuracy results, we show the overall accuracy of all tested data mining algorithms in Fig. 3. As this figure depicts the SVM algorithm which has been optimized using PSO has the highest accuracy.
Table 4. Performance of different data mining algorithms on Rapid Miner platform with 5- fold cross-validation method.
Training and testing data Parameter SVM Naive Bayes
Decision Tree
Auto MLP
Neural Net
SVM- PSO
FP 0 93 782 196 345 0
(Training data=517-2584) Accuracy (%) 84.11 83.00 74.92 81.59 79.52 84.11 (Testing data= 1-516) Computational
time (second)
10 11 42 690 3550 137
FP 0 764 471 90 336 0
(Training data=1-516, 1034-2584) Accuracy (%) 93.04 79.73 84.80 91.32 87.16 93.04 (Testing data=517-1033) Computational
time (second)
9 12 43 649 3513 134
FP 478 70 215 0 6 0
(Training data=1-1033, 1551- 2584)
Accuracy (%) 89.01 96.72 94.13 98.06 97.95 98.06
(Testing data=1034-1550) Computational time (second)
10 12 45 631 3659 135
FP 491 325 156 9 100 0
(Training data=1-1550, 2068- 2584)
Accuracy (%) 86.32 89.72 92.46 95.18 93.50 95.36
(Testing data=1551-2067) Computational time (second)
8 13 45 683 3814 135
FP 1492 199 123 0 12 0
(Training data=1- 2067) Accuracy (%) 68.70 93.09 94.22 96.52 96.32 96.52 (Testing data=2068-2584) Computational
time (second)
9 13 44 736 35.70 136
Average accuracy 84.236 88.452 88.106 92.534 90.89 93.418
Fig. 3. Comparison of accuracy of different data mining algorithms.
78 80 82 84 86 88 90 92 94 96
SVM Decision Tree
Naive Bayes
Neural Net
Auto MLP SVM-PSO Different data mining algorithms
Accuracy (%)
4. DISCUSSION
For this evaluation 2584 data from eighteen features has been used. All datasets have been participated in training and testing phases using 5-fold cross validation. In all steps, the SVM-PSO algorithm was almost the fastest algorithm. In other words, different data which have been used in each step of 5-fold cross validation do not change the algorithm accuracy. In order to evaluate the impact of different features on the performance, we have divided eighteen features into three categories. Each categories has six features which was depicted in Table 3. The obtained results for each category using 5-fold cross validation using whole dataset is depicted in Table 5. As can be seen in the table, each algorithm using each feature category has different accuracy. In other words, each feature category has different impact on the accuracy of the algorithms. For example, the SVM, decision tree, MLP, and neural network algorithms have the highest accuracy for third category compared to the other categories, while the Naive Bayes and SVM-PSO algorithms have the highest accuracy for the first category. The combination of two first categories, twelve features does not improve the performance of algorithms except for SVM algorithm. These results show that some features are more important than others in terms of impact on the accuracy.
Comparison of results of this paper and other related work has been presented in Table 6. The results show the ability of SVM-PSO algorithms to predict earthquakes using seismic hazard data in Poland coal mines is higher than other algorithms such as rule-based regression algorithm and rule-based classification algorithm in [23]. As can be seen the accuracy of the SVM-PSO algorithm is higher than [23].
Table 5. The priority results of each category feature simulated on Rapid Miner platform with 5-fold cross-validation method
CategoryAttributeSVMNaive Bayes Decision TreeMLPNeural NetSVM- PSO 1
Seismic Seismoacoustic Shift Genergy Gpuls Gdenergy
53.9 91.185.56 92.292.193.4 2
Gdpuls Ghazard Nbumps Nbumps2 Nbumps3 Nbumps4
73.0
80.3 88.7
87.7 90.4
88.27 93.2
92.7 92.7
91.6 80.5
93.41 3
Nbumps5 Nbumps6 Nbumps7 Nbumps89 Energy Maxenergy
90.8
77.3 90.7
88.2 91.5
90.2 93.4
93.2 93.4
92 60.4
60.3 Overall Accuracy84.288.488.192.590.893.4
Table 6. Comparing the algorithms mentioned in this paper with the other ones
5. Conclusions
Humans face with various natural crises such as earthquake, flood and volcano and unnatural crises such as war, road accident and refugees which impose irreparable damages to him. If we can predict the probability of crises, damages will certainly reduce. The researchers have used various techniques such as the study of animal behavior changes, ionosphere analysis and mathematical modeling for predicting of earthquakes.
Ionosphere analysis usually includes fewer properties, such as changes in water ions and, thus, it has low prediction accuracy. Expert systems which are knowledge-based, inference engine and user interface have recently been proposed for this goal. Various algorithms neural networks, decision tree, Naive Bayes and support vector machine were used in expert systems. These algorithms were evaluated in terms of accuracy and computational time. Each earthquake has some unique properties depending on the type of geology characteristics. The simulation results using real data in Rapid Miner platform show that support vector machine algorithm is the fastest and also has the lowest accuracy, while the multilayer perceptron network has the highest accuracy. Some reasons for the behavior of support vector machine are as follows. First, the number of false positives in support vector machine was more than other algorithms. Second, a kernel function of polynomial was used. In order to increase the accuracy of support vector machine algorithm, this algorithm was optimized using particle swarm optimization algorithm. The benefits of particle swarm optimization algorithm include memory utilization, sharing information, and high-speed convergence. Support vector machine optimized using particle swarm optimization algorithm has the highest accuracy compared to other algorithms.
REFERENCES
1. Monadi, A., B. Eslampour, N. Baghaeimehr and F. Abed, 2012. Crisis Management in Cities from Holey Quran Perspective. J. Appl. Environ. Biol. Sci., 2 (12): 606-608.
2. Solymani, A., A. Abdolahii and M. Rahimi, 2014. Evaluation and Improvement of Urban Worn against Earthquakes (Case Study Neighborhood of Shiraz Zvalanvar). J. Appl. Environ. Biol. Sci., 4 (1): 157- 165.
3. Abraham, A., 2005. Rule-based expert systems, Handbook of measuring system design, pp: 909-919.
4. Hadjimichael, M., A. P. Kuciauskas, P.M. Tag, R.L. Bankert and J.E. Peak, 2002. A meteorological fuzzy expert system incorporating subjective user input, Knowl Inf Syst, 4 (3): 350-369.
5. Mansiya, K., Z. Alma, M. Torgyn, M. Marzhan and N., Kanat, 2014. The methodology of expert systems, International Journal of Computer Science and Network Security, 14 (2): 62-66.
6. Meyer, D. and F.T. Wien, 2012. Support Vector Machines, The Interface to libsvm in package e1071, pp: 1-8.
7. Vapnik, V., 1995. The nature of statistical learning theory, Springer Verlag.
8. Suresh, G. and G. Zayaraz, 2015. Concept relation extraction using Naive Bayes Classifier for ontology-based question answering systems. Journal of King Sand University- Computer and Information Sciences, 27 (1): 13-24.
9. Giarratano, J. C. and G. D. Riley, 2004. Expert system: principles & programming, 4rd edn, PWS publishing company.
10. Ellahi, A. and W. Shahzad, 2016. Exploring and Analyzing Evolutionary Optimization in Different Environments. J. Appl. Environ. Biol. Sci., 6 (8): 98-111.
11. Ahumada, A., A. Altunkaynak and A. Ayoub, 2015. Fuzzy logic-based attenuation relationships of strong motion earthquake records. Expert Systems with Applications, 42 (3): 1287-1297.
12. Siler, W. and J.J. Buckley, 2005. Fuzzy expert system and fuzzy reasoning, John Wiley & Sons, Inc.
References Country Used data The probability of
occurring an earthquake using seismic hazard data
(percentage)
Simulation platform
SVM-PSO algorithm (this paper)
Polish In Polish coal mines (contains 2584 records)
93.42 Rapid Miner
[23] Polish In Polish coal mines (task 1 for the longwall Sc503 contains 1097records)
87.0 Weka
13. Ikram, A. and U. Qamar, 2014. A rule-based expert system for earthquake prediction, J Intell Inf Syst, 43 (2): 205-230.
14. Ikram, A. and U. Qamar, 2015. Developing and expert system based on association rules and predicate logic for earthquake prediction, Knowledge-Based Systems, 75, 87-103.
15. Abrahamson, N. and W. Silva, 2008. Summary of the Abrahamson and Silva NGA ground-motion relations. Earthquake Spectra, 24 (1): 67-97.
16. Basak, D. and S. Pal, 2007. Patranabis DC Support Vector Regression. Neural Inf. Process, 11 (10):
203-225.
17. Tezcan, J. and Q. Cheng, 2012. Support vector regression for estimating earthquake response spectra, Bulletin of Earthquake Engineering, 10 (4): 1205-1219.
18. Moustra, M., M. Avraamides and C., Christodoulou, 2011. Artificial neural networks for earthquake prediction using time series magnitude data or Seismic Electric Signals, Expert Systems with Applications, 38 (12): 15032-15039.
19. Dehbozorgi, L. and F. Farokhi, 2010. Effective feature selection for short-term earthquake prediction using neuro-fuzzy classifier. Second IITA International Conference on Geoscience and Remote Sensing, 2, 165-169.
20. Sikder, I.U. and T. Munakata, 2009. Application of rough set and decision tree for characterization of premonitory factors of low seismic activity. Expert Systems with Applications, 36 (1): 102-110.
21. Aminzadeh, F., S. Katz and K. Aki, 1994. Adaptive neural nets for generation of artificial earthquake precursors. IEEE Transactions on Geoscience and Remote Sensing, 32 (6): 1139-1143.
22. Http://rapid-i.com/ Accessed August 2015.
23. Kabiesz, J., B. Sikora, M. Sikora, and L. Wrobell, 2013. Application of rule-based models for seismic hazard prediction in coal mines, Acta Montanistica Slovaca, 18 (4): 262-277.
24. UCI, 2015. Machine learning Respository: Https://archive.ics.uci.edu/ml/datasets/seismicbumps.
25. Kawabe, I., I. Ohno and S. Nadano, 1988. Groundwater flow records indicating earthquake occurrence and induced Earth's free oscillation, Geophysical Research Letters, 15 (11): 1235-1238.
26. King, C.Y., S. Azuma, M. Ohno, Y. Asai, P. He, Y. Kitagawa, G. Igarashi and H. Wakita, 2000. In search of earthquake precursors in the water-level data of 16 closely clustered wells at Tono, Japan, Geophysical Journal International, 143 (2): 469-477.
27. Tertyshnikov, A.V., V.O. Skripachev and G.V. Chemyavskii, 2009. Variations in deceleration of space vehicles in the upper ionosphere before strong earthquake, Doklady Earth Sciences, 424 (1): 180- 184.
28. Lakshmi, K.R., Y. Nagesh and M.V. Krishna, 2014. Analysis on predicting earthquakes through an abnormal behavior of animals, International Journal of Scientific & Engineering Research, 5 (4): 845- 857.
29. Zoback, M.L., E. Geist, J. Pallister, D.P. Hill, S. Young and W. McCausland, 2013. Advances in natural hazard science and assessment, 1963-2013, The Impact of the Geological Sciences on Society:
Geological Society of America Special Paper 501, pp: 81-154.
30. Huang, C. and J. Dun, 2008. A distributed PSO-SVM hybrid system with feature selection and parameter optimization. Applied soft computing, 8 (4): 1381-1391.
31. Reyes, J., A.M. Esteban and F.M. Alvarez, 2013. Neural networks to predict earthquakes in Chile, Applied Soft Computing, 13 (2): 1314-1328.
Appendix A
