Accepted Manuscript

A Swarm-Optimized Fuzzy Instance-Based Learning approach for predicting

slope collapses in mountain roads

Min-Yuan Cheng, Nhat-Duc Hoang

PII: S0950-7051(14)00460-2
DOI: http://dx.doi.org/10.1016/j.knosys.2014.12.022
Reference: KNOSYS 3031

To appear in: Knowledge-Based Systems

Received Date: 11 August 2014

Revised Date: 22 December 2014
Accepted Date: 23 December 2014

Please cite this article as: M-Y. Cheng, N-D. Hoang, A Swarm-Optimized Fuzzy Instance-Based Learning approach
for predicting slope collapses in mountain roads, Knowledge-Based Systems (2014), doi: http://dx.doi.org/10.1016/
j.knosys.2014.12.022

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 TITLE PAGE
A Swarm-Optimized Fuzzy Instance-Based Learning Approach for Predicting Slope Collapses in

Mountain Roads

Reference: KNOSYS 3031

Editorial reference: KNOSYS_KNOSYS-D-14-00997

Author: Min-Yuan, Cheng, Ph.D.

Position: Professor

Affiliation: Department of Civil and Construction Engineering, National Taiwan University of Science and
Technology

Address: #43, Section 4, Keelung Road, Daan District, Taipei, Taiwan

Postal Code: 8862

Phone: +886 2 27336596 / +886 2 27301074

Email: myc@mail.ntust.edu.tw

Author: Nhat-Duc, Hoang, Ph.D. *

Position: Lecturer

Affiliation: Institute of Research and Development, Faculty of Civil Engineering, Duy Tan University

Address: K7/25 Quang Trung, Danang, Vietnam

Postal Code: 550000

Phone: +84 0511.3827111

Email: hoangnhatduc@dtu.edu.vn

*Corresponding author

A Swarm-Optimized Fuzzy Instance-Based Learning Approach for

Predicting Slope Collapses in Mountain Roads

Abstract. Due to the disastrous consequences of slope failures, forecasting their occurrences is a

practical need of government agencies to develop strategic disaster prevention programs. This

research proposes a Swarm-Optimized Fuzzy Instance-based Learning (SOFIL) model for

predicting slope collapses. The proposed model utilizes the Fuzzy k-Nearest Neighbor (FKNN)

algorithm as an instance-based learning method to predict slope collapse events. Meanwhile, to

determine the model’s hyper-parameters appropriately, the Firefly Algorithm (FA) is employed

as an optimization technique. Experimental results have pointed out that the newly established

SOFIL can outperform other benchmarking algorithms. Therefore, the proposed model is very

promising to help decision-makers in coping with the slope collapse prediction problem.

Keywords: Instance-Based Learning; Swarm Intelligence; Fuzzy K-Nearest Neighbor; Slope

Collapse Prediction; Firefly Algorithm

1. Background

Road is the major transportation construction that support modern society; many researchers

have found significant positive correlations between roads and economic growth at both local

and regional levels [1, 2]. Accordingly, in various countries around the world, extensive

networks of mountain roads has recently been built to catch up with the population expansion

and the economic development [3, 4].

 Furthermore, natural hazards coupled with rugged terrains lead to the fact that slope collapses

possibly occur in many sections of the road network. These catastrophic events are often

triggered by earthquakes or heavy rainfalls during typhoons or monsoon storms [5, 6]. Slope

collapses are very undesirable since they inflict damages to man-made structures, disruption of

traffic, and indispensible losses of human lives. Hence, slope stability assessment is an inevitable

task which should be regularly conducted by roadway maintenance authorities [7, 8]. The

analysis results can be utilized for identifying collapse-prone areas as well as allocating scarce

resources to establish an overall disaster prevention program [9]. In order to analyze slope

stability, physical model, expert evaluation, and machine learning are the three common methods

[10].

The physical model method is based on the slope displacement model which can analyze the

slope stability by identifying of the most dangerous sliding surface and calculating the factor of

safety [11]. Although this approach can deliver accurate analytical results, it requires input

parameters for every calculation point of the investigated area. Therefore, the physical model

method is only appropriate for evaluating stability in small areas, and its capacity for analysis

over large areas is inapplicable [12].

The expert evaluation approach utilizes expert judgments and information of slope collapse

events occurred in the past [8, 13]. Using expert knowledge, the main influencing features and

possible triggering factors can be identified [14]. Based on that information, the stability of a

slope can be evaluated by expert knowledge. Obviously, requiring many subjective judgments

and inconsistency of the prediction results are the main drawbacks of this method.

Recently, machine learning approaches have been utilized to automate the slope assessment

process due to their better flexibilities and prediction capabilities compared to the traditional
approaches. Generally, machine learning based models are established by combining artificial

intelligence (AI) techniques and historical databases [15]. Using these models, the slope

evaluation can be considered as a classification task in which prediction outputs are either “stable”

or “unstable”.

Lu and Rosenbaum [16], Zhou and Chen [17], Jiang [18], Das et al. [7], Cho [19], Lee et al.

[20], and Wang et al. [21] applied the Artificial Neural Network (ANN) to predict the slope

condition. Zhao et al. [22] employed the Relevance Vector Machine (RVM) to explore the

nonlinear relationship between slope stability and its influence factors. Slope stability forecasting

models based on the Support Vector Machine (SVM) were developed by Li and Wang [23],

Cheng et al. [24], Zhao [25], Samui [26], and Li and Dong [27]; these studies found that SVM

based models are very effective under the condition of limited data.

Although the ANN has been extensively applied for predicting slope collapse, the

implementation of this approach has several drawbacks. The major disadvantage of the ANN is

that its training process is achieved through a gradient descent algorithm on the error space,

which can be very complex and may contain many local minima [28]. Moreover, the SVM

training requires solving a quadratic programming problem subjected to inequality constraint;

this means that the training process for large data sets requires expensive computational cost [29].

Most importantly, the black box nature of the ANN, SVM, and RVM algorithms makes them

difficult for practical engineers or government agencies to comprehend how they predict slope

collapses.

Different from the aforementioned AI methods, the Fuzzy k-Nearest Neighbor (FKNN)

algorithm [30] belongs to the class of instance-based learning. This algorithm utilizes the whole

collected data to establish its memory. A FKNN classifier utilizes the information obtained from
the k nearest neighbors of a sample vector and assigns class memberships to it. The vector’s

membership values provide a level of assurance to accompany the resultant classification.

Moreover, the algorithm also assigns fuzzy memberships as a function of the vector’s distance

from its k nearest neighbors and those neighbors’ memberships in the possible classes [31].

Needless to say, this approach is simple to implement and its classification outcomes are also

easily interpretable. In addition, the competitive prediction performance of the FKNN has been

demonstrated in various studies [30, 32-34]. Nevertheless, none of previous works has evaluated

the capability of the FKNN method in slope collapse assessment.

Additionally, the implementation of the FKNN requires a proper setting of two tuning

parameters: the neighboring size (k) and the fuzzy strength (m). Furthermore, this parameter

selection process can be modeled as an optimization problem. Meta-heuristic approaches have

been illustrated to be feasible to tackle the optimization problem at hand [35-39]. Recently

developed by Yang [40], the Firefly Algorithm (FA) is a fast and effective meta-heuristic for

solving global optimization in continuous space. Numerical experiments in previous researches

have demonstrated the superior performance of the FA over other meta-heuristic methods [41-

43]. Nonetheless, few research works have investigated the capability of this algorithm in

optimizing the parameter selection process of the FKNN. Thus, this study proposes to hybridize

the FKNN with FA [40] to automatically search for appropriate hyper-parameters of the

prediction model.

Thus, this research employs the FKNN classifier as the machine learning technique to

construct a prediction model for slope collapse assessment. We propose to hybridize the FKNN

algorithm with the FA [40] to automatically search for an appropriate combination of tuning

parameters for the prediction model. The newly established approach is named as Swarm-
Optimized Fuzzy Instance-based Learning (SOFIL). The remaining part of this paper is

organized as follows. The second section of this paper presents the research methodology. The

framework of the proposed SOFIL is described in the third section. The fourth section

demonstrates the experimental results. Conclusions of the study are stated in the final section.

2. Methodology

2.1 Fuzzy k-Nearest Neighbor algorithm

The FKNN algorithm is an instance-based classifier that incorporates the fuzzy set theory

into the classification process [30]. In the FKNN, the fuzzy memberships of samples are assigned

to different classes. The class which possesses the maximum membership degree can be chosen

as the winner. The first step of the FKNN algorithm is to calculate the fuzzy partition matrix U =

[uij] from the memory which stores a set of n training sample vectors [x1,…,xn]. Herein, we

denote j as the vector index (j = 1, 2, …,n), where n is the number of training samples. And, the

variable i represents the class index (i = 1, 2, …,C) , where C is the number of classes. For each

training case x, we identify its k nearest neighbors by calculating Euclidean distances. The

membership degree of the sample vector xj in the class i is given as follows:

⎧⎪0.51 + (ni / k ) × 0.49, if c( x j ) = i

uij ( x) = ui ( x j ) = ⎨ (1)
⎪⎩(ni / k ) × 0.49, if c( x j ) ≠ i

where ni is the number of neighbors found which belong to the class i and c(xj) represents the

class label of the sample vector xj. It is obvious that uij is an element of the C-by-n matrix U.

Moreover, it is also worth noticing that the purpose of Eq. (1) is to assign higher fuzzy

membership grades to the training samples that stay away from the decision boundary and lower

fuzzy memberships grade to the patterns that lie in the vicinity of the decision boundary [30]. It
is because the information supplied by the samples in the region close to the decision surface is

more uncertain than that provided by other samples.

Since uij is a fuzzy membership grade of the sample xj in the class i, uij must satisfy the

following properties:

uij ∈[0,1] (2)

∑u
i =1
ij =1 (3)

n
0 < ∑ui j < n (4)
j =1

The second step of the FKNN approach is to assign fuzzy memberships of the unknown

sample x to different classes according to the following equation:

∑u j =1
ij (1 / || x − x j ||2 /( m−1) )
ui ( x ) = k (5)
∑ (1/ || x − x
j =1
j || 2 /( m −1)
)

where i = 1,2,…,C, and j = 1,2,..,k. j represents the jth sample vector among the k nearest

neighbors of x. C is the number of classes; k denotes the neighboring size. The fuzzy strength m

is used to determine how heavily the distance is weighted when computing each neighbor’s

contribution to the membership value. || x − x j || represents the distance between x and its jth

nearest neighbor xj. In this study, Euclidean metric is used as the distance measurement. uij,

denotes the membership degree of the sample vector xj in the class i and is computed in the first

step of the algorithm (refer to Eq. (1)).

2.2 Firefly Algorithm (FA)

 In order to commence the training process of the FKNN, two tuning parameters (k, m) are

required to be determined. A proper setting of these tuning parameters is necessary to achieve a

desirable performance of the prediction model [29]. Thus, in this study, we utilize the FA as a

means for tuning the FKNN parameters. The description of the FA algorithm is provided in the

following section of the article.

The FA is a population-based swarm intelligence which simulates the flashing and

communication behavior of fireflies [44]. In the natural world, a firefly is attracted to brighter

ones as it randomly explores the habitat. Based on that phenomenon in nature, the FA is

formulated as a global optimization method in which the brightness of fireflies characterizes the

value of the objective function. Previous studies have demonstrated that this advanced swarm

intelligence is fast and effective for locating the global optimum and superior performance of the

FA over other meta-heuristic algorithms has been proved in various applications [39, 41, 42].

The FA utilizes the following rules: (1) all fireflies are unisex, so each firefly is attracted to

other fireflies regardless of their sex, (2) the attractiveness of a firefly is proportional to its

brightness and decreases as the distance increases. A firefly moves randomly if no other firefly is

brighter, and (3) the brightness of a firefly is affected or determined by the landscape of the

objective function. The FA pseudo code is illustrated in Fig. 1.

<Insert Fig. 1 here>

The brightness of an individual firefly can be defined similarly to the fitness value in the

genetic algorithm. The light intensity I(r) varies according to the following equation:

I (r ) = I o exp(−γr 2 ) (6)

where Io denotes the light intensity of the source. γ is the light absorption coefficient. r represents

the distance from the source.

 As the attractiveness of a firefly is proportional to the light intensity seen by adjacent fireflies,

the attractiveness β of a firefly is defined as:

β = βo exp(−γr 2 ) (7)

In a D-dimensional space, the distance between any two fireflies i at xi and j at xj, is the

calculated as follows:

D
rij = xi − x j = ∑ (x
k =1
i ,k − x j ,k ) 2 (8)

Since a specific firefly xi is attracted to the brighter one xj, the movement of the ith firefly can

be expressed as:

xi = xi + βo exp(−γrij2 )(xi − x j ) + α (ω − 0.5) (9)

where γ is the light absorption coefficient, γ varies from 0.1 to 10; β0 represents the attractiveness

at rij = 0; α denotes a trade-off constant to determine the random behavior of movement; ω

represents a random number drawn from the Gaussian distribution.

2.3 Historical Records

The historical data utilized in this research contains 211 slope evaluation cases collected in

the Taiwan Provincial Highway No. 18 and No. 21 during the typhoons Herb (1996), Nari (2001),

and Toraji (2013). In this database, there are 105 failure and 106 non-failure cases. In this

research, a slope condition, either failure or non-failure, is determined and recorded during field

surveys. Specifically, a non-failure slope is determined when there is no movement of the soil in

the slope surface that affects the safety of road traffic. Moreover, a case of slope is characterized

by slope attributes and slope observations (failure/non-failure).

For the purpose of slope collapse prediction, this study employs 16 slope attributes divided

into 9 groups: landforms, geological structure, stratigraphy, rock properties, vegetation coverage,
water condition, road properties, earthquake, and rainfall. These attributes can be considered as

influencing factors that determine slope conditions and they are selected based on engineering

judgments, available statistical data, and findings from previous researches [4, 9, 45-47] .

<Insert Table 1 here>

<Insert Table 2 here>

Table 1 provides the information of the influencing factors and their statistical descriptions.

Illustration of the database is shown in Table 2 where the output of 1 indicates a failed slope and

the output of -1 represents a stable slope. In Table 1, the first group covers four factors: slope

aspect, slope gradient, slope height, and slope form. The slope aspect refers to the horizontal

direction to which a mountain slope faces; and it has an indirect impact on moisture content of

the soil, which is related to the reduction of the effective stresses at the potential failure surface.

The slope angle measures the steepness inclination of the slope. The slope height, defined as the

distance from crest to toe of a slope, is physically related to the magnitude of the stress and the

pore-water pressure in the lower slope [48]. The slope form describes the geometry of slope

surface which influences soil movement, rill patterns, and run-off production [49].

The second group depicts the geological characteristic of the area along the mountain roads.

The stratigraphic feature of the region is described in the third group; it includes two impact

factors, namely the angle between slope aspect and trend and the angle between gradient and

inclination. In addition, the fourth group of factors provides information of rock properties in

which the rock mass size and the rock mass volume are taken into account. Moreover, the

characteristic of vegetation on slope surface is also critical when assessing the slope stability; to
quantify this characteristic, the vegetation coverage percentage and the vegetation coverage

thickness are considered [46, 50].

Furthermore, in this research, the water condition of a slope is reflected by the size of

catchment area which is computed by a digital terrain model (DTM) [4]. Additionally, since road

construction is the artificial factor that can cause slope failure, our study considers two features

of mountain roads: excavation height at slope toe and change of slope gradient due to toe cutting.

On the other hand, earthquake and typhoon are generally considered as the two natural hazards

that trigger slope collapse events in many places (e.g. Taiwan). In our study, the maximum

ground acceleration at the slope location during earthquake and the maximum accumulated

rainfall during typhoon are taken into account to measure their effects on slope stability.

3. Swarm-Optimized Fuzzy Instance-based Learning (SOFIL) for

Predicting Slope Collapses in Mountain Roads

This section of the article describes the proposed slope collapse prediction method, named as

SOFIL, in detail. The model (see Fig. 2) is established by a fusion of the FKNN algorithm and

the FA optimization algorithm.

<Insert Fig. 2 here>

(1) Input Data: The input data provides the attributes of a slope. As mentioned earlier, the slope

attributes consist of influencing factors that impose significant impacts on the slope collapse

events. The data can be real values or integers and they should be normalized into a range of (0,

1). This transformation can help avoid numerical difficulties and prevent the situation in which

attributes with greater numeric magnitudes dominate those with smaller magnitudes.
(2) Tuning Parameter Initialization: The aforementioned tuning parameters of the model are

randomly generated within the range of lower and upper boundaries. In this study, the lower and

upper boundaries of the neighboring size (k) are 1 and 30, respectively. Meanwhile, these two

values of the fuzzy strength (m) are 1.0001 and 10. Moreover, the equation used for generating

the model tuning parameters can be shown as follows:

X i , 0 = LB + rand [ 0,1] × (UB − LB ) (10)

where X i , 0 is the tuning parameter i at the first generation. rand[0,1] denotes a uniformly

distributed random number between 0 and 1. LB and UB are two vectors of lower bound and

upper bound for any parameter.

(3) Class Membership Assignment: In this step, the FKNN algorithm is deployed to assign fuzzy

memberships of an input vector to different classes. This step requires two parameters (the

neighboring size and the fuzzy strength) that are acquired from the FA component. It is noted

that the slope assessment problem is a two-class classification problem with two labels: “collapse”

and “non-collapse”. Thus, for each input pattern x, there are two outputs, u1(x) and u2(x),

representing membership degrees of x in the two classes.

(4) Firefly Algorithm Searching: the FA optimization technique is applied to automatically

explore the various combinations of the tuning parameters (k and m). At each generation, the

optimizer carries out its searching process to guide the population of fireflies to the optimal

solution. By evaluating the fitness of each firefly, the algorithm discards inferior combinations of

m and k, and permits robust combinations of these parameters to be passed on the next

generations.
(5) Output Defuzzification: Because the FKNN yields fuzzy memberships of an input pattern in

the two classes (u1(x) and u2(x)), a step of defuzzification is employed to convert fuzzy outputs to

crisp outputs (Y(x)) as follows:

2
Y ( x) = arg max(ui ( x)) (11)
i =1

(6) Fitness Evaluation: In this step, the training data set is divided into five mutually exclusive

subsets. In each run, one subset is used as a validating set; meanwhile, the other subsets are used

for constructing the model memory. In order to determine the optimal tuning parameters of the

FKNN, the following objective function is used:

1
Ffitness = 5
(12)
∑ AR
k =1
k

where ARk denotes the classification accuracy of the validating set at the kth run. The

classification accuracy is calculated as the number of correct classifications divided by the

number of all data instances within a data set.

(7) Stopping Condition: The optimization process of the FA algorithm terminates when the

maximum number of generation is achieved. If the stopping condition is not met, the FA will

continue its searching progress.

(8) Optimal Prediction Model: When the program terminates, the optimal set of tuning

parameters has been successfully identified. The SOFIL is ready to predict new input patterns.

4. Experimental Results

To demonstrate the capability of the proposed SOFIL, its performance is compared to

results acquired from other benchmark approaches including the ANN, FKNN, RVM, and

SVM algorithms. As mentioned earlier, in the SOFIL, the neighboring size (k) and the fuzzy
strength (m) are automatically chosen by the FA optimization. In FKNN algorithms, the

neighboring size k is allowed to vary between 1 and 30; additionally, this parameter is also

selected via a five-fold cross validation process based on the training cases. The fuzzy strength

parameter (m) in the FKNN algorithm is set to be 2, as recommended by the previous work

[30]. Moreover, the parameters of the SVM is determined via the grid search approach [51, 52].

When using an ANN, it is needed to specify the number of hidden layers, the number of

neurons in the hidden layer, the learning rate, and the number of training epochs [53]. These

parameters of an ANN are generally selected via repetitive trial-and-error processes. The

network configuration is described as follows: the number of hidden layers is set to be 1; the

number of neurons in the hidden layer is 16; and the number of training epochs is selected to be

2000. The back-propagation approach is used as the method for training the ANN model [54].

In the experiment, the whole database is randomly divided into two set: set 1 (including 80%

of the cases) used to construct the prediction model, and set 2 (including 20% of the cases)

utilized for testing the model. To evaluate model performance, the classification accuracy rate

can be employed. The classification accuracy rate (CAR) is the ratio of correctly predicted cases

over the total number of cases, can be used to measure the classifier performance [55, 56].

Moreover, the predictive capability of the classifiers can also be assessed using the following

four metrics [56]: true positive rate (the percentage of positive instances correctly classified),

true negative rate (the percentage of negative instances correctly classified), false positive rate

(the percentage of negative instances misclassified), and false negative rate (the percentage of

positive instances misclassified). These four metrics can be summarized in a confusion matrix

[57]. A well-known approach to incorporate these four measures and to produce an evaluation
criterion is to employ the Receiver Operating Characteristic (ROC) curve [58]. Furthermore, the

area under the ROC curve, denoted as AUC, provides a single measure of a classifier’s

performance for evaluating which model is better on average [59, 60].

It is noted that a higher AUC value indicates a better predictive performance. Generally, a

classifier with perfect predictive ability has an AUC of 1; meanwhile, a poor classifier with

random predictions has an AUC of 0.5. Moreover, an AUC of the range (0.7, 0.8) indicates an

acceptable classification performance. If 0.8 ≤ AUC ≤ 0.9, an excellent classification

performance is attained. And, if AUC ≥ 0.9, the classifier has attained an outstanding

performance.

<Insert Fig. 3 here>

When the FA-based parameter tuning process terminates, the optimized hyper-parameters of

the SOFIL has been identified as: k = 5, m = 1.28. Furthermore, the evolutionary process of the

SOFIL can be observed in Fig. 3. It can be seen that the value of the fitness function (shown in

Eq. 12) gradually improved and it reached the best value at iteration 102. During the latter part of

the searching process, the fitness value remains the same until the stopping condition (the

maximum number of generation = 300) is satisfied.

The experimental result has demonstrated that the FA is a very effective meta-heuristic since

it can help the proposed model to converge quickly toward the most desirable set of hyper-

parameters. Detailed of the SOFIL’s prediction results for the testing data is shown in Table 3.

Herein, µ 1(X) and µ 2(X) represent the membership degrees of the input vector X in the two

classes (collapse and non-collapse).

 <Insert Table 3 here>

<Insert Table 4 here>

The confusion matrices of the SOFIL and other methods are described in Table 4. In the

training process, the numbers of common false positives and false negatives of the five models

are 1 and 3, respectively. Meanwhile, in the testing process, the proposed approach and the SVM

do not commit any false positive. The RVM, ANN, and FKNN algorithms have 1 overlapped

false positive. In addition, the number of overlapped false negatives of the five methods is 2. It

can be observed that the proposed SOFIL has achieved the lowest false positives and false

negatives in both the training and testing processes.

<Insert Table 5 here>

Table 5 provides the result obtained from the training and testing processes of the SOFIL and

other benchmark methods. The CAR results of the SOFIL, SVM, RVM, ANN, and FKNN

methods are 95.24%, 92.85%, 88.01%, 88.10%, and 85.71% in the testing process, respectively.

When predicting testing samples, the AUC values of the five methods are 0.95, 0.94, 0.89, 0.88

and 0.86, respectively. Observably, the proposed SOFIL can deliver the best result of slope

collapse prediction in both training and testing processes. Notably, the newly established method

can properly classify 40 testing cases with only 2 misclassifications. Thus, the SOFIL deems best

suited for the slope collapse prediction problem at hand.

5. Conclusion

In this research, a new slope collapse prediction model, named as SOFIL, has been proposed.

Experimental results obtained from both training and testing processes have verified that the new

model can outperform other benchmark methods in terms of all performance measurements. This
demonstrates that the SOFIL is a very promising alternative to support decision makers in slope

collapse assessment.

The newly built approach is established by a hybridization of the FKNN - an instance-based

learning classifier and the FA - a swarm intelligence optimization technique. The SOFIL utilizes

the FKNN algorithm to assign a membership grade in each class to an unknown pattern of slope

attributes. Additionally, the FA searching algorithm is deployed to identify the most appropriate

set of the FKNN’s tuning parameters. As a result, the proposed method can eliminate the need of

human effort or domain knowledge for parameter setting. Since the SOFIL is an effective

classifier, it can be applied for solving other problems in the field of civil engineering. Moreover,

investigating a mechanism of adaptive weightings for calculating distances can be a promising

future enhancement of the current prediction model.

References

[1] J. Bryan, S. Hill, M. Munday, A. Roberts, Road infrastructure and economic development in the periphery: the
case of A55 improvements in North Wales, J. Transp. Geogr., 5 (1997) 227-237.
[2] A.H. Munnel, Policy Watch - Infrastructure Investment and Economic Growth, J. Econ. Perspect., 6 (1992) 189–
198.
[3] S. Yang, C. Shen, C. Huang, C. Lee, C. Cheng, C. Chen, Prediction of Mountain Road Closure Due to Rainfall-
Induced Landslides, J. Perform. Constr. Fac., 26 (2012) 197-202.
[4] J. Ching, H.-J. Liao, J.-Y. Lee, Predicting rainfall-induced landslide potential along a mountain road in Taiwan,
Geotechnique 61, No. 2, 153–166, (2011).
[5] H.-M. Lin, S.-K. Chang, J.-H. Wu, C.H. Juang, Neural network-based model for assessing failure potential of
highway slopes in the Alishan, Taiwan Area: Pre- and post-earthquake investigation, Eng. Geol., 104 (2009)
280-289.
[6] H.A. Nefeslioglu, E. Sezer, C. Gokceoglu, A.S. Bozkir, T.Y. Duman, Assessment of Landslide Susceptibility by
Decision Trees in the Metropolitan Area of Istanbul, Turkey, Math. Probl. Eng., 2010 (2010).
[7] S.K. Das, R.i. Biswal, N. Sivakugan, B. Das, Classification of slopes and prediction of factor of safety using
differential evolution neural networks, Environ. Earth Sci., 64 (2011) 201-210.
[8] M.Y. Cheng, C.H. Ko, Automated Safety Monitoring and Diagnosis System for Unstable Slopes, Comput-aided
Civ. Inf., 18 (2003) 64-77.
[9] M.-Y. Cheng, A.F.V. Roy, K.-L. Chen, Evolutionary risk preference inference model using fuzzy support vector
machine for road slope collapse prediction, Expert Syst. Appl., 39 (2012) 1737-1746.
[10] A. Ahangar-Asr, A. Faramarzi, A.A. Javadi, A new approach for prediction of the stability of soil and rock
slopes, Eng. Computation, 27 (2010) 878 - 893.
[11] R. Baker, Sufficient conditions for existence of physically significant solutions in limiting equilibrium slope
stability analysis, Int. J. Solids. Struct, 40 (2003) 3717-3735.
[12] Y. Song, J. Gong, S. Gao, D. Wang, T. Cui, Y. Li, B. Wei, Susceptibility assessment of earthquake-induced
landslides using Bayesian network: A case study in Beichuan, China, Comput. Geosci., 42 (2012) 189-199.
[13] A.K. Sinha, M. Sengupta, Expert system approach to slope stability, Mining Science and Technology, 8 (1989)
21-29.
[14] K. Muthu, M. Petrou, Landslide-Hazard Mapping Using an Expert System and a GIS, IEEE Trans. Geosci.
Remote Sens., 45 (2007) 522-531.
[15] M.-Y. Cheng, N.-D. Hoang, A novel groutability estimation model for ground improvement projects in sandy
silt soil based on Bayesian framework, Tunn.Undergr. Sp. Tech., 43 (2014) 453-458.
[16] P. Lu, M.S. Rosenbaum, Artificial Neural Networks and Grey Systems for the Prediction of Slope Stability, Nat.
Hazards, 30 (2003) 383-398.
[17] K.-p. Zhou, Z.-Q. Chen, Stability Prediction of Tailing Dam Slope Based on Neural Network Pattern
Recognition, in: In Proc. of the Second International Conference on Environmental and Computer Science,
Dubai, 2009, pp. 380-383.
[18] j. Jiang, BP neural networks for Prediction of the factor of safety of slope stability, In Proceedings of the
International Conference on Computing, Control and Industrial Engineering (CCIE), Wuhan, China, (2011).
[19] S.E. Cho, Probabilistic stability analyses of slopes using the ANN-based response surface, Comput. Geotech.,
36 (2009) 787-797.
[20] T.-l. Lee, H.-m. Lin, Y.-p. Lu, Assessment of highway slope failure using neural networks, J. Zhejiang Univ.
Sci. A, 10 (2009) 101-108.
[21] H.B. Wang, W.Y. Xu, R.C. Xu, Slope stability evaluation using Back Propagation Neural Networks, Eng. Geol.,
80 (2005) 302-315.
[22] H. Zhao, S. Yin, Z. Ru, Relevance vector machine applied to slope stability analysis, Int. J. Numer. Anal. Meth.
Geomech., 36 (2012) 643–652.
[23] J. Li, F. Wang, Study on the Forecasting Models of Slope Stability under Data Mining, in: In Proc. of the Earth
and Space 2012: : Engineering, Science, Construction, and Operations in Challenging Environments, Honolulu,
Hawaii, United States, ASCE, 2010, pp. 765-776.
[24] M.-Y. Cheng, Y.-W. Wu, K.-L. Chen, Risk Preference Based Support Vector Machine Inference Model for
Slope Collapse Prediction, Autom. Constr., 22 (2012) 175-181.
[25] H.-b. Zhao, Slope reliability analysis using a support vector machine, Comput. Geotech., 35 (2008) 459-467.
[26] P. Samui, Slope stability analysis: a support vector machine approach, Environ Geol, 56 (2008) 255-267.
[27] J. Li, M. Dong, Method to Predict Slope Safety Factor Using SVM, in: In Proc. of the Earth and Space 2012:
Engineering, Science, Construction, and Operations in Challenging Environments, Pasadena, California, United
States, ASCE, 2012, pp. 888-899.
[28] S. Kiranyaz, T. Ince, A. Yildirim, M. Gabbouj, Evolutionary artificial neural networks by multi-dimensional
particle swarm optimization, Neural Net., 22 (2009) 1448-1462.
[29] M.-Y. Cheng, N.-D. Hoang, Interval Estimation of Construction Cost at Completion Using Least Squares
Support Vector Machine, J. Civ. Eng. Manag., 20 (2013) 223-236.
[30] J.M. Keller, M.R. Gray, J.A. Given, A Fuzzy K-Nearest Neighbor Algorithm, IEEE T. Syst. Man Cy., 15 (1985)
580-585.
[31] S.-T. Li, H.-F. Ho, Predicting financial activity with evolutionary fuzzy case-based reasoning, Expert Syst.
Appl., 36 (2009) 411-422.
[32] M. Govindarajan, R.M. Chandrasekaran, Evaluation of k-Nearest Neighbor classifier performance for direct
marketing, Expert. Syst. Appl., 37 (2010) 253-258.
[33] H.-L. Chen, C.-C. Huang, X.-G. Yu, X. Xu, X. Sun, G. Wang, S.-J. Wang, An efficient diagnosis system for
detection of Parkinson’s disease using fuzzy k-nearest neighbor approach, Expert. Syst. Appl., 40 (2013) 263-
271.
[34] M. Cheng, N. Hoang, Groutability Estimation of Grouting Processes with Microfine Cements Using an
Evolutionary Instance-Based Learning Approach, J. Comput. Civ. Eng., ASCE, 28 (2014) 04014014.
[35] M.-Y. Cheng, N.-D. Hoang, L. Limanto, Y.-W. Wu, A novel hybrid intelligent approach for contractor default
status prediction, Knowl.-Based Syst., 71 (2014) 314-321.
[36] H.-L. Chen, B. Yang, G. Wang, J. Liu, X. Xu, S.-J. Wang, D.-Y. Liu, A novel bankruptcy prediction model
based on an adaptive fuzzy k-nearest neighbor method, Knowl.-Based Syst., 24 (2011) 1348-1359.
[37] G. Kim, D. Seo, K. Kang, Hybrid Models of Neural Networks and Genetic Algorithms for Predicting
Preliminary Cost Estimates, J. Comput. Civ. Eng., ASCE, 19 (2005) 208-211.
[38] W. Zhang, P. Niu, G. Li, P. Li, Forecasting of turbine heat rate with online least squares support vector machine
based on gravitational search algorithm, Knowl.-Based Syst., 39 (2013) 34-44.
[39] T. Xiong, Y. Bao, Z. Hu, Multiple-output support vector regression with a firefly algorithm for interval-valued
stock price index forecasting, Knowl.-Based Syst., 55 (2014) 87-100.
[40] X.-S. Yang, Firefly algorithm, Luniver Press, Bristol, UK, (2008).
[41] B. Amiri, L. Hossain, J.W. Crawford, R.T. Wigand, Community Detection in Complex Networks: Multi–
objective Enhanced Firefly Algorithm, Knowl.-Based Syst., 46 (2013) 1-11.
[42] I. Fister, I. Fister Jr, X.-S. Yang, J. Brest, A comprehensive review of firefly algorithms, Swarm. Evol. Comput.,
13 (2013) 34-46.
[43] L.d.S. Coelho, V.C. Mariani, Improved firefly algorithm approach applied to chiller loading for energy
conservation, Energ. Buildings, 59 (2013) 273-278.
[44] A. Baykasoğlu, F.B. Ozsoydan, An improved firefly algorithm for solving dynamic multidimensional knapsack
problems, Expert Systems with Applications, 41 (2014) 3712-3725.
[45] S.Z. Chen, L.J. Hu, G.H. Chen, The investigation and remedy on the slope failure of Mt. So-San, Hazard
Mitigation Report No. 77-12, National Science Council, Taiwan, Republic of China (1988).
[46] J.B. Hsu, A study of landslide damage risk evaluation model for mountain roads, MS Thesis, Department of
Construction Engineering, National Taiwan Univ. of Sci. and Tech., (2006).
[47] D.H. Lee, A study on mechanical properties of representative rocks at Mt. So-San and in situ measurements of
slope movements, Report to the National Science Council No. 77-0414-P006-12B, National Cheng Kung
University, Taiwan, Republic of China, (1989).
[48] C.T. Lee, C.C. Huang, J.F. Lee, K.L. Pan, M.L. Lin, J.J. Dong, Statistical approach to storm event-induced
landslides susceptibility, Nat. Hazards Earth Syst. Sci., 8(4), 941–960, (2008).
[49] D.H. Rieke-Zapp, M.A. Nearing, Slope Shape Effects on Erosion: A Laboratory Study, Soil Sci. Soc. Am. J.
69:1463–1471, (2005).
[50] W.-T. Lin, W.-C. Chou, C.-Y. Lin, Earthquake-induced landslide hazard and vegetation recovery assessment
using remotely sensed data and a neural network-based classifier: a case study in central Taiwan, Nat. Hazards,
47 (2008) 331-347.
[51] J. Suykens, J.V. Gestel, J.D. Brabanter, B.D. Moor, J. Vandewalle, Least Square Support Vector Machines,
World Scientific Publishing Co. Pte. Ltd., Singapore, (2002).
[52] C.W. Shu, C.C. Chang, C.J. Lin, A practical guide to support vector classification, Technical Report.
Department of Computer Science, National Taiwan University, (2010).
[53] S. Samarasinghe, Neural Networks for Applied Sciences and Engineering, Taylor and Francis (2006).
[54] S.J. Russell, P. Norvig, Artificial Intelligence A Modern Approach, 2nd Edition, Prentice Hall, Person
Education, Inc, (2003).
[55] M.-Y. Cheng, N.-D. Hoang, Groutability prediction of microfine cement based soil improvement using
evolutionary LS-SVM inference model, J. Civ. Eng. Manag., (2014) 1-10.
[56] V. López, A. Fernández, S. García, V. Palade, F. Herrera, An insight into classification with imbalanced data:
Empirical results and current trends on using data intrinsic characteristics, Inform. Sciences, 250 (2013) 113-
141.
[57] T. Fawcett, An introduction to ROC analysis, Pattern Recognition Letters, 27 (2006) 861-874.
[58] A.P. Bradley, The use of the area under the ROC curve in the evaluation of machine learning algorithms,
Pattern Recognit., 30 (1997) 1145-1159.
[59] H. Jin, C.X. Ling, Using AUC and accuracy in evaluating learning algorithms, IEEE Trans. Knowl. Data Eng.,
17 (2005) 299-310.
[60] H.P. Tserng, G.-F. Lin, L.K. Tsai, P.-C. Chen, An enforced support vector machine model for construction
contractor default prediction, Autom. Constr., 20 (2011) 1242-1249.
 List of Figures

Fig. 1 The Firefly Algorithm pseudo code ................................................................................................. 21

Fig. 2 Swarm-Optimized Fuzzy Instance-based Learning (SOFIL) model for predicting slope collapses 22
Fig. 3 The result of the SOFIL’s evolutionary process .............................................................................. 23
Begin FA
Define objective function f(x), where x=(x1,...,xd)
Generate an initial population of fireflies
Formulate the light intensity I
Define the absorption coefficient γ
While (t < Max_Generation)
For i = 1 to n (all n fireflies)
For j=1 to n (all n fireflies)
If (Ij > Ii), move firefly i towards firefly j
End if
Evaluate new solutions and update light intensity;
End for j
End for i
Rank the fireflies and find the current best
End while
End FA
Fig. 1 The Firefly Algorithm pseudo code
Fig. 2 Swarm-Optimized Fuzzy Instance-based Learning (SOFIL) model for predicting slope
collapses
0.0125
Best Fitness

0.0120

0.0115

0.0110

0.0105

0.0100
0 50 100 150 200 250 300
Iterations

Fig. 3 The result of the SOFIL’s evolutionary process

 List of Tables

Table 1 Influencing factors and statistical description ............................................................................... 25

Table 2 Historical database ........................................................................................................................ 26
Table 3 Prediction results of the SOFIL for the testing data ...................................................................... 27
Table 4 Confusion matrices ....................................................................................................................... 29
Table 5 Result comparison ........................................................................................................................ 29
 Table 1 Influencing factors and statistical description

Group Notation Description Min Average Std. Dev. Max

Landforms X1 Slope aspect 0.0 154.1 99.8 355.0
X2 Slope gradient 10.0 60.7 14.2 90.0
X3 Slope height 5.0 20.7 16.8 150.0
X4 Slope form -55.4 1.0 18.7 50.0
Geological Structure X5 Formation type 1.0 4.8 0.9 6.0
Stratigraphy X6 Angle between slope aspect and trend 0.0 89.7 32.1 180.0
X7 Angle between gradient and inclination -20.0 46.2 23.8 85.0
Rock Properties X8 Rock mass size 0.1 0.5 0.4 2.5
X9 Rock mass volume 20.0 68.0 17.2 100.0
Vegetation coverage X10 Vegetation coverage percentage 5.0 59.8 27.6 98.0
X11 Vegetation coverage thickness 0.1 1.4 1.1 4.5
Water condition X12 Catchment area 197.0 23553.5 69365.2 888751.0
Road Properties X13 Excavation height at slope toe 0.0 4.7 5.4 50.0
X14 Change of slope gradient due to toe cutting 0.0 9.8 11.6 45.0
Earthquake X15 Maximum ground acceleration 0.0 249.1 84.1 391.9
Rainfall X16 Maximum accumulated rainfall 384.6 1238.0 525.8 1947.3
 Table 2 Historical database

Case X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 Output

1 0.27 0.69 0.24 0.50 0.80 0.50 0.33 0.06 0.69 0.70 0.33 0.01 0.10 0.00 0.69 0.28 1
2 0.56 0.63 0.07 0.61 0.80 0.56 0.48 0.17 0.75 0.32 0.06 0.00 0.02 0.11 0.62 0.12 1
3 0.79 0.88 0.07 0.31 0.80 0.50 0.95 0.08 0.88 0.48 0.01 0.00 0.03 0.00 0.60 0.63 -1
4 0.65 0.25 0.08 0.46 0.80 0.50 0.48 0.08 0.38 0.70 0.55 0.00 0.06 0.00 0.89 1.00 -1
5 0.35 0.75 0.06 0.58 0.80 0.61 0.71 0.25 0.69 0.81 0.21 0.00 0.16 0.33 0.95 1.00 -1
… … … … … … … … … … … … … … … … … …
207 0.54 0.56 0.07 0.45 0.80 0.50 0.71 0.21 0.69 0.65 0.21 0.00 0.20 0.44 0.98 1.00 -1
208 0.37 0.19 0.03 0.79 1.00 0.50 0.43 0.04 0.38 0.91 0.33 0.03 0.02 0.00 0.61 0.63 -1
209 0.28 0.63 0.04 0.79 0.80 0.50 0.76 0.17 0.63 0.75 0.55 0.00 0.06 0.00 0.50 1.00 -1
210 0.51 0.50 0.24 0.62 0.80 0.42 0.05 0.13 0.75 0.43 0.08 0.05 0.04 0.67 0.64 0.12 1
211 0.27 0.75 0.28 0.47 0.80 0.94 0.76 0.06 0.69 0.38 0.21 0.01 0.16 0.11 0.69 0.04 1

26
 Table 3 Prediction results of the SOFIL for the testing data
Influencing factors Membership Results
No
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 µ 1(X) µ 2(X) YA YP
1 0.73 0.75 0.06 0.46 0.80 0.50 0.86 0.21 0.69 0.81 0.55 0.02 0.06 0.00 0.47 1.00 1.00 0.00 -1 -1
2 0.54 0.50 0.07 0.49 0.00 0.50 0.67 0.06 0.38 0.81 0.21 0.00 0.04 0.00 0.52 1.00 0.78 0.22 -1 -1
3 0.37 0.63 0.66 0.30 0.80 0.86 0.32 0.27 0.75 0.05 0.06 0.28 0.10 0.33 0.64 0.46 0.03 0.97 1 1
4 0.20 0.75 0.10 0.62 0.80 0.50 0.48 0.25 0.75 0.65 0.06 0.03 0.02 0.00 0.68 0.63 0.73 0.27 -1 -1
5 0.39 0.44 0.10 0.57 1.00 0.50 0.62 0.00 0.69 0.00 0.01 0.05 0.16 0.78 0.75 0.03 0.00 1.00 1 1
6 0.27 0.81 0.38 0.10 0.80 0.67 0.71 0.06 0.50 0.16 0.21 0.06 0.26 0.00 0.87 0.10 0.01 0.99 1 1
7 0.13 0.75 0.07 0.49 0.80 0.50 0.38 0.42 0.88 0.75 0.44 0.00 0.10 0.00 0.53 1.00 0.99 0.01 -1 -1
8 0.44 0.75 0.09 0.58 1.00 0.50 0.86 0.04 0.88 0.00 0.10 0.02 0.08 0.22 0.66 0.38 0.04 0.96 1 1
9 0.27 0.69 0.24 0.50 0.80 0.50 0.33 0.06 0.69 0.70 0.33 0.01 0.10 0.00 0.69 0.28 0.20 0.80 1 1
10 0.08 0.63 0.10 0.48 0.40 0.50 0.76 0.08 0.56 0.81 0.44 0.00 0.10 0.22 0.52 0.37 0.22 0.78 1 1
11 0.44 0.63 0.17 0.34 0.80 0.50 0.76 0.06 0.38 0.75 0.44 0.03 0.16 0.33 0.90 0.32 0.08 0.92 1 1
12 0.42 0.50 0.24 0.35 0.80 0.50 0.67 0.02 0.56 0.86 1.00 0.05 0.06 0.00 0.52 0.01 0.21 0.79 1 1
13 0.28 0.69 0.10 0.42 0.80 0.50 0.81 0.02 0.38 0.48 0.06 0.02 0.04 0.33 0.63 0.25 0.10 0.90 1 1
14 0.18 0.69 0.03 0.72 0.80 0.50 0.81 0.21 0.38 0.81 0.10 0.00 0.04 0.00 0.65 0.63 0.95 0.05 -1 -1
15 0.82 0.31 0.14 0.63 0.80 0.50 0.52 0.13 0.75 0.91 0.06 0.04 0.02 0.33 0.67 0.63 0.98 0.02 -1 -1
16 0.38 0.38 0.14 0.18 0.80 0.89 0.10 0.04 0.31 0.65 0.06 0.08 0.03 0.56 0.62 0.11 0.15 0.85 1 1
17 0.56 0.63 0.05 0.66 0.80 0.50 0.76 0.25 0.69 0.27 0.10 0.00 0.14 0.44 0.98 1.00 1.00 0.00 -1 -1
18 0.55 0.63 0.12 0.67 0.80 0.50 0.76 0.04 0.38 0.97 0.55 0.02 0.03 0.00 0.63 0.63 0.99 0.01 -1 -1
19 0.61 0.44 0.17 0.28 0.40 0.50 0.62 0.04 0.00 0.16 0.33 0.09 0.12 0.00 0.58 0.39 0.02 0.98 1 1
20 0.25 0.94 0.09 0.63 0.80 0.19 0.57 0.58 0.88 0.22 0.06 0.01 0.02 0.22 0.64 0.10 0.04 0.96 1 1
21 0.00 0.75 0.45 0.28 0.80 0.14 0.57 0.17 0.44 0.05 0.10 1.00 0.20 0.56 0.66 0.04 0.02 0.98 1 1
22 0.72 0.50 0.31 0.69 0.80 0.50 0.67 0.38 0.63 0.27 0.06 0.01 0.02 0.22 0.56 0.33 0.60 0.40 1 -1
23 0.15 0.55 0.14 0.58 1.00 0.50 0.70 0.06 0.50 0.00 0.10 0.01 0.04 0.00 0.59 0.63 0.07 0.93 1 1
24 0.28 0.63 0.04 0.79 0.80 0.50 0.76 0.17 0.63 0.75 0.55 0.00 0.06 0.00 0.50 1.00 1.00 0.01 -1 -1
25 0.82 0.50 0.06 0.61 0.80 0.50 0.67 0.13 0.63 0.81 0.21 0.00 0.10 0.33 0.95 1.00 1.00 0.00 -1 -1
26 0.46 0.75 0.10 0.68 0.40 0.33 0.38 0.08 0.63 0.22 0.33 0.00 0.24 0.22 0.50 0.21 0.02 0.98 1 1
27 0.49 0.63 0.14 0.25 0.40 0.28 0.29 0.06 0.38 0.81 0.44 0.01 0.14 0.00 0.50 0.36 0.09 0.91 1 1
28 0.00 0.63 0.14 0.55 0.20 0.50 0.76 0.02 0.38 0.86 0.78 0.01 0.16 0.00 0.50 0.36 0.26 0.74 1 1
29 0.25 0.50 0.04 0.64 0.80 0.08 0.26 0.02 0.38 0.91 0.21 0.06 0.02 0.00 0.67 0.63 0.93 0.07 -1 -1
30 0.15 0.63 0.07 0.61 0.80 0.50 0.76 0.29 0.69 0.97 0.66 0.00 0.06 0.00 0.47 1.00 1.00 0.00 -1 -1
31 0.56 0.63 0.07 0.61 0.80 0.56 0.48 0.17 0.75 0.32 0.06 0.00 0.02 0.11 0.62 0.12 0.02 0.98 1 1

27
 32 0.89 0.56 0.05 0.55 0.80 0.50 0.71 0.25 0.69 0.86 0.44 0.00 0.06 0.00 0.47 1.00 1.00 0.00 -1 -1
33 0.54 0.75 0.31 0.50 0.80 0.50 0.81 0.08 0.69 0.59 0.21 0.00 0.60 0.22 0.00 0.72 0.19 0.81 1 1
34 0.04 0.94 0.06 0.64 0.80 0.58 0.71 0.58 0.94 0.38 0.06 0.00 0.02 0.11 0.64 0.63 0.66 0.34 -1 -1
35 0.23 0.63 0.05 0.72 0.80 0.50 0.76 0.25 0.75 0.86 0.33 0.00 0.06 0.00 0.83 1.00 1.00 0.00 -1 -1
36 0.24 0.63 0.08 0.45 0.80 0.50 0.76 0.21 0.63 0.81 0.55 0.01 0.06 0.00 0.53 1.00 1.00 0.00 -1 -1
37 0.37 0.19 0.03 0.79 1.00 0.50 0.43 0.04 0.38 0.91 0.33 0.03 0.02 0.00 0.61 0.63 1.00 0.00 -1 -1
38 0.00 0.75 0.14 0.27 0.80 0.50 0.86 0.25 0.50 0.81 0.66 0.13 0.18 0.67 0.00 0.98 0.80 0.20 1 -1
39 0.39 0.63 0.03 0.45 1.00 0.50 0.76 0.04 0.00 0.05 0.21 0.06 0.05 0.44 0.64 0.39 0.03 0.97 1 1
40 0.34 0.63 0.10 0.54 0.40 0.00 0.29 0.08 0.56 0.81 0.44 0.01 0.10 0.00 0.00 0.36 0.00 1.00 1 1
41 0.00 0.31 0.09 0.15 0.80 0.67 0.24 0.08 0.38 0.59 0.01 0.01 0.07 0.00 0.62 0.12 0.20 0.80 1 1
42 0.06 0.69 0.10 0.41 0.80 0.50 0.81 0.02 0.13 0.16 0.10 0.04 0.03 0.00 0.64 0.45 0.06 0.94 1 1
Note: YA and YP represent actual and predicted slope collapse, respectively. The output of 1 and -1 indicate a failed slope and a stable slope, respectively.

28
 Table 4 Confusion matrices

Training Predicted collapsed slope Predicted stable slope

Actual collapsed slope 76 3
Actual stable slope 1 89
SOFIL
Testing Predicted collapsed slope Predicted stable slope
Actual collapsed slope 18 2
Actual stable slope 0 22
Training Predicted collapsed slope Predicted stable slope
Actual collapsed slope 68 8
Actual stable slope 3 90
SVM
Testing Predicted collapsed slope Predicted stable slope
Actual collapsed slope 22 3
Actual stable slope 0 17
Training Predicted collapsed slope Predicted stable slope
Actual collapsed slope 73 6
Actual stable slope 5 85
RVM
Testing Predicted collapsed slope Predicted stable slope
Actual collapsed slope 21 4
Actual stable slope 1 16
Training Predicted collapsed slope Predicted stable slope
Actual collapsed slope 73 10
Actual stable slope 4 82
ANN
Testing Predicted collapsed slope Predicted stable slope
Actual collapsed slope 17 3
Actual stable slope 2 20
Training Predicted collapsed slope Predicted stable slope
Actual collapsed slope 72 12
Actual stable slope 5 80
FKNN
Testing Predicted collapsed slope Predicted stable slope
Actual collapsed slope 16 2
Actual stable slope 4 20

Table 5 Result comparison

29
 Models FKNN ANN RVM SVM SOFIL
Training
CAR 89.94% 91.71% 93.49% 93.49% 97.50%
AUC 0.90 0.92 0.94 0.93 0.98
No. of correct cases 152 155 158 158 165
No. of wrong cases 17 14 11 11 4
Testing
CAR 85.71% 88.10% 88.10% 92.85% 95.24%
AUC 0.86 0.88 0.89 0.94 0.95
No. of correct cases 36 37 37 39 40
No. of wrong cases 6 5 5 3 2

30
 Research highlights

• A novel AI model (SOFIL) for predicting slope collapses is proposed.

• The SOFIL integrates the Fuzzy K-NN and the Firefly algorithms.

• Slope observations were collected to construct the prediction model.

• The new method can outperform other benchmarking algorithms.

31