Mining critical least association rules from students suffering study anxiety datasets

Abstract

In data mining, association rules mining is one of the common and popular techniques used in various domain applications. The purpose of this study is to apply an enhanced association rules mining method, so called significant least pattern growth for capturing interesting rules from students suffering from exam, family, presentation and library anxiety datasets. The datasets were taken from a survey among engineering students in Universiti Malaysia Pahang. The results of this research will provide useful information for educators to make more accurate decisions concerning their students, and to adapt their teaching strategies accordingly. Moreover, it can also highlight the role of non-academic staff in supporting learning environments for students. The obtained findings can be very helpful in assisting students to handle their fear and anxiety, and, finally, increasing the quality of the learning processes at the university.

1 Introduction

Anxiety is a psychological and physical response to treat a self-concept characterized by subjective, consciously perceived feelings of tension (Spielberger and Vagg 1995). The root meaning of the word anxiety is ‘to vex or trouble’; in either presence or absence of psychological stress, anxiety can create feelings of fear, worry, uneasiness, and dread (Davison et al. 2008). It is also associated with feelings of restlessness, fatigue, concentration problems, and muscle tension (Bouras and Holt 2007). Anxious students have experience of cognitive deficits like misapprehension of information or blocking of memory and recall (Vitasari et al. 2010, 2011).

Many educators find themselves overwhelmed with data concerning students suffering anxiety, but lack the information they need to make informed decisions. Currently, there is an increasing interest in data mining and educational systems, making educational data mining a new growing research community (Romero and Ventura 2007). One of the popular educational data mining techniques is Association Rules Mining (ARM). Association rules mining has been widely studied in the knowledge discovery community (Ceglar and Roddick 2006). It aims to discover the interesting correlations, frequent patterns, associations or causal structures among sets of items in the data repositories. The problem of association rules mining was first introduced by Agrawal for market-basket analysis (Agrawal et al. 1993a, b; Aggarwal and Srikant 1994). There are two main stages involved before producing the association rules: first, find all frequent items from transactional database, and, second, generate the common association rules from the frequent items.

Generally, an item is said to be frequent if it appears more than the minimum support threshold. These frequent items are then used to produce the ARs. In addition, confidence is another measure that is always used in tandem with the minimum support threshold. By definition, the least item is an itemset that is rarely found in the database but may produce interesting and useful ARs. These types of rule are very meaningful in discovering rarely occurring but significantly important, such as air pollution detection, critical fault detections, and network intrusions, and their possible causes. For past developments, many series of ARs mining algorithms used the minimum support-confidence framework to avoid the overloading of ARs. The challenge is that by increasing or decreasing the minimum support or confidence values, the interesting rules might be missing or untraceable. Because of the complexity of study and difficulty in algorithms (Agrawal et al. 1993a) it may require excessive computational cost, and, furthermore, very limited attention has been given to discover the highly correlated least ARs. In terms of relationship, the frequent and least ARs have a different degree of correlation. Highly correlated least ARs refer to the itemsets the frequency of which does not satisfy a minimum support but are very highly correlated. ARs are classified as highly correlated if there is positive correlation, and, at the same time, they fulfil a minimum degree of predefined correlation. Until this moment, the statistical correlation technique has been successfully applied in the transaction databases (Agrawal et al. 1993b), to find a relationship among pairs of items and determine whether they are highly positive or negatively correlated. As a matter of fact, it is not absolutely true that the frequent items have a positive correlation as compared to the least items. In previous papers, we address the problem of mining least ARs with the objective of discovering significant least ARs, which, surprisingly, are highly correlated (Abdullah et al. 2010a, b). A new algorithm named significant least pattern growth (SLP-Growth) to extract these ARs is proposed (Abdullah et al. 2010a). The proposed algorithm imposes interval support to extract all least item set families first before continuing to construct a significant least pattern tree (SLP-tree). The correlation and critical relative support measures for finding the relationship between itemsets are also embedded into this algorithm (Abdullah et al. 2010a, b). In this paper, we explore the SLP-Growth algorithm for capturing interesting rules in students suffering anxiety from exams, family, presentation and library datasets. The datasets were taken from a survey on exploring several types of anxiety among engineering students in Universiti Malaysia Pahang (UMP). The results of this research will provide useful information for educators to make more accurate decisions concerning their students, and to adapt their teaching strategies accordingly. In addition, the good support from non-academic staff is also important in providing a better learning environment for the students. It can also be helpful in assisting students to handle their anxiety and be useful in increasing the quality of learning.

The remainder of this paper is organized as follows. Section 2 describes the related work and the basic concepts and terminology of association rules mining. Section 3 describes the proposed association rules method, the so-called SLP-Growth algorithm. Section 4 describes the scenario in mining information from the four types of anxiety datasets. This is followed by performance analysis through the four types of anxiety datasets and the results are presented in Sect. 3. Finally, the conclusions of this work are reported in Sect. 4.

2 Methodology

2.1 Association rules mining

For the past two decades, several attempts have been made to discover scalable and efficient methods for mining frequent ARs. However, mining least ARs still lags behind. As a result, ARs that are rarely found in the database are pruned out by the minimum support-confidence threshold. As a matter of fact, the rare ARs can also reveal useful information for detecting the highly critical and exceptional situations. Zhou et al. (2007) suggested a method to mine the ARs by only considering infrequent itemsets. The drawback is that the matrix-based scheme (MBS) and hash-based scheme (HBS) algorithms are very expensive in terms of hash collision. Ding (2005) proposed the transactional co-occurrence matrix (TCOM) for mining association rule among rare items. However, implementation wise it is quite complex and costly. Yun et al. (2003) introduced the relative support apriori algorithm (RSAA) to generate rare itemsets. The challenge is that it takes a similar time as performed by Apriori if the allowable minimum support is set very low. Koh and Rountree (2005) suggested the Apriori–Inverse algorithm to mine infrequent itemsets without generating any frequent rules. However, it suffers from candidate itemset generations and is costly in generating the rare ARs. Liu et al. (1999) proposed the multiple support apriori (MSApriori) algorithm to extract the rare ARs. In actual implementation, this algorithm faces the “rare item problem”. From the proposed approaches (Zhou and Yau 2007; Yun et al. 2003; Ding 2005; Koh and Rountree 2005; Liu et al. 1999), many of them use the percentage-based approach to improve the performance as faced by the single minimum support based approaches. In terms of measurement, Brin et al. (1997) introduced the objective measure called lift and chi-square as a correlation measure for ARs. Lift compares the frequency of pattern against a baseline frequency computed under the statistical independence assumption. Omiecinski (2003) proposed two interesting measures based on the downward closure property called all confidence and bond. Lee et al. (2003) suggested two algorithms for mining all confidence and bond correlation patterns by extending the pattern-growth methodology Han et al. (2000). In terms of mining algorithms, Agrawal et al. (1993a) proposed the first ARs mining algorithm called Apriori. The main bottleneck of Apriori is that it requires multiple scanning of the transaction database and also generates a huge number of candidate itemsets. Han and Kamber (2006) suggested the FP-Growth algorithm, which, amazingly, can overcome the two limitations faced by the Apriori series algorithms. Currently, FP-Growth is one of the fastest approaches and most benchmarked algorithms for frequent itemsets mining. It is derived based on a prefix tree representation of database transactions (called an FP-tree).

2.2 Association rules (ARs)

ARs were first proposed for market basket analysis to study customer purchasing patterns in retail stores (Agrawal et al. 1993a, b). Recently, ARs have been used in many applications or disciplines, such as customer relationship management (Au and Chan 2003), image processing (Aggarwal and Yu 1998; Li et al. 2008), mining air pollution data (Mustafa et al. 2006), educational data mining (Abdullah et al. 2011a, b; Herawan et al. 2011), information visualization (Herawan et al. 2009; Abdullah et al. 2011c) and text mining. Typically, association rule mining is the process of discovering associations or correlation among itemsets in transaction databases, relational databases and data warehouses. There are two subtasks involved in ARs mining: generating frequent itemsets that satisfy the minimum support threshold and generating strong rules from the frequent itemsets.

Throughout this section the set, $I=\{i_1 ,i_2 , \ldots , i_{|A|} \}$ for $|A| > 0$ refers to the set of literals called set of items and the set $D=\{t_1 , t_2 , \ldots , t_{|U|} \}$, for $|U|>0$ refers to the data set of transactions, where each transaction $t\in D$ is a list of distinct items $t=\{i_1 ,i_2 , \ldots , i_{|M|} \},1\le |M| \le |A|$ and each transaction can be identified by a distinct identifier TID.

Definition 1

A set $X\subseteq I$ is called an itemset. An itemset with k-items is called a k-itemset.

Definition 2

The support of an itemset $X\subseteq I$, denoted $\hbox {supp}(X)$ is defined as a number of transactions contain X.

Definition 3

Let $X,Y\subseteq I$ be itemset. An association rule between sets X and Y is an implication of the form $X\Rightarrow Y$, where $X\cap Y=\phi$ The sets X and Y are called antecedent and consequent, respectively.

Definition 4

The support for an association rule $X\Rightarrow Y$, denoted $\hbox {supp}(X\Rightarrow Y)$, is defined as a number of transactions in D contain $X\cup Y$

Definition 5

The confidence for an association rule $X\Rightarrow Y$ , denoted $\hbox {conf}(X\Rightarrow Y)$ is defined as a ratio of the number of transactions in D containing $X\cup Y$ to the number of transactions in D containing X. Thus

$$\begin{aligned} \hbox {conf}(X\Rightarrow Y)=\frac{\sup \hbox {p}(X\Rightarrow Y)}{\sup \hbox {p}(X)}. \end{aligned}$$

An item is a set of items. A $k-$ itemset is an itemset that contains $k$ items. An itemset is said to be frequent if the support count satisfies a minimum support count (minsupp). The set of frequent itemsets is denoted as $L_k$ .The support of the ARs is the ratio of transaction in $D$ that contain both $X$ and $Y$ (or $X\cup Y)$. The support can also be considered as probability $P(X\cup Y)$. The confidence of the ARs is the ratio of transactions in $D$ containing $X$ that also contains $Y$. The confidence can also be considered as conditional probability $P(Y|X)$. ARs that satisfy the minimum support and confidence thresholds are said to be strong.

2.3 Correlation analysis

After the introduction of ARs, many researches including Brin et al. (1997) had realized the limitation of the confidence-support framework. When utilizing this framework alone it is quite impossible to discover the interesting ARs. Therefore, the correlation measure can be used as a complementary measure together with this framework. This leads to correlation rules as:

$$\begin{aligned} A\Rightarrow B (\hbox {supp}, \hbox {conf}, \hbox {corr}) \end{aligned}$$

(1)

The correlation rule is a measure based on the minimum support, minimum confidence and correlation between itemsets $A$ and $B$. There are many correlation measures applicable for ARs. One of the simplest correlation measures is lift. The occurrence of itemset $A$ is independent of the occurrence of itemset $B$ if $P(A\cup B)=P(A)P(B)$; otherwise itemset $A$ and $B$ are dependent and correlated. The lift between occurrence of itemset $A$ and $B$ can be defined as:

$$\begin{aligned} \hbox {lift}(A,B)=\frac{P(A\cap B)}{P(A)P(B)} \end{aligned}$$

(2)

The equation of (4) can be derived to produce the following definition:

$$\begin{aligned} \hbox {lift}(A,B)=\frac{P(B|A)}{P(B)} \end{aligned}$$

(3)

or

$$\begin{aligned} \hbox {lift}(A,B)=\frac{\hbox {conf}(A\Rightarrow B)}{\hbox {supp}(B)} \end{aligned}$$

(4)

The strength of correlation is measured from the lift value. If $\hbox {lift}(A,B)=1$ or $P(B|A)=P(B) (\hbox {or }P(A|B)=P(B))$ then $B$ and $A$ are independent and there is no correlation between them. If $\hbox { lift}(A,B)>1$ or $P(B|A)>P(B) (\hbox {or }P(A|B)>P(B))$, then A and B are positively correlated, meaning the occurrence of one implies the occurrence of the other. If $\hbox {lift}(A,B)<1$ or $P(B|A)<P(B) (\hbox {or }P(A|B)<P(B))$, then A and B are negatively correlated, meaning the occurrence of one discourages the occurrence of the other. Since the lift measure is not downward closed, it will definitely not suffer from the least item problem. Thus, least itemsets with low counts, which perchance occur a few times (or only once) together can produce enormous lift values.

2.4 Critical relative support

Definition 6

(Critical Relative Support). A critical relative support (CRS) is a formulation of maximizing relative frequency between itemset and their Jaccard similarity coefficient.

The value of Critical Relative Support denoted as CRS and

$$\begin{aligned} \hbox {CRS}(I)=\max \left( {\left( {\frac{\hbox {supp}(\hbox {A)}}{\hbox {supp}(\hbox {B)}}} \right) , \left( {\frac{\hbox {supp}(\hbox {B)}}{\hbox {supp}(\hbox {A)}}} \right) } \right) \times \left( {\frac{\hbox {supp}(A\Rightarrow B)}{\hbox {supp}(\hbox {A})+\hbox {supp}(\hbox {B})-\hbox {supp}(A\Rightarrow B)}} \right) \end{aligned}$$

The CRS value is between 0 and 1, and is determined by multiplying the highest value, either support of antecedent divide by consequence or alternatively with their Jaccard similarity coefficient. This is a measurement to show the level of CRS between combinations of both the least items and frequent items either as antecedent or consequence, respectively.

2.5 FP-growth

Candidate set generation and tests are two major drawbacks in Apriori-like algorithms. Therefore, to deal with this problem, a new data structure called frequent pattern tree (FP-Tree) was introduced. FP-Growth was then developed based on this data structure, and, currently, is a benchmark and the fastest algorithm in mining frequent itemsets (Han et al. 2000). The advantages of FP-Growth are that it requires scanning the transaction database two times. Firstly, it scans the database to compute a list of frequent items sorted by descending order and eliminates rare items. Secondly, it scans to compress the database into a FP-Tree structure and mines the FP-Tree recursively to build its conditional FP-Tree.

A simulation of data (Au and Chan 2003) is shown in Table 1. Firstly, the algorithm sorts the items in the transaction database with the infrequent items removed. Let us say a minimum support is set to 3, therefore alphabets f, c, a, b, m, p are only kept. The algorithm scans the entire transactions starting from T1 until T5. In T1, it prunes from {f, a, c, d, g, i, m, p} to {f, c, a, m, p, g}. Then, the algorithm compresses this transaction into the prefix tree in which f becomes the root. Each path on the tree represents a set of transactions with the same prefix. This process will execute recursively until the end of the transactions. Once the complete tree has been built, then the next pattern mining can be easily performed.

Table 1 A simple dataset of transactions

2.6 SLP growth algorithm

2.6.1 Algorithm development

Determine interval support for least itemset

Let $I$ be a non-empty set such that $I=\{i_1 ,i_2 , \ldots , i_n \}$, and $D$ is a database of transactions where each $T$ is a set of items such that $T\subset I$ An item is a set of items. A $k$-itemset is an itemset that contains $k$ items. An itemset is said to be least if the support count satisfies in a range of threshold values called Interval Support (ISupp). The Interval Support is a form of ISupp (ISMin, ISMax) where ISMin is a minimum and ISMax is the maximum value, respectively, such that $\hbox {ISMin}\ge \phi , \hbox {ISMax} \hbox {>} \phi$and $\hbox {ISMin}\le \hbox {ISMax}$. The set is denoted as $L_k$ . Itemsets are said to be least significant if they satisfy two conditions. First, support counts for all the items in the itemset that must be greater than ISMin. Second, those itemsets must consist of at least one of the least items. In brevity, the significant least itemset is a union between the least items and frequent items, and the existence of an intersection between them.

Construct significant least pattern tree

A Significant Least Pattern Tree (SLP-Tree) is a compressed representation of significant least itemsets. This tree data structure is constructed by scanning the dataset of a single transaction at a time and then mapping onto a path in the SLP-Tree. In the SLP-Tree construction, the algorithm constructs a SLP-Tree from the database. The SLP-Tree is built with only the items that satisfy the ISupp. In the first step, the algorithm scans all transactions to determine a list of least items, LItems and frequent items, FItems (least frequent item, LFItems). In the second step, all transactions are sorted in descending order and mapped against the LFItems. It is essential that the transactions consist of at least one of the least items, otherwise the transactions are disregarded. In the final step, a transaction is transformed into a new path or mapped into the existing path. This final step continues until the end of the transactions. The problem of the existing FP-Tree is that it may not fit into the memory and is expensive to build. The FP-Tree must be built completely from all the transactions before calculating the support for each item. Therefore, the SLP-Tree is an alternative and more practical to overcome these limitations.

Generate Least Pattern Growth ( LP-Growth)

SLP-Growth is an algorithm that generates significant least itemsets from the SLP-Tree by exploring the tree based on a bottom-up strategy. The ‘divide and conquer’ method is used to decompose the task into a smaller unit for mining desired patterns in conditional databases, which can optimize the searching space. The algorithm will extract the prefix path sub-trees ending with any least item. In each of the prefix path sub-trees, the algorithm will recursively execute to extract all frequent itemsets, and, finally, build a conditional SLP-Tree. A list of least itemsets is then produced based on the suffix sequence and also sequence in which they are found. The pruning process in the SLP-Growth is faster than in the FP-Growth since most of the unwanted patterns are already cut-off during the construction of the SLP-Tree data structure. The complete SLP-Growth algorithm is shown in Fig. 1.

Fig. 1

SLP-Growth algorithm

2.7 Weight assignment

Apply correlation

The weighted ARs (ARs value) are derived from formula (4). This correlation formula is also known by lift. The processes of generating weighted ARs take place after all the patterns and ARs are completely produced.

Discovery of highly correlated least ARs

From the list of weighted ARs, the algorithm will begin to scan all of them. However, only those weighted ARs with a correlation value that is more than one are captured and considered as highly correlated. ARs with a correlation less than one will be pruned and classified as low correlation.

2.8 Scenario on capturing rules

2.8.1 Dataset

The dataset was taken from a survey exploring four types of anxiety among engineering students in Universiti Malaysia Pahang (Vitasari et al. 2010) . A total 770 students participated in this survey. The respondents were 770 students, consisting of 394 males and 376 females. The respondents were undergraduate students from five engineering faculties at Universiti Malaysia Pahang, i.e. 216 students from the Faculty of Chemical and Natural Resources Engineering (FCNRE), 105 students from the Faculty of Electrical and Electronic Engineering (FEEE), 226 students from the Faculty of Mechanical Engineering (FME), 178 students from the Faculty of Civil Engineering and Earth Resources (FCEER), and 45 students from the Faculty of Manufacturing Engineering and Technology Management (FMETM).

2.8.2 Exam anxiety

The survey’s findings indicate that exam anxiety among engineering students is manifested through seven dimensions: (a) Lack of preparation, (b) Feel depressed after test, (c) Lost concentration during exam, (d) Prepared for exam, (e) Do not understand the test question, (f) Important exam, and (g) Take a surprise test. For this, we have a dataset, which comprises the number of transactions (student), which is 770, and the number of items (attributes), which is 5 (refer to Table 2).

Table 2 Exam anxiety dataset

2.8.3 Family anxiety

Family anxiety among engineering students focuses on five dimensions: (a) Insufficient income, (b) Childhood experiences, (c) Many members in family, (d) Parents disappointed, and (e) Family problems. For this, we have a dataset, which comprises the number of transactions (student), which is 770, and the number of items (attributes), which is 5 (refer to Table 3).

Table 3 Family anxiety dataset

2.8.4 Presentation anxiety

Presentation anxiety among engineering students focuses on five dimensions: (a) Anxious time presentation, (b) Lack of confidence, (c) Heart beating very fast, (d) Tongue tied, and (e) Felt that presentation did not contribute to your study. For this, we have a dataset, which comprises the number of transactions (student), which is 770, and the number of items (attributes), which is 5 (refer to Table 4).

Table 4 Presentation anxiety dataset

2.8.5 Library anxiety

Library anxiety among engineering students focuses on six dimensions: (a) Unable to use the library, (b) Lack of references, (c) Feel confused to find the references, (d) Feel uncomfortable in the library, (e) Library is unimportant, and (f) Library staff are unwilling to assist. For this, we have a dataset, which comprises the number of transactions (student), which is 770, and the number of items (attributes), which is 5 (refer to Table 5).

Table 5 Library anxiety dataset

2.8.6 Design

The design for capturing interesting rules on students suffering study anxiety is described in Fig. 2 as follow.

Fig. 2

The procedure of mining critical least association rules

In order to capture the interesting rules and make a decision, the experiment using SLP-Growth method was conducted on $\hbox {Intel}{\circledR }$ $\hbox {Core}^\mathrm{TM}$ 2 Quad CPU at 2.33 GHz speed with 4 GB main memory, running on Microsoft Windows Vista. The algorithm was developed using C# as a programming language. The study anxiety datasets used and the SLP-Growth produced in this model were in a flat file format.

3 Results and discussion

3.1 Exam anxiety

We evaluate the proposed algorithm to exam the anxiety dataset as in Table 2. For this, we have a dataset, which comprises the number of transactions (student), which is 770, and the number of items (attributes), which is 7. Table 6 displays the mapping of the original survey dimensions, and the Likert scale and new attribute ID.

Table 6 The mapping between survey dimensions of exam anxiety, Likert scale and new attribute ID

The item is constructed based on the combination of the survey dimension and its Likert scale. For simplicity, let us consider the survey dimension “Lack of preparation” with Likert scale “1”. Here, item “11” will be constructed by means of a combination of an attribute ID (first characters) and its survey dimension (second character). Different Interval Supports were employed for this experiment.

By embedding FP-Growth algorithm, 7,897 ARs are produced. ARs are formed by applying the relationship of an item or many items to an item (cardinality: many-to-one). Figure 3 depicts the correlation’s classification of interesting ARs. For this dataset, the rule is categorized as significant and interesting if it has a positive correlation and the CRS value should be equal to 1.0. These categorization criterion also will be employed to another four more experiments.

Fig. 3

Classification of ARs using correlation analysis

Only 6.56 % from the total of 7,897 ARs are classified as interesting. Table 7 shows the top 20 interesting ARs with numerous types of measurement. The highest correlation value from the selected ARs is 5.66. From these ARs, one of the dominant items of consequence is item 45 (Prepared for exam is 5). Item 45 appears within the entire dataset 22.47 %. Besides item 22, another item of consequence that occurs in the top 20 interesting ARs is item 32. Item 32 is “Lost concentration during exam is 2”. For item 32, its occurrence in the dataset is 17.66 %. Table 4 also indicates that all interesting ARs have a CRS value equal to 1 and the confidence is 100 %. Therefore, further analysis and study can be used to identify other interesting relationships, such as self-awareness, people’s perception, and classroom climate.

Table 7 Top 20 highest correlations of interesting association rules sorted in descending order of correlation

The results illustrate that CRS is successful in producing a lower number of ARs compared to the other measures. The typical support measure alone is not a suitable measure to be employed to discover the interesting ARs. Although the correlation measure can be used to capture the interesting ARs, its ratio is still nearly 8 times larger than the CRS measure. Therefore, CRS is proven to be more efficient and outperformed the benchmarked measures for discovering the interesting ARs. Figure 4 presents the correlation classification of the interesting ARs using the various Interval Supports.

Fig. 4

Correlation analysis of interesting ARs using a variety of Interval Supports

Generally, the total number of ARs is kept decreased when the predefined Interval Support thresholds are increased. The results show that CRS is successful in reducing the volume of ARs as compared to the others measures. Despite the fact that the correlation measure alone can also extract the interesting ARs, it ratios is 8 times more than the CRS measure. Thus, CRS is more effective in capturing the interesting ARs.

3.2 Family anxiety

The second experiment was done to the family anxiety dataset as in Table 3. To this, we also have the same number of transactions (student) and the number of items (attributes) is 5. Table 8 displays the mapping of the original survey dimensions, Likert scale and new attribute ID.

Table 8 The mapping between survey dimensions of social anxiety, Likert scale and new attribute ID

The construction of the item is similar to that performed on the exam anxiety dataset. For this experiment, different Interval Supports were used. By embedding FP-Growth algorithm, 1,938 ARs are produced. Again, ARs are formed by applying the relationship of an item or many items to an item (cardinality: many-to-one). Figure 5 shows the correlation’s classification of interesting ARs. For this dataset, the evaluation of the significance of the interesting rules is similar to the previous experiment.

Fig. 5

Classification of ARs using correlation analysis

Only 2.01 % from the total of 1,938 ARs are classified as interesting ARs. Table 9 shows the top 20 interesting ARs with numerous types of measurement. The highest correlation value from the selected ARs is 6.94. From these ARs, two dominant items of consequence are item 21 (Childhood experiences is 1) and item 51 (Family problems is 1), respectively. From the entire dataset, item 21 only occurs about 1.43 % compared to item 51, which appears about 14.29 %. Table 9 also shows that all interesting ARs have a CRS value equal to 1. Therefore, further analysis and study can be used to identify other interesting relationships, such as academic performance, personality and attitude. Table 9 displays the correlation analysis based on several Interval Supports (Fig. 6).

Table 9 Top 20 highest correlations of interesting association rules sorted in descending order of correlation

Fig. 6

Correlation analysis of interesting ARs using variety Interval Supports

Generally, the total numbers of ARs are kept decreased when the predefined Interval Support thresholds are increased. The results illustrate that CRS is also successful in extracting a lower number of ARs compared to the other measures. It also indicates that the single measure alone is not a suitable measure to be used in mining interesting ARs. Although, correlation measure can be used to capture the interesting ARs, its ratio is still 28 times larger than the CRS measure. Therefore, CRS is proven to be more appropriate for discovering the interesting ARs.

3.3 Presentation anxiety

The third experiment was done to the presentation anxiety dataset, as in Table 10. To this, we also have the same number of transactions (student) and the number of items (attributes) is 5. Table 10 displays the mapping of the original survey dimensions, Likert scale and new attribute ID.

Table 10 The mapping between survey dimensions of social anxiety, Likert scale and new attribute ID

The construction of the item is similar to that performed to the exam and family anxiety datasets. In this experiment, different Interval Supports were also used. By embedding the FP-Growth algorithm, 1,422 ARs are captured. Again, ARs are formed by applying the relationship of an item or many items to an item (cardinality: many-to-one). Figure 7 shows the correlation’s classification of interesting ARs using the several Interval Supports. For this dataset, the evaluation of the significances of the interesting rules is similar to the previous experiments.

Fig. 7

Classification of ARs using correlation analysis

From the total of 1,422 ARs, only 2.39 % are classified as interesting ARs. Table 11 shows the top 20 interesting ARs with numerous measurements. The highest correlation value from the selected ARs is 8.95. From these ARs, one of the dominant items of consequence is item 12 (Lack of confidence is 2). In fact, item 12 only appears 12.99 % in the entire dataset. Besides item 12, another item of consequence that occurs in the top 20 interesting ARs is item 25. Item 25 is “Lack of confidence is 5”. For item 25, it appears around 13.25 %. Table 11 also shows that all interesting ARs have a CRS value equal to 1. Therefore, further analysis and study can be used to identify other interesting relationships, such as environmental, medical, and genetics. Figure 8 depicts the correlation analysis of ARs against a few Interval Supports,

Table 11 Top 20 highest correlations of interesting association rules sorted in descending order of correlation

Fig. 8

Correlation analysis of interesting ARs using variety Interval Supports

Generally, the total number of ARs decreases when the predefined Interval Supports thresholds are increased. The results also illustrate that CRS is successful in generating fewer ARs compared to the others measures. Although the correlation measure can be used to capture the interesting ARs, its ratio is still 19 times larger than the CRS measure. Therefore, CRS is proven to be more practical in extracting the interesting ARs.

3.4 Library anxiety

The fourth experiment was done on the Library anxiety dataset, as in Table 12. For this, we also have the same number of transactions (student) and the number of items (attributes) is 7. Table  12 displays the mapping of the original survey dimensions, the Likert scale and new attribute ID.

Table 12 The mapping between survey dimensions of social anxiety, Likert scale and new attribute ID

The construction of the item is similar to that performed to previous datasets. For this experiment, different Interval Supports were also employed. By embedding the FP-Growth algorithm, 4,933 ARs are extracted. Again, ARs are formed by applying the relationship of an item or many items to an item (cardinality: many-to-one). Figure 9 shows the correlation classification of interesting ARs. For this dataset, the evaluation of the significance of the interesting rules is similar to the previous experiments.

Fig. 9

Classification of ARs using correlation analysis

Only 3.89 % from the total of 4,933 ARs are classified as interesting ARs. Table 13 shows the top 20 interesting ARs with numerous types of measurement. The highest correlation value from the selected ARs is 9.51. From these ARs, there is only one dominant item of consequence, item 61 (Library staff is unwilling to assist is 1). In fact, item 61 only appears 10.52 % from the entire dataset. Table 13 also shows that all interesting ARs have a CRS value equal to 1. Therefore, further analysis and study can be used to identify other interesting relationships, such as hostel facilities, internet connectivity, and friendship. Figure 10 presents the correlation analysis of ARs according to variety of the Interval Supports.

Table 13 Top 20 highest correlations of interesting association rules sorted in descending order of correlation

Fig. 10

Correlation analysis of interesting ARs using variety Interval Supports

Generally, the total number of ARs decreases when the predefined Interval Supports thresholds increase. The results also show that CRS successfully produced the lowest number of ARs compared to the other measures. The support measure alone is not a suitable measure to be employed to discover the interesting ARs. Although, the correlation measure can be used to capture the interesting ARs, its ratio is still 13 times larger than the CRS measure. Therefore, CRS is proven to be more suitable and scalable in mining the interesting ARs.

4 Conclusion

Many educators find themselves overwhelmed with students suffering anxiety data, but lack the information they need to make informed decisions. Currently, there is an increasing interest in data mining and educational systems, making educational data mining a new growing research community (Romero and Ventura 2007). One of the popular data mining methods is Association Rules Mining (ARM). In this paper, we have successfully applied an enhanced association rules mining method, the so called SLP-Growth (significant least pattern growth) proposed by (Abdullah et al. 2010a), for capturing interesting rules in students suffering study anxiety datasets. The datasets were taken from a survey exploring anxiety pertaining to exams, family, presentation and library among engineering students in Universiti Malaysia Pahang (UMP). A total of 770 students participated in this survey. The respondents comprise 394 males and 376 females. The respondents are undergraduate students from five engineering faculties at Universiti Malaysia Pahang. It is found that the SLP_Growth method is suitable to mine the interesting rules that provide faster and accurate results. Based on the results, educators can obtain recommendations from the rules captured. The results of this research will provide useful information for educators to make a decision on their students more accurately, and to adapt their teaching strategies accordingly. The obtained findings can also be helpful in assisting students to handle their fear and anxiety, and in increasing the quality of the learning.