Research on Data Mining Algorithms for Vocational Education Based on Student Behaviour Analysis

The data used in this paper comes from the open-source dataset of a third-party platform, which consists of data on the behaviour of students using campus one-card swipe cards in two academic years in X university and data on grade ranking in the teaching management system. Specifically, it includes book borrowing data, one-card data, dormitory access data, library access data, and student achievement data. The student information in the dataset has been anonymised through data desensitisation.

In this paper, we study students' campus card behavioural data based on learning and summarising previous experiences to understand that the side-by-side assessment of students' performance rankings related to behaviour is multifaceted. Firstly, statistical analysis methods are used to select the behavioural attributes related to performance ranking. In this paper, 18 features are manually extracted, and based on the existing machine learning algorithms, the most suitable parameters are searched for to form the optimal state for predicting and evaluating the performance of students with different ranking grades. The feature fields about Studying Habits (SH), Living Habits (LH) and ConsumptionHabits (CH) are shown in Table 1:

The study habits of college students are formed over a long period of time, and their tendencies and behaviours cannot be easily changed with external factors. Characteristics 1-8 represent the number of times students borrowed books during non-exam time, the number of times they borrowed books during exam time, the number of times they entered the library earlier than 8 o'clock, the number of times they left the library later than 22 o'clock, the average daily stay in the library, the number of times they went to the library during non-exam time, the number of times they went to the library during exam time, and the number of times they spent money on the POS during class time, which were categorised as study habits. Good and regular habits of living behaviour are beneficial to academic performance, and the regularity of these habits is closely related to students' ability of self-control and self-discipline. In this paper, we choose indicators such as the average number of times per day to fetch water and dormitory stay time as the characteristics of living habits. Characteristics 9-15 represent the number of times students leave the dormitory earlier than 8:00, the number of times they enter the dormitory later than 22:00, the average daily stay in the dormitory, the number of weekly showers, the number of times they eat breakfast earlier than 8:00, the number of times they go to the copy centre in non-examination time periods, and the number of times they go to the copy centre in examination time periods, respectively. There are differences in the amount of campus card spending among college students with different academic achievements, reflecting the different consumption needs and consumption psychology of college students. Characteristics 16-18 represent the average daily consumption in the canteen, the average daily consumption in the supermarket and the average daily number of times to fetch water, respectively.

In order to better interpret the behavioural data and to have a more comprehensive understanding of the data in advance, this paper firstly classifies the students' performance ranking into three categories, i.e. good, fair and poor, based on the normal distribution. The labels were set as "1", accounting for 19.84% of the total number of students, "2", accounting for 60.02% of the total number of students, and "3", accounting for 20.14% of the total number of students, respectively. 20.14 per cent. The campus behavioural and ranking data were then analysed to convert the raw behavioural data into behavioural characteristics related to academic performance. This paper compares the differences in the number of visits to the library, books borrowed and visits to the literature and printing centre between exam and non-exam periods for the three categories of students. Library borrowing data were collected from students of all colleges over a period of two academic years. The results show that the average number of books borrowed by the students of category I is 52, the average number of books borrowed by the students of category II is 48 and the average number of books borrowed by the students of category III is 42. The behaviour of studying and borrowing books in the library decreases in the order of student category. Category 1 students are most equipped with good study and borrowing habits, while Category 3 students have relatively few of the above three behaviours and lack practical action in their daily campus routine.

In order to demonstrate the correlation between students' behavioural traits and academic performance, this paper further uses the statistical methods of factor analysis and principal component analysis to explain the correlation. Firstly, factor analysis is used to extract the common factors, and the factors affecting academic performance are rationalised by rotating the component matrix to explore the weight of each factor and common factor on the impact of performance. The basic structure of the observed data was explored by examining the internal dependencies of students' behavioural characteristics and several hypothetical variables were used to represent their basic data structure.

Before factor analysis, the selected characteristics were tested using KMO and Bartlett's test.KMO test was used to check the correlation and partial correlation between variables.The closer the KMO statistic is to 1, the stronger the correlation between variables and the weaker the partial correlation, the better the factor analysis. KMO result of this research is 0.720 (>0.6), which is suitable for factor analysis.The Bartlett's spherical test determines that if the correlation array is a unit array, the independent factor analysis method for each variable is invalid. The results showed that the approximate chi-square value of Bartlett's test was 47483.939 and the concomitant probability value of Sig. < 0.05 reached the significance level. It indicates that the correlation coefficient matrix of the factors is not a unit matrix and there is correlation among the variables. Therefore, based on the results of the above analysis, the original variables are suitable for factor analysis.

Principal component analysis is to find out the independent composite indicators that reflect multiple variables and how the internal structure between multiple variables can be revealed by several principal components. Using principal component analysis to examine the correlation between multiple variables and the predictive function for 18 factors affecting student achievement rankings, as shown in Tables 3, the first seven factors were able to explain 69.942% of the overall variance, effectively reflecting the overall information and having a significant relationship with academic achievement.

Campus card swiping behaviour belongs to students' spontaneous behaviour and is also an important factor reflecting students' academic performance. Due to the individual differences of each student and the diversity and complexity of environmental factors, it is not possible to generalise and explain the model by extracting campus card swiping behavioural characteristics alone. Behavioural indicators affecting academic ranking can be discussed in the context of student performance classification, and data mining algorithms can be used to explore the practical value of these behavioural data in management, teaching and learning.

In the first stage of the sequence-based performance prediction task, this paper uses a hybrid attention-based encoder-decoder model as the first classifier to learn sequential features of student behaviour. The inputs of the top and bottom encoders\(x = [ {{x}_1,{x}_2, \ldots ,{x}_{t - 1},{x}_t} ]\) are the same data of students' historical behaviours (e.g., x₁ for "cafeteria" and x_t for "shower"), while the outputs are the implicit representations of the students' sequential behavioural features h_i . Notice that\(h_t^g\) and \(h_t^l\) play different roles, but they have the same value. The upper encoder encodes the entire initial sequence information through the last hidden layer state\(h_t^g\), while the lower encoder computes the attentional weights with the previous hidden states through the last hidden state\(h_t^l\), which then unifies the input and output high-dimensional vectors.

The encoder of the basic sequence is connected to the output of the attention-based sequence encoder to be fed to the sequence feature generator. Both use GRU as the basic network unit. This is because the GRU not only maintains important features in short-term propagation, but also solves the gradient vanishing problem well. The specific steps of the algorithm are, for each behavioural sequence xi, the GRU takes the historical campus behavioural sequences as inputs, and the output is a linear transformation between the previous implied state ht-1 and the candidate implied state\(h_t^{'}\) . This process can be expressed as:

Where ht-1 and \(h_t^{'}\) represent the previous hidden state and current hidden state, respectively. Equations 3-1 to 3-4 represent update gates, reset gates, new network units and hidden states, respectively. In particular, the update gate zt controls how much information from the previously hidden state ht-1 needs to be forgotten and how much information from the currently hidden state\(h_t^{'}\) needs to be remembered. The reset gate rt determines how much previously memorised information needs to be retained. The process interpolates linearly between the existing hidden states and the current hidden state, and the final hidden state of the encoder carries the information of the entire initial sequence. The final hidden state ht (later denoted as\(h_t^g\) ) is actually used as a representation of the student's behavioural sequence features, i.e. the basic sequence encoder.

Noting that not all students' behaviours are related to their academic performance, the prediction is made with the hope that the SPC model will pay more attention to behavioural interactions that are related to their achievement performance. Therefore, this paper proposes a DEVICE-level attention mechanism to model this assumption, called Attention-Based Sequence Encoder.

Where \(\sigma\) is the sigmoid function which converts ht and hi into potential space via a matrix. This function does not directly add up all the hidden states learned by the RNN network to represent the sequence of student behaviours. Rather, a weighting factor\({\alpha }_{ti}\) is used to indicate which hidden state/interaction is important to the encoder. Then, the weighted sum of the above hidden states is computed to represent the student's behavioural sequence features. To better understand \(c_t^l\), hi can also be represented as the last hidden state hi (i.e.,\(h_t^l\) ) at time step t. Therefore,\(c_t^l\) is modified as,

The \(h_t^g\)are incorporated into\(c_t^g\), and \(h_t^l\) and \({\alpha }_{ti}\) are incorporated into\(c_t^l\), which together provide a behavioural sequence representation for the SPC model. Specifically, the last hidden state of the sequence-based encoder\(h_t^g\) is responsible for encoding the entire sequence behaviour, while the attention-based sequence encoder\(h_t^l\) is responsible for computing the attention weights of the previous hidden state. With this hybrid scheme, the basic sequence encoder and the attention-based sequence encoder can eventually be modelled as a unified representation ct, i.e., the sequence feature generator, which is represented as a splice of the vectors\(c_t^g\) and \(c_t^l\),

For better student performance prediction, this paper applies a bilinear decoding mechanism with selectivity between the current sequence implicit representation and each campus card device to compute the similarity score Si .

Where T is a\(| D | * | H |\) matrix,\(| D |\) denotes the dimension of each campus card device embedding for mapping each behavioural vector to a low-dimensional space, and \(| H |\) is the dimension of the sequence representation. The similarity score of each campus card end device is then fed to the softmax layer to obtain the probability of a deep behavioural feature (denoted as F afterwards) for model decoding. For the sequence-based prediction task, the basic sequence encoder summarises the behaviours produced by the whole student each week, whereas the attention-based sequence encoder can adaptively select relevant behaviours to capture the student's main intention, which can focus on the most recent actions performed by the student. Therefore, a representation of the sequence behaviours and the previous hidden state is used to calculate the attentional weights for each occurring behaviour. The sequence behavioural features are then combined with the student's main intention features to form an extended representation of each timestamp, which allows for the learning of deeply periodic sequence features from the student's behaviour.

In the linearly separable case, the support vector classifier tries to find an optimal classification hyperplane\({W}^T \cdot x + b = 0\) to maximise the separation interval. To find this hyperplane, the following quadratic programming problem needs to be solved,

Where w is the normal vector, b is the bias term, and x is denoted as the identity. The solution of the above quadratic programming problem by the Lagrangian duality,

This equation is the original problem, and its dual problem is obtained from the differential formula and the simplification of the relationship between w and \(\alpha\), b and \(\alpha\) as,

The solution steps are similar to the linearly divisible case, where \(C\sum\limits_{}^l {{\xi }_i}\) is the penalty term. The classification hyperplane at this point is. \(f(x) = \sum\limits_{}^l {\alpha _i^ * } {y}_iK( {{x}_i,{x}_j} ) + {b}^ *\) The nature of the kernel function corresponding to the inner product of the high-dimensional space is exploited to enable the linear classifier to implicitly build the classification plane of the high-dimensional space. The support vector machine constructs the optimal segmentation hyperplane in the feature space based on the theory of structural risk minimisation, which allows the learner to be globally optimised and the expected risk of the entire sample space to satisfy some upper bound with a certain probability.

In solving the multi-classification problem, this paper chooses a one-to-one approach, i.e., classifying each two categories separately.The input to the SVM is the complete sequence of behavioural features formed in the HRNN network model\(F = [ {{f}_1,{f}_2,...,{f}_{t - 1},{f}_t,} ]\), and the output is the real student's performance ranking classification\(y \in \{ {1,2,3} \}\) . Let\(( {{f}_i,{y}_i} )\) be the sample set and y denote the category. The academic performance of the students is classified into three main categories, students belonging to the first category A rank account for about 20% of the total number of students, students belonging to the second category B rank account for about 60% of the total number of people, and the last category C rank accounts for about 20% of the total number of people. Therefore, the method can be used to consider the prediction of students' academic performance as a short-term sequence modelling problem. If weekly behavioural data of a target student is provided, the student's academic performance can be predicted based on the SPC two-stage classifier in order to identify students with learning crisis in time.