Research on the optimisation of music education curriculum content and implementation path based on big data analysis

Data mining has unknown potential and value. The mining process is generally divided into four stages, including data selection, preprocessing, data mining, and data analysis and evaluation. The data mining model can be seen in Figure 2. 1)

Data Selection

Data selection, also known as data integration, entails bringing together data from various operational scenarios. Data mining presupposes the availability of a large amount of rich data, which can be obtained from either an online processing system or an offline data warehouse. This step is to identify the data set to be analyzed, reduce the scope of data processing, and improve the efficiency of the process.

Figure 2.

Data mining model diagram

Data mining is aimed at mining interesting, useful and implicit information, with the task of finding human-acceptable patterns to describe the data, using some variables to predict the future values of unknown or other variables, to find the knowledge that can be used from a large amount of data that is large, varied and complex, different mining tasks can be analysed using different algorithms. 1)

Description

Algorithms such as clustering, discovery of association rules, discovery of sequence patterns, regression, and bias detection are generally used.

HDFS is a Hadoop distributed file system, which can be used to store more than a million size number of files, mainly using streams to access the data and batch processing of data. Its best feature is that it is more fault-tolerant and can be automatically recovered when the saved copy is lost, so it is currently widely used mainly on commercial hardware. The HDFS architecture is shown in Figure 3. The HDFS file system is guaranteed to provide services by the following procedures, including NameNode is the master node and is the daemon running on the master node, and Secondary Namenode is the secondary node of NameNode.DataNode is the data node that runs on each slave node. 1)

NameNode

NameNode is mainly responsible for managing the directory structure of HDFS, including the number of file directories, file data block index, and file directory metadata information. The second is responsible for managing block files, and when reading data, it will contact the NameNode to obtain the data node where it is located, and then contact the DataNode to retrieve the data. The NameNode manages and controls how to decompose files into blocks, identifies the nodes that should be subordinated to store these blocks, and overall status and applicability of the distributed file system. The tasks performed by NameNode take a lot of memory and I/O.

Figure 3.

HDFS Architecture

MapReduce is a software framework for writing applications that can process as well as analyse large datasets in parallel and enables multi-node clustering through features such as reliability, scalability and fault tolerance [22]. It consists of two main phases, i.e. Map phase and the Reduce phase. 1)

Map

The Map phase involves analysing and extracting data, processing it into key-value pairs, and sending it out using parallel computing.

Figure 4.

MapReduce calculation process

The Node2vec algorithm models the use of a random wandering method with bias. Random wandering method with bias is used as a node information mining strategy for graphs, by adjusting hyperparameters to have preferred node information acquisition for the graph structure and generating a sequence of node’s nearest neighbours using the mined information [23].

The Node2vec algorithm uses a random wandering strategy as shown in Fig. 5, assuming that the current random wandering path passes through link (t, n) to node n, at which point the bias operator α_pq is computed from node n to node x: (1)αPq={1/pdΔx=01dss=11/qdΔx=2$${\alpha _{Pq}} = \left\{ {\begin{array}{*{20}{l}} {{1 /p}}&{{d_{\Delta x}} = 0} \\ 1&{{d_{ss}} = 1} \\ {{1 /q}}&{{d_{\Delta x}} = 2} \end{array}} \right.$$

Where d_tx is the shortest path between vertex t and vertex x.

Figure 5.

Schematic diagram of random walk strategy

The hyperparameter p controls the probability of repeated visits to the node just visited. From the above equation, it can be seen that p is involved in the computation only if d_tx is 0. If p is lower then the probability of visiting the vertex just visited becomes higher and vice versa. The hyperparameter q controls the direction of the wandering and determines whether the wandering is inward or outward. When q > 1, random wandering tends to visit vertices close to t. When q < 1, the random walk strategy tends to visit vertices farther away from t.

When the paranoid operator αpq(t,x)$${\alpha _{pq}}\left( {t,x} \right)$$ is obtained after passing through a certain edge (n,x)$$\left( {n,x} \right)$$, follow: (2)πnx=αpq(t,x)⋅wnx$${\pi _{nx}} = {\alpha _{pq}}\left( {t,x} \right) \cdot {w_{nx}}$$

Calculate the transfer probability π_nx of randomly wandering past an edge (t,n)$$\left( {t,n} \right)$$ and reaching node x. where w_nx is the weight of edge (n,x)$$\left( {n,x} \right)$$. Then, calculate the probability P(x|n)$$P\left( {x\left| n \right.} \right)$$ of the current node n, visiting the next vertex x, based on this transfer probability: (3)P(x|n)={πnx/Zif(n,x)∈E0otherwise$$P\left( {x\left| n \right.} \right) = \left\{ {\begin{array}{*{20}{l}} {{{{\pi _{nx}}} /Z}}&{if\left( {n,x} \right) \in E} \\ 0&{otherwise} \end{array}} \right.$$

Where Z is a normalised constant.

In summary, the random walk to obtain the node near-neighbour sequence of a total of n nodes, each node according to the random walk strategy (formulated by p, q) on the sampled node information is extracted. r then represents each need to carry out random walk r times, each time the final integration of each node after r times of random walk to obtain the node information constitutes the node sequence of the node.

In the Node2vec model, in order to calculate the conditional probability between words, each word is considered as two n-dimensional vectors. Assuming that the index of a word in the whole dictionary is i, when it is the centre word, its word vector is v_i ∈ R_N. When it is the context word, its word vector is u_i ∈ R_N. Assuming that the index of the centre word “GCN” in the corpus is c, and the index of the context word in the corpus is o, according to Softmax operation, we can get the conditional probability of the context word under the condition of the occurrence of “GCN”, and we can obtain the conditional probability of the context word under the condition of the occurrence of “GCN”. According to Softmax operation, we get the probability of the occurrence of the context word under the condition of “GCN”: (4)P(wo|wc)=exp(uoTvc)∑itexp(uiTvc) $$P\left( {{w_o}\left| {{w_c}} \right.} \right) = \frac{{\exp \left( {u_o^T{v_c}} \right)}}{{\sum\limits_{it} {\exp } \left( {u_i^T{v_c}} \right)}}$$

This is then converted into the form of a probabilistic concatenation: (5)P(wo)=∏i=IT∏−m≤j≤m,j=0P(w(i+j)|wt)$$P\left( {{w_o}} \right) = \prod\limits_{i = I}^T {\prod\limits_{ - m \leq j \leq m,j = 0} P } \left( {{w^{(i + j)}}\left| {{w^t}} \right.} \right)$$

Where T denotes the position of the window centre word and m denotes the size of the window. This allows the probability of inferring a background word for each centre word to be calculated.

The goal of Node2vec is then to find the background word with the highest probability of occurrence. Therefore, the next step is to consider only result maximisation. In this paper, the great likelihood method is used: (6)−∑t=IT∑−m≤j≤m,j=0logP(wt+1|w(i))$$ - \sum\limits_{t = I}^T {\sum\limits_{ - m \leq j \leq m,j = 0} {\log } } P\left( {{w^{t + 1}}\left| {{w^{(i)}}} \right.} \right)$$

The loss function needs to be minimised in order to make the gap between the predicted result and the true result as small as possible: (7)logP(wo|)=uoTvc−log(∑i∈Vexp(uiTvc))$$\log P\left( {{w_o}\left| {} \right.} \right) = u_o^T{v_c} - \log \left( {\sum\limits_{i \in V} {\exp } \left( {u_i^T{v_c}} \right)} \right)$$

Finally, in order to improve the efficiency of iterative solution in large-scale numerical matrices, as well as to address the inability of the least squares method to compute the unique optimal solution in the whole domain, the gradient descent method is used to perform the minimisation operation on the modified loss function: (8)∂logP(wo|wc)∂vc=uo−∑j∈VP(wj|wc)uj$$\frac{{\partial \log P\left( {{w_o}\left| {{w_c}} \right.} \right)}}{{\partial {v_c}}} = {u_o} - \sum\limits_{j \in V} P \left( {{w_j}\left| {{w_c}} \right.} \right){u_j}$$

In this way, the model parameters are iteratively updated to achieve optimisation.

The scenario for the use of the Node2vec model is shifted from contextual text prediction to mining the learning context for feature relationships between students and courses. In this context, its role is to calculate the probability of other nodes appearing under the condition that the target node appears.

After mining the feature relationships between students and courses, the feature relationships obtained from mining are used as a priori knowledge to improve the representation of learning ability. Therefore, in this paper, two types of (i.e., student and course) relationship graphs are constructed to mine patterns in a data-driven manner.

Their collaborative interactions can well reflect the information on the student side of the student relationship graph. Therefore, complex patterns can be learned by exploiting the collaborative interaction behaviour of users. Mathematically, there exists a boundary euvU,*$$e_{uv}^{U,*}$$ between student u and course v, additionally in this paper, we use Jaccard similarity calculation to represent the strength of correlation between students, which is calculated as follows: (9)wuvU,*=|eu*∩ev*||eu*∪ev*|$$w_{uv}^{U,*} = \frac{{\left| {e_u^* \cap e_v^*} \right|}}{{\left| {e_u^* \cup e_v^*} \right|}}$$

Unlike student relationships, courses have directed relationships due to the presence of temporal information. Therefore, a directed course relation graph is constructed to obtain a better representation of learning. Specifically, order the interactions of each student u into a sequence S_u based on the interaction timestamps, and count the number of times each pair of items from all sequences (i,j)$$\left( {i,j} \right)$$ appears in a particular order. Formally, under behaviour *iffCnt(i→j)>0$$*iff\;Cnt\left( {i \to j} \right) > 0$$, the course-relationship graph G^I has a boundary euvI,*$$e_{uv}^{I,*}$$ from course i to course j. where Cnt(⋅)$$Cnt\left( \cdot \right)$$ is a counter that calculates sequence strength using Jaccad similarity: (10)wuvI,*=Cnt(i→j)*Cnt(i→j)*+Cnt(j→i)*$$w_{uv}^{I,*} = \frac{{\mathop {Cnt(i \to j)}\limits^* }}{{\mathop {Cnt(i \to j)}\limits^* + \mathop {Cnt(j \to i)}\limits^* }}$$

The constructed student/course relationship diagram is shown in Figure 6.

Figure 6.

Student/course relationship

In order to verify the effectiveness of the Node2vec algorithm applied to music education course content recommendation after 308,225 student behaviour data training. In this paper, Pre@K, Recall@K, NDCG@K, MRR and AUC are used as evaluation metrics to assess the performance of the Node2vec algorithm on the test set, while K is set to 20 in order to have a more intuitive feeling of the gap between different models on the same metrics, and the experimental results are shown in Table 1. In the last row of the table, Node2vec algorithm comparisons are calculated for Matrix Factorization (MF), Heterogeneous Graph (HERec), Graph Convolutional Networks (NGCF), Contextual Information (ACKRec), Heterogeneous Information Networks (MOOCIR), Meta-Paths (HFCNqh), Knowledge Concepts (HFCNqk) and Behavioural Sequence Mining (HFCNqb) for a total of eight baselines The lift rate of the best performing model among the models on each evaluation metric. The following four conclusions can be drawn from the experimental results: 1)

Firstly, from the experimental results, it can be seen that MOOCIR, ACKRec model, HERec model and Node2vec algorithm outperform the MF algorithm in all the indicators, which indicates that the rich auxiliary data in the course recommender system characterised by heterogeneous graphs can be utilised to improve the recommender performance of the recommender system. Thus, it validates the importance of using heterogeneous graphs in recommender systems to mine entities and relationships between them.

Currently, one of the major problems facing course content recommendation systems is the problem of sparse student behaviour data. In this section, the performance of the experimental model when the data sparsity varies is discussed. The learners are divided into four groups (0,5], (5,15], (15,30] and (30,100) according to the number of interactions between learners and courses in the training set, and the division results are shown in Table 2. It can be seen that the first group contains the largest number of learners, and the last group contains the smallest number of learners, only 42 student users, indicating that most of the learners have studied courses concentrated in less than 5 courses, with fewer historical interaction behaviours, and the whole dataset is relatively sparse, which is more in line with the current status quo, and close to the reality.

Using accuracy and AUC as evaluation metrics, the performance of Node2vec, NGCF, and ACKRec algorithms on four sets of sparsity datasets is compared with the model HERec, and the enhancement rate of evaluation metrics based on the HERec algorithm is calculated for Node2vec, NGCF, and ACKRec algorithms and the result is visualised as a bar chart as shown in Figure 10. It can be seen that the Node2vec algorithm proposed in this paper achieves the best performance on all four sparsity datasets. Thus, it shows that the Node2vec algorithm can, to a certain extent, alleviate the problem of data sparsity.

Figure 10.

Model performance on data sets with different sparsity