Resource Classification and Knowledge Aggregation of Library and Information Based on Data Mining

Since it was conceptualized in 2008, big data has become a hot topic in the academia. In the meantime, data mining has been increasingly applied in various industries [1-3]. In particular, the application of data mining in library and information (L&I) attracts much attention from experts and scholars [4-6]. With the help of data mining, researchers have optimized the aggregation and retrieval of massive L&I data, and acquired better capability to retrieve, identify, and make intelligent analysis of such data. Hence, data mining brings new opportunities to the informatization and intellectualization of L&I management system.

Traditionally, L&I resources are classified based on access control and optimal configuration [7-9]. Raflesia et al. [10] extracted and vectorized the attributes of L&I resources, in the light of the text documents about these attributes. Based on support vector machine (SVM) classification algorithm, Antoniy et al. [11] established an automatic classification model for L&I resources, integrated the sequential minimal optimization (SMO) to effectively improve the classification efficiency, and optimized the classification effect through grid search of the optimal algorithm parameters. Using the real-time information of the L&I resource set during the update, Losee [12] constructed a resource classification model, and verified its feasibility and effectiveness through experiments on multi-source L&I resource data. After exploring deep into the unified management of L&I resources, Tella et al. [13] highlighted the importance of resource management to real-time L&I resource classification, and put forward clear standards for resource classification, principles for differentiating between new and old resources, and effective measures to link up the two kinds of resources; in addition, an L&I resource classification system was developed for the unified management of L&I resources, including 4 A-level classes, 12 B-level classes, and 25 C-level classes. Considering the similarity between same-class L&I resources in content, theme, and features, Jerrett et al. [14] constructed a thematic L&I resource classification model based on long short-term memory (LSTM) network, and demonstrated the superiority and feasibility of the model through experiments on the CNKI database for the Belt and Road Initiative (BRI).

The information knowledge generated from L&I resources face several problems: the knowledge points are scattered and fragmented, the quality is uneven, and the contents are complex and redundant. In addition, there is a lack of direct channels between multi-source L&I resource platforms. It is time-consuming to browse and acquire knowledge on multiple platforms [15-17]. Many scholars have explored the ways to aggregate the knowledge in L&I resources, aiming to scientifically organize, mine, and manage the knowledge, and to innovate the knowledge service model [18-21]. For example, Kankonsue et al. [22] defined the connotations of knowledge aggregation of multi-source L&I resources, effectively organized the knowledge contained in L&I resources, and mined the associations between the knowledge. Borrego [23] proposed a knowledge aggregation strategy based on topic-generated multi-source L&I resources: the topic probability model of latent Dirichlet allocation (LDA) was combined with the hybrid neural network BiLSTM-CNN-CRF (bidirectional LSTM-convolutional neural network-conditional random field) to learn and segment the contents, and to generate knowledge topics. Kalenov et al. [24] produced knowledge summaries of multiple L&I resources, using maximal marginal relevance (MMR) algorithm and word2vec model. After mining user interests, Ammar et al. [25] provided a knowledge aggregation and accurate recommendation strategy for multi-source L&I resources, and calculated the user similarity between multi-source L&I resource platforms, creating a robust user network.

Big data technology makes it possible to classify the resources and optimize the configuration of L&I resource platforms according to user demand, and unify the nonuniform data structures (structured, semi-structured, or non-structured) of traditional knowledge service systems. With the aid of data mining, this paper introduces the new information service model of big data resources and knowledge services to the processing of L&I data. Firstly, the keywords of resource retrieval were sampled and integrated based on the data storage structure and relationship model of the L&I resource platform. Next, an SVM-based L&I resource classification model was constructed to extract and quantify the attributes of the keywords of resource retrieval. Then, a knowledge aggregation model was developed for a complex network of multiple L&I resource platforms, and proved effective through experiments.

Figure 3 explains the workflow of L&I resource classification. The rapid development of information technology (IT) has diversified the types and structure of L&I resources. Therefore, the keywords of L&I data need to be defined uniformly, according to the shared features of L&I resources, and prepared into a standardized description template.

3.jpg

Figure 3. The workflow of L&I resource classification

Figure 4 shows the formal description model of L&I resources. The model mainly consists of a basic information description module, an online information description module, an implicit knowledge and utility description module, and a retrieval state description module.

4.jpg

Figure 4. The formal description model of L&I resources

Based on the integration of keywords, an extensible markup language (XML) file of the L&I resources can be formulated from the information provided by the description modules. The term frequency–inverse document frequency (TF-IDF) model was adopted to quantify the retrieval keywords of L&I resources. The word frequency in the TF-IDF model can be expressed as:

$W F(w, K)=\frac{A F(w, K)}{\max A F(K)}$  (9)

where, WF(w, K), AF(w, K), and max AF(K) are the normalized frequency, absolute frequency, and peak frequency of keyword w in the semantic information set K of the L&I resources, respectively. The anti-document frequency can be expressed as:

$A D F(w)=\log \frac{N R}{N_{i}+1}$  (10)

where, NR is the number of L&I resources; n_i is the number of resources containing keyword i. From formulas (9) and (10), the weight coefficient of each keyword can be quantified by:

$W E I(w, K)=W F(w, K) \times A D F(w)$  (11)

SVM is a generalized linear classifier with strong generalization ability. This classifier can learn the features of input data, while minimizing structural risks. Considering the advantages of SVM in data mining, this paper applies SVM to classify L&I resources.

In the objective function of SVM, the characteristic parameters have the same variance and the same mean (zero). Therefore, any single feature that does not obey standard normal distribution might dominate the objective function, causing errors in the classification results. To solve the problem, the eigenvalues should be normalized by:

$W E I_{N O R}=\frac{W E I-\overline{W E I}}{W E I_{v a r}}$  (12)

where, the numerator is the difference between WEI and its mean; WEI_var is the variance of WEI.

For the SVM, the keyword set of the L&I resources in the feature space can be described as E={(a₁,b₁),(a₂, b₂),…,(a_N,b_N)}, where a_i is an r-dimensional vector in R^r, and b_i is the class tag (b_i=1, or -1). Then, the classification hyperplane of the feature space can be expressed as:

$b(a)=\phi^{T} a_{i}+c$   (13)

where, c is a constant; φ is an r-dimensional vector. The distance between a data point in the feature space to the hyperplane can be described by the function interval b_ii((φ^Ta_i+c)/ǁφǁ). Under the premise of maximizing the interval, the search for the optimal hyperplane can be transformed into the optimization of the following constraint:

$\min _{\phi, c} \frac{1}{2}\|\phi\|^{2} s . t . b_{i}\left(\phi^{T} a_{i}+c\right)-1 \geq 0$  (14)

In the real world, some samples cannot be classified by linear classifiers. That is, the distance of some points to the hyperplane is smaller than 1. Thus, a nonnegative slack variable γ_i was introduced to make b_ii(φ^Ta_i+c)≥1-γ_i. Adding a penalty to γ_i, the objective function can be transformed into:

$\frac{1}{2\|\phi\|^{2}}+P \sum_{i=1}^{N} \gamma_{i}$  (15)

where, P is the penalty function (P>0). The greater the P value, the stricter the penalty on misclassification. Formula (15) aims to maximize the interval, i.e. minimizing ǁφǁ, while minimizing the number of misclassified points. Hence, the constraint optimization problem can be rewritten as:

$\min _{\phi, c} \frac{1}{2}\|\phi\|^{2}+P \sum_{i=1}^{N} \gamma_{i}$ s.t. $b_{i}\left(\phi^{T} a_{i}+c\right)-1$$\geq 0$ and $\gamma_{i} \geq 0$  (16)

The optimal solution of the original optimization problem can be obtained by solving the dual problem in the above formula. Then, a positive Lagrangian multiplier τ_i, υ_i≥0 was introduced to the above inequality. The Lagrangian function can be defined as:

$L(\phi, c, \gamma, \tau, v)=\frac{1}{2}|| \phi||^{2}$$+P \sum_{i=1}^{N} \gamma_{i}$$-\sum_{i=1}^{N} \tau_{i}\left(b_{i}\left(\phi^{T} a_{i}+c\right)-1+\gamma_{i}\right)$$-\sum_{i=1}^{N} v_{i} \gamma_{i}$  (17)

To obtain the optimal solution to the original problem, the feasible solution of the dual problem needs to satisfy the Karuch-Kuhn-Tucker (KKT) conditions. That is, solving the minimum of the Lagrangian function L relative to φ, c, and γ:

$\left\{\begin{array}{c}\min _{\phi} L(\phi, c, \gamma, \tau, v)=\phi-\sum_{i=1}^{N} \tau_{i} b_{i} a_{i}=0 \\ \min _{c} L(\phi, c, \gamma, \tau, v)=-\sum_{i=1}^{N} \tau_{i} b_{i}=0 \\ \min _{\gamma} L(\phi, c, \gamma, \tau, v)=P-\tau_{i}-v_{i}=0\end{array}\right.$  (18)

The results of formula (18) were simplified and substituted into formula (17). Then, the dual problem can be obtained by maximizing τ value:

$\max _{\tau} \frac{1}{2} \sum_{i=1}^{N} \sum_{j=1}^{N} \tau_{i} \tau_{j} b_{i} b_{j}\left(a_{i} \cdot a_{j}\right)-\sum_{i=1}^{N} \tau_{i}$ s.t. $\sum_{i=1}^{N} \tau_{i} b_{i}$$=0 \quad 0 \leq \tau_{i} \leq P$   (19)

The classification of multi-source L&I resources is a hard nonlinear problem. Here, the nonlinear problem is converted into a linear problem with the Gaussian kernel function:

$G\left(a, a^{\prime}\right)=e^{-\frac{|| a-a^{\prime}||^{2}}{2 \sigma^{2}}}$  (20)

To reduce the heterogeneity of multi-source L&I resources, this paper uses the one-against-one method to set up a binary classifier between any L&I resource samples, thereby building up a multi-class SVM, which outputs the classification decision function based on the input: keyword set of L&I resources E={(a₁,b₁),(a₂,b₂),…,(a_N, b_N)}. The penalty coefficient P and bandwidth σ were selected rationally, and introduced to formula (20). Then, the optimal solution τ^*=(τ^*₁,τ^*₂,…τ^*_N) to the optimization problem can be obtained by:

$\max _{\tau} \frac{1}{2} \sum_{i=1}^{N} \sum_{j=1}^{N} \tau_{i} \tau_{j} b_{i} b_{j} e^{-\frac{|| a_{i}-a_{j}||^{2}}{2 \sigma^{2}}}$$-\sum_{i=1}^{N} \tau_{i} s . t . \sum_{i=1}^{N} \tau_{i} b_{i}=0$  $0 \leq \tau_{i}\leq P$  (21)

Taking a positive component τ^*_j of the optimal solution τ* that is smaller than P, the following formula can be calculated:

$c^{*}=b_{j}-\sum_{i=1}^{N} \tau_{i}^{*} b_{j} e^{-\frac{|| a_{i}-a_{j}||^{2}}{2 \sigma^{2}}}$  (22)

The classification decision function can be established as:

$C D F(x)=\operatorname{sign}\left(\sum_{i=1}^{N} \tau_{i}^{*} b_{j} e^{-\frac{|| a_{i}-a_{j}||^{2}}{2 \sigma^{2}}}+c^{*}\right)$  (23)

The following experiments were conducted to test the performance of our method in the collection and aggregation of L&I data. The algorithms were programmed on MATLAB in C++. The RFID tag conversion accuracy was assumed to be 36 bits. The training set and test set include 3,000 and 500 keyword attributes of L&I resources, respectively. The correlation coefficient between different types of keyword attributes was set to 0.25. The context matching degree was set to 0.61.

Figure 6 compares the recalls of the retrieval keywords of L&I resources aggregated by different methods, including our method, time series reconstruction and adaptive balanced retrieval (TSC-ATR), and adaptive screening of distributed structure (ASDS). It can be seen that our method achieved higher accuracy in keyword retrieval of L&I resources than the other two methods. The superior retrieval accuracy comes from the aggregation of high-precision collected data by the RFID tag recognition model based on keyword attributes.

6.png

Figure 6. The comparison of the recall

To verify its feasibility and effectiveness, the SVM-based L&I resource classification model was programmed with Spyder compiler. The resource dataset was randomly divided into a training set and a test set. The classification effect of the proposed model was compared with that of mainstream algorithms, namely, the k-nearest neighbors (k-NN) rough set algorithm, the k-modes k-NN, the kernel SVM, and the multi-class SVM (Figure 7; Table 1). The common metrics like accuracy, recall, F1-score, and receiver operating characteristic (ROC) curve were selected to measure the classification effect. It can be seen that, facing the multi-class, high-dimensional L&I resources, our classification model achieved an accuracy of 84.21%, a recall of 85.61%, an F1-score of 86.71%, and a test value of 83.92%. These results are much better than those of other algorithms.

Table 1. The classification effects of different methods

7.png

Figure 7. The classification results of different methods

Figure 8 presents the learning curves of the proposed SVM-based classification model in the training test and cross-validation test. It can be seen that the training test value was relatively stable, while the cross-validation test value tended to be stable with the growing number of samples, indicating that the learning effect of the model gradually improves. This proves the overall good classification effect of our model.

8.png

Figure 8. The learning curves of our model

9.png

Figure 9. The ROCs of our model

To disclose the classification effect of our model on each type of L&I resources, the ROCs of our model correctly classifying 100%, 99%, 85%, and 80% of test samples, and the mean ROC are plotted as Figure 9. It can be seen that the mean classification accuracy of our model surpassed 92%, suggesting that our model boasts a good classification effect.

10a.png

(a)

10b.png

(b)

Figure 10. The relationship between knowledge aggregation performance and the number of platform nodes

The final task is to verify the effectiveness of the proposed knowledge aggregation model in the complex network of multiple L&I resource platforms. The node strength and mean path length between nodes in global and local networks were tested under the multi-polar knowledge interaction model. Figure 10(a) provides the curve between the number of local network nodes and node strength, and Figure 10(b) displays the curve between the mean path length of the global network and node strength. It can be seen that the node strength of our model obeys the power-law distribution, reflecting the features of node strength distribution of complex networks. This means our knowledge aggregation model has certain credibility. In addition, it was learned that the mean path length slowly increased and then grew linearly, with the growing number of nodes, indicating the knowledge aggregation model adapts to the small-world features of ultrashort mean path length.