Cluster analysis is a data mining technique that maximises object similarity within each group and minimises object similarity between groups to explore patterns and structures in datasets, aiding in the classification or grouping of data and facilitating decision-making (Zhang & Zhang, 2023).

K-Means clustering is an iterative algorithm (Mao et al., 2021). The main concept is to separate the data into k groups by first choosing k objects at random to serve as the initial centroids for clustering. Next, each object is assigned to the closest cluster based on the distance to each seed cluster centre. The cluster centre will be recalculated using the objects that are currently in the cluster each time a sample is assigned.

Assume that the initial dataset is D = {X1, X2, ..., Xn}, where Xi, Xj are the data objects, and Xil is the “l” characteristic attribute of Xi(Fan et al., 2022). The definition of the Euclidean distance is shown in Equation 1.

The choice of k has a significant impact on the results from K-means clustering. In this study, the K-means using the silhouette score and elbow method are employed to select the k value based on the sum of squares of errors (SSE) (Fan et al., 2022). The idea is that SSE describes each cluster’s compactness to a certain extent. Therefore, if the SSE result is small, it will lead to better clustering, which is determined by Equation 2.

Regarding the application of the K-means results in library work to organise reading promotion activities or library services, Wang et al. (2022) classified users into four categories. They identified which classification easily utilises the information resources and which one hardly does, along with which clusters have longer loan terms. Guzman et al. (2025) classified the library’s collection of books to demonstrate which ones are frequently borrowed, sometimes borrowed, and seldom borrowed.

There are several types of algorithms for recommending books to customers of a library, and the most used approach is association rules-based recommendations (Monika et al., 2025). The recommendation algorithm-based, on association rules, focuses on mining association rules between borrowed books from the library database. The general mining process can be divided into two steps: searching for frequent sets in the database and searching for strict rules within these frequent sets (Duchen, 2022).

Association rules are logical rules developed by data mining. Agrawal et al. (1993) proposed them to reveal potential relationships between data items. The concept of support is fundamental to understanding association rules. In the context of book borrowing, using literature books and computer books as examples, the support of a rule connecting these two indicates the probability that literature books and computer books will appear simultaneously throughout the entire data item set, while the confidence indicates the probability that literature books will be borrowed after computer books have been borrowed (Zhou, 2022).The basic steps of mining the association rule between borrowed books and the association rule-based recommendation algorithm are shown in Figure 1.

In Figure 1, the database of book borrowing records is simplified for illustrative purposes. The book titles are represented by letters, and the record number is reduced to four users and five books.

The flow steps are as follows:

For example, take the item set {B, C} in Figure 1. The possible strong rules are {B}⇒{C} and {C}⇒{B} (Duchen, 2022). The confidence degree for calculating a strong rule X⇒Y is calculated according to Equation 4.

Where Confidence is the confidence degree of the strong rule, |X ꓵ Y| is the number of records that have both X and Y, and |X| is the number of records that have X. When the confidence degree exceeds a certain threshold, the association rule is considered a strong rule. The recommendation algorithm provides book recommendations based on the user’s book borrowing records and strong rules (Duchen, 2022).

Library management decision-making is supported by association rules. Liu (2018) developed a library management system based on association rules and described the modules required for mining, and used association rule mining to identify relationships among the most borrowed books each semester. They also used association rule mining to examine the books borrowed by users from the same group.

The Apriori algorithm was the first association rule mining method to utilize support-based pruning to manage the exponential increase in candidate item sets (Mariana et al., 2017). Support-based pruning involves removing item sets that are infrequent in the dataset, specifically eliminating any (k + 1)-length item set if its k-length subset is infrequent. This method also introduces the apriori principle (Mohanty et al., 2023), which states that every non-empty subset of a frequent item set must also be frequent. The algorithm helps find all frequent item sets that commonly appear in a database. Data processing was performed across various minimum support thresholds until frequent k-item sets with k of at least two were identified. The apriori algorithm is widely used to analyse customer behaviours for product and service recommendations across different domains. For example, it is applied in business to analyse social media and online shopping habits (Liao et al., 2021) and to understand the relationship between customers and products (Qisman et al., 2021). This data is then used to predict trends and guide marketing strategies. In education, the algorithm has been employed to analyse borrowing behaviours to recommend resources aligned with user interests (Zhou, 2020). It is also utilised in health science applications to analyse patterns within health-related data and to inform recommendations for the prevention of dengue (Kajornkasirat et al., 2025).

The Frequent Pattern Growth (FP-Growth) algorithm is an improvement over the apriori algorithm in terms of execution speed, thereby addressing the shortcomings of the apriori algorithm. FP-Growth is one alternative that can be used to identify the most frequently occurring patterns (frequent item sets) within a dataset (Zhang & Zhang, 2023). The apriori algorithm needs to generate candidates to obtain a frequent item set. However, for the FP-Growth algorithm, generating candidates is not done because FP-Growth uses the concept of a development tree in search of a frequent item set. Characteristics of the FP-Growth algorithm are that the data structure used is a tree called the FP-Tree (Lakshika & Caldera, 2022). The FP-Growth algorithm can extract a frequent item set directly from the FP-Tree. The extraction of frequent item sets using the FP-Growth algorithm is performed by constructing a data structure called an FP-Tree (Lael & Pramudito, 2023).

Library automation employs information and communications technologies (ICT) to reduce manual processing in libraries, saving librarians significant time and effort. Integrating recommendation systems into these automated environments improves user satisfaction while delivering scalable efficiencies for large user bases. Researchers and librarians in academic settings have examined such systems, particularly their application to the OPAC module, which serves as the primary public interface for searching and retrieving resources (Puritat & Intawong, 2020; Puritat et al., 2021; Ikwuanusi et al et al., 2024). Open-source platforms such as Koha, Evergreen, and OpenBiblio support modular integration of advanced algorithms directly into bibliographic databases (Mamoriya, 2021). This flexibility enables the use of collaborative filtering and content-based techniques to deliver tailored discovery services aligned with academic research patterns (Alaoui et al., 2024). Numerous studies apply association rule mining to circulation logs, identifying recurring borrowing behaviours to predict and recommend relevant titles (Bhajantri et al., 2024; Xie & Liu, 2025). These approaches collectively advance personalised library services.

Recent studies indicate that K-Means Clustering is effective in classifying users according to their service usage behaviour, borrowing frequency, and specific interests. Meanwhile, association rule mining, along with the FP-Growth algorithm, has proven to be a valuable approach for identifying patterns of relationships among books and co-occurring borrowing behaviours. Integrating techniques with data from an automated library system can significantly enhance the accuracy of book recommendation systems. This advancement enables users to more efficiently discover books that align with their interests;, improves access to pertinent information resources;, and enhances the overall quality of library services.

The dataset used in this study was obtained from the daily borrowing behaviour records of the Library of Prince of Songkla University, Pattani Campus, through the Automated Library System for Thai Higher Education Institutes (ALIST). This dataset covers the period from January 1, 2020, to June 30, 2025.The data include as variables daily borrowing, number of book borrowing, book data, book title, call number, book category, book collection, and number of copies.

This study followed established ethical principles and data governance frameworks by employing thorough anonymisation and de-identification procedures. All personally identifiable information was permanently removed before data processing and analysis. Additionally, techniques such as pseudonymisation, data masking, and tokenisation were utilised to reduce the risk of re-identification. The data were protected through encryption, strict access controls, and secure storage systems. There was no intention to identify individuals or infringe on library users’ rights. These measures ensure the protection of participant confidentiality and compliance with relevant data protection regulations and institutional review requirements.

After collecting the data, the next step was data pre-processing, which involved data cleaning aimed at removing unused or duplicate data during the process of building a recommendation model. Additionally, to address data inconsistencies, various issues were identified that required correction, such as inconsistent data types and missing data, which were filled in with null values.

Then we identified the association in the book data, which includes a book collection and book classification, based on the book borrowing records. This stage produced four variables that serve as inputs for K-Means clustering.

Cluster analysis is defined as the process of grouping a set of similar objects into clusters (Bhatia, 2019). The main idea of clustering is to try to divide instances into clusters so that instances within a cluster are highly similar but with high dissimilarity between clusters (Bhatia, 2019; Silwattananusarn et al., 2020). , The K-Means clustering algorithm was used to analyse user characteristics. In the K-Means clustering algorithm, objects are grouped into k clusters based on their attributes or features.

After data was pre-processed, we checked for recent transactions, frequency and the amount spent by the customers. To create a recency variable, decide the reference date, that is, one day prior to the last transaction. RFM (recency, frequency, monetary) analysis is a very popular customer segmentation and identification technique in database marketing (Christy et al., 2018; Patil & Anita, 2022). RFM has been applied to analyse the borrowing behaviour of users as follows: Recency (R) implies the number of days since the last borrowing records; Frequency (F) refers to the number of times or records that users make when borrowing books; Monetary (M) is defined as the total quantity of books borrowed.

K-Means algorithm is applied using the Euclidean distance metric to partition the users for RFM values. K-Means is used twice to analyse the amount obtained for Recent and Frequent transactions as mentioned below: i) To partition the customers based on the amount generated with recent transactions, and ii) To group the customers by the amount generated with frequent transactions. Clusters obtained in RFM analysis are evaluated using the silhouette score and elbow method (Cui, 2020; W3Schools, 2025), which analyses show well the resulting clusters are separated based on their book borrowing records. The elbow method analysis results combined with silhouette score values are compared for k values to identify the optimal number of clusters in this study. For each cluster, a book recommendation model is developed using the association rules from the FP-Growth algorithm.

Association rule mining is a commonly used data mining technique to explore the relationships between items in large data sets. The rule has the form ‘cause’ ⇒ ‘consequence’, which means “if the cause is satisfied, the consequence is likely to also occur”. In this study, association rules are used to recommend books. For example, the association rule ‘book 1’, ‘book 2’ ⇒ ‘book 3’ means that if a reader reads book 1 and book 2, they are likely to read book 3 as well. Rules like this can be used for book recommendations (Zhang et al., 2024). The FP-Growth algorithm is used in order to convert the raw borrowing logs into useful association rules; by setting strict thresholds for support, confidence, and lift, the system eliminates insignificant patterns so as to focus on the most important associations, thus enabling the identification of new titles that are tailored to individual user behaviours Books are recommended based on a set of highly prioritised criteria, ensuring that users are introduced to new titles that they have not previously read.

A confusion matrix is used to summarise the performance of the classification model. The Confusion matrix holds Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN) counts, which will be used to calculate precision and recall scores, as shown in Table 1.

The metrics used to assess the performance of the research method includes Precision, Recall, and F-measure (Sundari &Subaji, 2022). The F-measure is defined as the harmonic average of precision (P) and recall (R).

In this study, the hit rate is a binary metric that indicates whether relevant items are included in a user’s recommendation list (Annisa et al., 2024). The higher the hit rate, the better the quality of the recommendations generated. The hit rate indicates that developing a high-quality recommendation system requires each recommendation target to receive at least one relevant item aligned with their interests (Annisa et al., 2024). The hit rate is defined by Equation 8.

Figure 3 shows the distribution of book borrowing events across Dewey Decimal Classification (DDC) classes from 2020 to 2024, showing both annual totals and the proportion of each category. Overall, book borrowing fluctuated slightly over the five years. The class with the highest number of borrowing was Languages [400], followed by History & Geography [900] and Science [500], reflecting the popularity of media in linguistics, history, and science among users. In contrast, the categories Philosophy and Psychology [100] and Religions [200] saw low and relatively stable borrowing, while the category Technologies [600] showed an increasing trend after 2022, indicating a growing interest in technology and its applications. The results from this graph highlight trends and differences in book borrowing behaviour, which can be used to effectively plan the development of information resources to meet future user needs.

The results of the trained algorithm are then used to predict clusters in the test data. To find the most appropriate number of clusters from the results of K-means, the variance of the best cluster is searched using the silhouette score and the elbow method.

Figure 4 shows the relationship between the number of clusters (k) and the distortion score obtained from the K-Means clustering process. The left y-axis shows the distortion score, which is the sum of the squares of the distances between each data point and the centre of that cluster. The distortion score decreases rapidly as the number of clusters increases from 1 to 3, and after that, the rate of decrease begins to slow down. This turning point, or “elbow point,” indicates that the optimal number of clusters is k = 3, which gives a distortion score of approximately 5, 382.175. This suggests that dividing the library user data into three groups is the most appropriate grouping in terms of model accuracy and simplicity.

In addition, the right y-axis shows the model’s fit time, which changes slightly as the number of clusters increases, reflecting the variation in the computation time of each clustering round. The elbow method analysis results combined with silhouette score values confirm that k = 3 is the most appropriate number of clusters for library user behaviour.

From Table 2 on the performance metrics of library user groupings using K-Means with the number of clusters equal to, three, six or ten, it is seen that all four metrics, namely precision, recall, F-measure and hit rate, were at a high level, indicating accuracy and stability of the model used to cluster the data. However, the model with three clusters was the most appropriate choice, with precision at 82.98%, recall at 98.55%, F-measure at 0.8815, and hit rate at 97.57%.

Based on the results of the user clustering, the researchers defined the behaviour of the three user groups using RFM, (Recency, Frequency, and Monetary) analysis, as shown in Figure 5.

Figure 5 (a), 5(b) and 5(c) illustrates the behaviour of each cluster as follows: Cluster 1 is defined as dormant, low-activity users. This group reflects users with the lowest level of library engagement across all dimensions, including both borrowing frequency and volume. They also have the longest gap between their last borrowing visits compared to other groups, indicating a tendency towards churn or being one-time users. ; Cluster 2 is defined as loyal high-value users. This group represents the highest quality target group, exhibiting significantly higher borrowing frequency and volume of resources borrowed compared to other groups (High Frequency & Monetary). Furthermore, they have the lowest Recency, indicating consistent return visits, reflecting their engagement and satisfaction with library services. Cluster 3 is defined as drifting moderate users. This group has a moderate borrowing history, neither too high nor too low, but their Recency is similar to Cluster 1, indicating that although they used the service moderately in the past, recent usage has become less frequent, and they may be heading towards cessation of service if not stimulated.

Association rule mining was used to identify patterns in book borrowing within each user cluster, with a minimum support value of 0.15and a minimum confidence value of 0.90.The results from all three clusters were similar, demonstrating a highly consistent pattern of book borrowing behaviour across the clusters, as shown in Table 3.

The analysis demonstrates significant interdisciplinary borrowing patterns among library users (Table 3).The association rules identified thirteen significant rules meeting the criteria for Support, Confidence, Lift, and Conviction.

Strongest relationship. Rule 1: Economics [330], Social Sciences [300], History of Europe [940] ⇒ History of Asia [950] has a confidence of 100%. This means that whenever a book from each of the premises has been borrowed, the user also borrows one from History of Asia. Although the Support is only 1.6%, this perfect Confidence indicates a strong and consistent connection in historical data. This rule may reflect a group of researchers or students studying social science and economics topics in a comparative context between Europe and Asia.

Relationship between disciplines. Economics & History: There is a strong relationship with Asian History (see Rules 1, 3, 8, and 13 in Table 3), indicating that those interested in Economics are also likely to be interested in the History of Asia. Science-related subjects (Life Sciences, Physics, and Chemistry) in rules 2, 6, and 7demonstrate the relationship between these subjects. For example, those who borrow Life Sciences and Biology books are likely to also borrow Chemistry, which is very reasonable because these subjects are related. Political Science, Public Administration, and Social Sciences show strong connections with History of Europe, History of Asia, and Law (see Rules 5, 9, 10, and 12), indicating that those interested in politics and government tend to study history and law simultaneously.

The association rule results in Table 3, showing the relationships among library users’ book borrowings, are presented as a knowledge graph, as shown in Figure 6.