Agent-Based Architecture for Developing Recommender System in Libraries

Abstract

The volume of data in terms of library resources is big and continuously increasing, therefore manipulation, searching, and suggesting these resources to user have become very difficult. The automation in libraries along with intelligence can solve the problem by easy and correct recommendation system. The intelligence can be incorporated into the system by implementing agent-based recommender system for libraries. An efficient intelligent agent-based library recommender system is developed to generate suggestions to end user by understanding his/her interest automatically by learning his/her profile. The manual recommendation process includes understanding the need of the user, solutions available; decision has to be made about which solution(s) can satisfy the user’s need and finally provide the solution(s) by filtering. The needs of user and available solutions are adaptive. It is clear that recommendation process involves the task of decision making and performing action based upon perception of recommender. The term software agent can be referred as a self-contained program capable of controlling its own decision making and acting, based on its perception of its environment, in pursuit of one or more objectives (Jennings and Wooldridge in IEE Rev 42(1):17–20, [1]). Therefore the framework of recommender system is designed using intelligent agent. The framework consists of agent, viz. profile agent, content-based agent, and collaborative agent. Profile agent automatically extracts the input from profile of user. Users may also have tried to search some resources. The search query and the results of the search are used by profile agent while collecting more information about active user. The task of identification of subject area(s) of active user is performed by profile agent. Then profile agent updates the user profile. Content-based agent and collaborative agent perform the main tasks of filtering and providing recommendations. In general, the perfect simulation of the natural recommendation process that human being performs to recommend to someone can provide better results. In this scenario, the main problem to the researchers is to think about the framework which can provide such perfect simulation. The main objective of the proposed architecture is to develop an agent-based recommender system for providing effective and intelligent use of library resources such as finding right book(s), relevant research journal papers and articles. The various issues with the dataset and library recommender system are identified. There are many recommendation domain and applications where content and metadata play a key role can be seen in the literature. In domains such as movies, the relationship between content and usage data has seen thorough investigation done already but for many other domains such as books, news, scientific articles, and Web pages, it is not still known if and how these data sources should be combined to provide the best recommendation performance. This has motivated to develop recommender system using library domain specifically. Some of the datasets such as Book-Crossing dataset, Techlens + dataset, ACL anthology reference corpus (Bird in The ACL anthology reference corpus: a reference dataset for bibliographic research in computational linguistics [2], Torres et al. in Proceedings of the 4th ACM/IEEE-CS joint conference on digital libraries, pp. 228–236 [3], Ziegler in current trends in database technology-EDBT 2004 workshops [4]) are available. They are not very useful since only the use of title of the book is not sufficient to improve the performance (Ziegler in current trends in database technology-EDBT 2004 workshops, pp. 78–89 [4]), not made publically available (Torres et al. in Proceedings of the 4th ACM/IEEE-CS joint conference on digital libraries, pp. 228–236 [3]), limited domain (NLP only) (Bird in The ACL anthology reference corpus: a reference dataset for bibliographic research in computational linguistics [2]). Since no benchmarking dataset is suitable, the proposed research contributes by creating dataset for library recommender systems. Some of the library resources are added from library of congress using Z39.50 protocol. Others are from catalogues of various publishers’, viz. PHI learning, Laxmi, Pearson education, Cambridge, and McGraw Hill publications. User profiles, Library Resources, and Knowledge base (ACM Computing Classification System 2012 used as Ontology) are important components of dataset. Profile agent is a goal-based agent which auto extracts more information about user and updates profile by removing noise. The results for the task of identification of user’s interested subject area(s) are improved using n-Gram and Jaccard’s measure (Kern et al. recommending scientific literature: Comparing use-cases and algorithms [5]). The comparison made with the similar work in the literature gives that ‘implicit interest extraction of the user,’ ‘implicit feedback mechanism,’ and ‘weighted profiles updation using semantic concepts in ACM CCS 2012’ are novel factors of profile agent (Kim and Fox in Digital libraries: international collaboration and cross-fertilization. Springer, Berlin, pp. 533–542 [6]). Library recommender agent is utility-based agent implemented using hybrid approach (combining results of content based and collaborative agent). Agent performs tasks such as classifying library resources, filtering and providing recommendations. Library resources are classified among the category in ACM CCS 2012. The task of classification is well performed using PU learning (Naive Bayes classifier) since no negative records exist in any real-world library (Wang and Blei in Proceedings of the 17th ACM SIGKDD international conference on knowledge discovery and data mining, pp. 448–456 [7]). Filtering library resources require to measure similarity between user profile and library resources. It is found that cosine distance outperformed with other measures such as Euclidean, Manhattan, and Chebyshev. Singular value decomposition technique helped to reduce the final computations needs to perform while providing recommendations. The ability to provide semantically related library resources by retrieval of related concepts in ACM CCS 2012 and weight assignment have encouraged the performance of library recommender agent (Morales-del-Castillo et al. in Int J Technol Enhanced Learn 2(3):227–240 [8]). The results of the library recommender system are evaluated using precision, recall, and F1. This proposed agent-based library recommender system improved performance up to the precision 81.66, recall 83.19, and F1 82.28 percentages (Morales-del-Castillo et al. in Int J Technol Enhanced Learn 2(3):227–240 [8], Porcel et al. in international conference on education and new learning technologies [9], Tejeda-Lorente et al. in Procedia Comput Sci 31:1036–1043 [10]). The research concludes that the discovery of the user interest and retrieval of semantically related recommendations are the important factors on which overall performance of recommender system depends. Integration of intelligent agent-based recommender system in existing library system is focused as the future work.