University and Public Library Recommendation Dataset for Sparse and Cold-Start Research

Description

This datasets contain anonymized user borrowing transactions and associated metadata collected from a university and public library system. The University Library dataset contains 14,148 users, 8,369 library items, and 52,394 borrowing interactions, exhibiting an extreme sparsity level of 99.97%. The Public Library dataset contains 53,008 users and 54,563 items with similarly sparse interaction patterns of 99.98%. The dataset was developed to support research on recommender systems, particularly studies addressing the challenges of extreme interaction sparsity and cold-start conditions. The collection consists of three interconnected data components: 1. User Data Anonymized user attributes including academic program, academic year, cumulative GPA, gender, and other demographic characteristics that can be used as auxiliary information for recommendation modeling. 2. Item Data Bibliographic metadata describing library resources, including title, author, publisher, publication year, ISBN, DDC classification, and location information. These attributes can support content-based recommendation and metadata-driven cold-start strategies. 3. Borrowing Transactions Historical borrowing records representing user–item interactions. These records can be transformed into user–item interaction matrices for collaborative filtering, hybrid recommendation systems, and interaction reconstruction research. The dataset exhibits characteristics commonly observed in real-world recommendation environments, including: - Extreme interaction sparsity (>99%) - High proportions of cold-start users and cold-start items - Heterogeneous user and item metadata - Long-tail borrowing behavior All personally identifiable information has been removed or anonymized prior to publication. The dataset is provided exclusively for research and educational purposes. Methods: Data were extracted from the university library management system and subsequently anonymized. User identifiers were replaced with anonymous IDs, while personal information was removed to ensure privacy protection. Data preprocessing included: 1. feature selection by the domain expert. 2. Data cleansing, to addressing missing data and outlier, removed duplication, and standardization). 3. Feature transformation. The technique used is One-Hot Encoding, which is used to convert categorical data into a numeric representation so that it can be used by machine learning algorithms. 4. Metadata integration and validation Keywords: Recommender Systems; Cold-Start Problem; Data Sparsity; Digital Library; Collaborative Filtering; Metadata-Based Recommendation; Matrix Densification; Library Analytics

Steps to reproduce

The datasets are extracted from the anonymous university and the public libraries’ management systems. Data that could reveal personal information has been removed to maintain privacy. Each original dataset has three metadata domains: borrowing transactions, user profile information, and item (bibliographic) metadata. Borrowing metadata records user-item interactions, while user profiles provide demographic and academic attributes of users, and item profile metadata describes library collection attributes. These three datasets are then integrated using unique user and item identifiers and consolidated into a unified dataset. Then, data preparation is carried out by domain expert analysis, including the cleansing process and the transformation of features, both text and numeric, into a format ready for processing. The resulting integrated dataset enables analyses that combine interaction data with user and item characteristics, supporting research in recommendation systems, user behaviour analytics, information retrieval, and digital library services.

Institutions

• Binus University

  Jakarta, Jakarta

Categories

Licence