Hasil Pencarian  ::  Simpan CSV :: Kembali

"In this book, we survey the most recent and state-of-the-art work about a whole new generation of recommender systems built to serve social tagging systems. The book is divided into self-contained chapters covering the background material on social tagging systems and recommender systems to the more advanced techniques like the ones based on tensor factorization and graph-based models."

"This book aims to help readers to discover and understand the interplay among legal issues such as privacy, technical aspects such as interoperability and scalability, and social aspects such as the influence of affinity, trust, reputation and likeness."

"Social tagging (including hashtags) is used over platforms such as Twitter, Facebook, Instagram, Pinterest, WordPress, Tumblr and YouTube across countries and cultures, meaning that one single hashtag can link information from a variety of resources. This new book explores social tagging as a potential form of linked data and shows how it can provide an increasingly important way to categorize and store information resources. The internet is moving rapidly from the social web embodied in Web 2.0, to the semantic web (Web 3.0), where information resources are linked to make them comprehensible to both machines and humans. Traditionally, library discovery systems have pushed information, but did not allow for any interaction with the users of the catalogue, while social tagging provides a means to help library discovery systems become social spaces where users could input and interact with content. The editors and their international contributors explore key issues including the use of hashtags in the dissemination of public policy, the use of hashtags as information portals in library catalogues, social tagging in enterprise environments, the linked data potential of social tagging, [and] sharing and disseminating information needs via social tagging. Social Tagging in a Linked Data Environment will be useful reading for practising library and information professionals involved in electronic access to collections, including cataloguers, system developers, information architects and web developers"

"Sistem rekomendasi menjadi salah satu kebutuhan utama bagi penyedia layanan e-commerce untuk memberikan saran rekomendasi produk sesuai dengan apa yang diinginkan oleh pengguna. Salah satu pendekatan yang paling banyak dilakukan dalam membangun sistem rekomendasi adalah collaborative filtering, dengan menggunakan data explicit feedback, yang dapat berupa review atau rating. Sistem rekomendasi dengan pendekatan collaborative filtering telah banyak dikembangkan dengan menggunakan metode machine learning dan metode deep learning. Penelitian ini berfokus untuk mengembangkan sistem rekomendasi dengan pendekatan collaborative filtering berbasis deep learning dengan menggunakan data gabungan review dan rating. Teknik deep learning yang digunakan diperkaya dengan word embeddings untuk dapat menangkap interaksi yang terdapat dalam data review. Penelitian ini menggunakan arsitektur yang diadopsi dari CARL. Modifikasi yang dilakukan pada CARL meliputi pengubahan optimizer dan penggunaan beberapa pretrained word embedding yang berbeda. Selain itu, penelitian ini juga membandingkan performa sistem rekomendasi yang diusulkan antara dataset berbahasa Inggris dan berbahasa Indonesia. Untuk melakukan evaluasi performa sistem rekomendasi yang dikembangkan, digunakan metrik evaluasi mean squared error (MSE). Hasil penelitian menunjukkan modifikasi model CARL (Review-based) dengan menggunakan optimizer Adam (CARL (Review-based) – Adam) menunjukkan performa terbaik dan dapat mengalahkan performa dari baseline model.

---

Recommender systems are one of the main needs for e-commerce to provide product recommendations according to what the users want. One of the most widely used approaches in developing recommender systems is collaborative filtering, using explicit feedback data, which can be in the form of reviews or ratings. Various collaborative filtering methods have been developed using machine learning and deep learning methods. This study focuses on developing deep learning-based recommender systems with collaborative filtering approach using combined reviews and ratings data. The deep learning technique that being used is enriched with word embeddings to capture the interactions contained in the review data. This study uses an architecture adopted from CARL. Modifications made to CARL include changing the optimizer and using several different pretrained word embeddings. This study also compares the performance of the proposed recommender systems between English datasets and Indonesian datasets. To evaluate the performance of the recommender systems, the mean squared error (MSE) evaluation metrics is used. The results showed that the modification of CARL (Review-based) model using Adam optimizer (CARL (Review-based) – Adam) showed the best performance and could beat the performance of the baseline model."

"Sistem perekomendasi semakin menjadi bagian yang tak terpisahkan dengan sistem komersial online, seperti toko online atau layanan film/IPTV on-demand. Tugas utama sistem perekomendasi adalah memberikan rekomendasi produk atau konten kepada pelanggan. Sistem perekomendasi bekerja berdasarkan informasi pribadi pengguna seperti riwayat belanja, item yang dilihat atau diinginkan, se- hingga timbul resiko hilangnya privasi karena penggunaan sistem tersebut. Penelitian sebelumnya telah mempelajari potensi untuk mempertahankan baik privasi pengguna maupun akurasi rekomendasi, namun masih terbatas pada algoritma sistem perekomendasi tertentu saja. Sistem perekomendasi terkini yang menghasilkan rekomendasi paling akurat, menggunakan teknik faktorisasi matriks, dan sejauh ini, sepengetahuan kami belum pernah dipelajari dalam penelitian privasi. Dalam penelitian ini, kami mencoba menerapkan kerangka privasi diferensial ke dalam faktorisasi matriks. Privasi diferensial memberikan jaminan privasi yang terbukti secara teoritis, misalnya dalam kondisi pengetahuan awal apapun, data masing-masing individu tidak dapat diketahui berdasarkan output agregat (sistem perekomendasi). Kami menganalisa beberapa cara untuk menerapkan privasi diferensial dalam konteks ini, yaitu menambahkan derau pada input; di dalam proses (menggunakan gradient descent); dan pada output proses. Kami mengimplementasikan dan mengevaluasi semua metode pendekatan. Di akhir, kami membahas dan memberikan hasil perbandingan tingkat kegunaan dan privasi. Hasil evaluasi menunjukkan bahwa meski perturbasi input lebih baik dibanding perturbasi gradient descent dan output, seluruhnya menunjukkan tingkat kegunaan yang baik hanya dapat diperoleh pada tingkat privasi yang kurang dapat diterima.

---

Recommender systems are becoming an integral part of commercial online systems, e.g., shopping websites or on-demand movie / IPTV services. The main task of a recommender system is to provide recommendations of products or content to customers. As recommender systems are based on personal information about users’ prior purchases, views or wish lists, there is an inherent loss of privacy resulting from the use of such systems.

Prior works explored to some extent the potential of attaining both users’ privacy and good recommendations, however only for a limited set of recommender system algorithms. The state-of-the-art recommender systems that provide the most accurate recommendations are based on the technique of matrix factorization, and so far, to the best of our knowledge, were not addressed in privacy research.

In this project, we address this gap by applying the differential privacy framework to matrix factorization. Differential privacy provides theoretically provable privacy guarantees, i.e., that under any conditions of prior knowledge, individuals data cannot be derived from the aggregated (recommender system) output. We analyze different ways of applying differential privacy in this context, including introduction of noise to the input; within the mechanism (using gradient descent); and at the output of the mechanism. We implement and evaluate all of the approaches.

Finally, we discuss and provide insights into the resulting utility and privacy tradeoffs. The evaluation shows that although input perturbation is superior to gradient descent and output perturbation, all demonstrate that satisfactory utility levels can be obtained only at the expense of unacceptable privacy levels."