Mahta Bakhshizadeh

• This PhD thesis explores how Recommender Systems (RSs) can be leveraged to develop Personal Knowledge Assistance (PKA), aimed at improving the productivity of Knowledge Workers (KWers) such as programmers and scientists who primarily handle information rather than manual labor and spend significant time searching for information at the expense of productivity. The thesis introduces this novel application of RSs to create a more intelligent and supportive working environment for KWers and identifies key challenges such as dealing with heterogeneous information items and capturing dynamic needs. Based on a conducted systematic literature review, the thesis distinguishes the unique characteristics that set this domain of RSs apart from others and highlights the potential of the most relevant RS categories to address these challenges. It explores technological foundations developed over two decades of PKA research that can support the development of the targeted RSs, and also introduces the concept of Contextual States (CSs) as multi-dimensional sessions representing relevant contextual information. This study divides the information space of KWers into three layers (personal, corporate, and global spaces) and proposes a framework to integrate these layers along with the concept of CSs into PKA technologies. It then conducts a case study based on the proposed framework with real-world data from the DFKI's Smart Data and Knowledge Services (SDS) department. Using a TF-IDF-based method, 1,987 recommendations were generated across 128 contexts, with 54% of these recommendations evaluated as relevant and half of them considered helpful. Recognizing the lack of comprehensive, public datasets for PKA research, the thesis introduces the Real-Life Knowledge Work in Context (RLKWiC) dataset. RLKWiC provides extensive contextual information and annotations, including over 61,000 desktop events, 211 personal concepts, 393 DBpedia resources, and personal KGs with over 6,400 nodes and 3,100 inter-relations. It aims to support benchmarking and evaluating PKA services. The study establishes a benchmark on RLKWiC for context-based Entity Recommendation (ER), offering full transparency and reproducibility with over a thousand entities labeled with explicit relevance scores. The baseline recommendation scenario achieved 56% precision for relevant entities and 25% for representative entities. This performance was subsequently improved by integrating a semantic-based approach with an Adaptive Relevance Prediction (ARP) module, which increased the precision for representative entities by 20%. Semantic methods, particularly those using Laplacian kernel and Euclidean distance metrics, were shown to effectively maintain context-based relevance. Finally, the thesis explored integrating Large Language Models (LLMs) into RSs for KWers. By using Mistral 7B and an ARP mechanism, the LLM-based approach significantly outperformed both the baseline and semantic-based methods in ER, demonstrating LLMs' potential for enhancing PKA recommendations. In conclusion, this research highlights RSs as a promising way to mitigate information overload in KW scenarios by delivering relevant information to enhance productivity. It includes a comprehensive literature review, a proposed framework, the creation of RLKWiC, a benchmark for ER evaluation, and incremental improvements using semantic and LLM-based methods.