A bunch of (mostly) rendered Jupyter notebooks about my thesis work. The blog is hosted on GitHub Pages and is available at jindrich.bar/thesis-blog.
While I try to make the posts make sense on their own, it’s possible they will be hard to understand without the context at times. To learn more about the motivation behind the problems solved here, read the thesis first and treat these posts as the supportive “behind the scenes” material.
Part of the thesis work is to evaluate the performance of the basic ranking algorithm - and see whether we can improve it by utilizing the graph structure of the data. This post is about creating the benchmark we will use to evaluate the performance of the algorithm.
A short post about collecting, cleaning and exploring the Elsevier Scopus data based on the queries from the previous post.
With the collected Scopus and Charles Explorer data, in this post we are exploring the possibilities of the automated relevance feedback system based on LLM-embeddings based similarity search and we see whether (and how much) do the measures of the local graph structure improve the search result ranking.
The data exports from the university systems are missing detailed data about external authors from other universities. One large problem is that the external authors are only identified by their name. To automatically merge the graph nodes representing the same person, we try to employ a hierarchical merging algorithm. In this blog post, we benchmark different ways of calculating the distance matrix between the mergeable nodes.
In this short blog post, we evaluate the performance of the hierarchical merging approach (based on the graph structure) and compare this with the naïve approach (based on the normalized names only).
In this post, we explore the possibilities of inferring the citation count of the papers based on the local graph structure. Using various local graph structure measures (publication node degree, various centrality measures, node cut for the neighborhood, etc.), we try to predict the citation count of the papers. We also benchmark the performance of the prediction models and assess the usability of the local graph structure measures for ranking the publications based on their predicted citation count.