Add graph-based paper recommendations with PageRank

[grpinto]

Made with @raeesay

• Added a new recommendations API endpoint to suggest papers using Personalized PageRank.
• Integrated recommendation functionality into the web UI with toggle options for graph recommendations and explanations.
• Enhanced paper card display to show recommendation scores and reasons.
• Updated search functionality to support recommendations and explanations.
• Introduced tests for recommendation logic and behavior.

[grpinto]

Hello, I was at your workshop today and thought that this would be a cool idea to add. A recommendation paper system. I changed the code and made a PR, feel free to reject it if you think it does not suit your current goals.

[thawn]

@grpinto thank you so much ❤️
I will look into it after HAICON

[thawn]

@grpinto, @raeesay

Thanks again for your PR. I just had a look at the code changes and tried it out on my machine.

The feature looks super promising, however I would like to understand fully what is happening under the hood and I also have some ideas how we could integrate it with the "interesting papers" feature.

Understanding the feature

let me summarize how I understand the feature. Please correct me where I am wrong and extend where I missed things.

1. a list of candidate papers is fetched based on the filter settings (to ensure that only papers from the currently selected conference are in the search results)
2. a semantic search based on the user query is performed (same as before)
3. the results from the semantic search are used as seeds to construct a graph based on authorships, shared keywords and shared sessions
4. The graph is used to calculate a "personalized page rank" I do not understand yet, how that rank is calculated and would be super grateful for an explanation
5. The search results are re-ranked based on the page rank and displayed
6. Optional: the chat model is used to explain the ranking - I did not understand these explanations at all. The chat model likely needs more context about how the ranking algorithm works (just like me ;-) )

Suggestions from my side

1. have you tested whether the personalized ranking makes the results more relevant for you?
2. I think this feature could be super useful, if we combined it with the "interesting papers" feature. What I had in mind, would be to re-rank the semantic search results based on how well they align with the papers that are already selected in the interesting papers list - that way, we would get a truly personalized page rank

Cheers and thanks again for contributing to the project!