RA Notes

/ 2021-09-30

Notes: This post is written to keep the ideas alive and document what I learn from the research work with Nate . Since Aug. 2021, I began to contribute to the project of Nate which is intended to find the connections between two online communities - Fandom wiki (which was previously known as Wikia) and Reddit . I make a role in the dataset construction.

  • Wikidata and SPARQL language: Wikidata is an open source knowledge graph that has structured data about relationships between different entities. Here is an example: One Piece . Wikidata provides a query service which can run by SPARQL language, to find the qualified entities and properties in the dataset. I have learned the rules for SPARQL, and pulled the entities that have both a fandom wiki and a subreddit ID.

  • Linux: To be more efficient to collaborate, I set the Linux and began to join the git of Community Data Science Collective , which is a research group focused on online communities.

    • use explorer.exe . to find the Linux files in Windows
    • if sudo apt install ... fails, it can probably be solved by sudo apt install -y ... --fix-missing(see in Help Docs )
    • to find the ip for linux, use ifconfig -a
    • to open xrdp, use sudo service xrdp restart

  • Wikidata: According to the query outcome in Wikidata, it seems like dataset in Wikidata does not have a good coverage for the match between a fandom wiki and a subreddit, but include almost no error due to the properties for Wikipedia. We are going to deal with more data to get back at the dataset in Wikidata.

  • Machine Learning Classification: To find out whether a subreddit and a fandom wiki is really similar or the mapping between them is correct, there can be some ideas from machine learning classification .

  • Precision and recall: How good the mapping in Wikidata is to be checked. Precision and recall rate can be used to examine it. In the supervised machine learning classifier, we have gotten a precision of 95%!

  • Machine Learning for Social Sciences: Machine learning models used in social sciences has triggered a heated discussion: whether it is “garbage in, garbage out” (Geiger et al., 2019 ) or “theory in, theory out” (Radford & Joseph, 2020 )?

  • Rules for prediction: Simpler than real ML classifier, if we give a set of rules to predict the correctnesss of mappings it would also make sense. An example rule Nate gave is:

    The wiki is the top linked wiki from the subreddit AND
(The subreddit and wiki are linked in wikidata OR
    The subreddit and wiki names are similar OR
    The subreddit-link zscore for the wiki is large
    )

  • Validation Set

  • More text features: We expanded the rules to check the text match ratio besides edit distance by using SequenceMatcher , to make a more complicated and accurate prediction. This is quite connected with Natural Language Processing . I have labeled 200 more cases and done a logistical regression model for testing if the above rules of prediction work well for a larger set.

  • MapReduce: MapReduce is a programming model. It is composed of a map procedure, which conducts filtering and sorting, and a reduce method, which conducts a summary operation.

  • Simulation: Reality is recursive. The macro outcome can be often surprising even if we have known the micro behavior well. Simulation methods, especially agent-based models, can be used to study the emergence in the complex systems, such as how individual behavior bias can accumulate to the group size difference (Foote et al., 2020 ). Simulation can be a promising field and method to address the micro-macro link in social scientific research.

  • Hyak: High performance computer cluster at UW. It can be used to conduct large amount of computing.

Last modified on 2021-09-30