Conspiracists Also Viewed: ‘Problematic’ Networks of Recommendation on Amazon.com

Team Members

Dylan O'Sullivan, Ekaterina Khryakova, Jingyi Wu, Mingzhao Lin, Yingying Chen

Facilitators: Tommaso Elli and Jonathan Gray

Essay Summary

The whole project is about relationships between a list of problematic books selling on Amazon and their “also-viewed” recommendation books. The main tools we used in the research are DMI recommendations scraper (Peeters, 2020) and Gephi. At the beginning, we manually collected the “also-viewed” recommendation books of the conspiratorial books (problematic status marked as 5) in the Problematic Infomedic Amazon Book list, which is provided by the infodemic amazon project team, on Amazon.com. By doing that, we can interface the recommendation algorithms directly on the platform. DMI recommendations scraper was used to enlarge the “COVID” direct search database. Gephi helped us to visualise the data we scraped from amazon and allows both researchers and readers to see the networks between different books clearly.

Since the conspiratorial books formed a network, we found that a non-conspiracy book can be a bridge between two conspiracy books. That is, on the one hand, we can reach a nonconspiratorial book via the recommendation section of a conspiratorial book. On the other hand, by looking at the recommendation section of the non-conspiracy book, customers can also find a book which is conspiratorial.

According to the results we have collected and the related analysis, readers will develop a better understanding of the recommendation system and algorithm of books selling on Amazon. In addition, they can also see what genres of books are most related to the COVID-19 conspiracy theory.

Research Questions

RQ 1: How easily might one enter a conspiratorial network on Amazon.com, via recommendations or otherwise?

RQ 2: How interconnected are conspiratorial books and genres—nodes and clusters— within the broader conspiratorial network?

RQ 3: What is the connectivity—at an individual and generic level—between conspiratorial and non-conspiratorial books?

Methodology

In this research, we first looked into the interface of Amazon.com. We manually scraped all the information from the platform to get a more comprehensive understanding of the more extensive infrastructure: namely, the “behaviour” of the platform, the topics/themes of books, the logistics of the recommendation system “on the field,” amongst other things. It is vital to fully understand the logic of the recommendation system and look into the interface of the platform directly.

Subsequently, we found that the recommendation methods of Amazon can be divided into six categories: “Frequently Bought Together”, “Viewed Also Viewed”, “Bought Also Bought”, “Buy After Viewing”, “More To Explore”, and “Items Related,” we ultimately chose “Viewed Also Viewed” as our core algorithm because it appeared most often across the recommendation system as a whole.

To analyze the connections of conspiracy books—“exploring associations around single actors” and “detecting key players”—narrative readings of the networks were applied (Liliana et al., 2016). We started from the list of Problematic Infomedic Amazon Books (Appendix 1), which was provided by the Infodemic Amazon Project team, and manually collected the “Viewed Also Viewed” recommendation books of the 12 conspiratorial books (problematic status marked as 5) in the list on Amazon.com. In order to avoid personalization settings and location influences, our team used the Hoxx VPN Proxy Chrome plugin on Google Chrome (private browser mode) with the VPN set to the United States of America.

Figure 1: The 12 Conspiratorial Seed Books in the Problematic Infomedic Amazon Book list

After finishing the manual data collection, based on 628 books that have been recorded, we merged books which appeared repetitively, as well as different editions of the same book, to avoid errors and improve graph readability. Ultimately, we were left with 396 books in total.

The second step of this project is to categorise the 396 books from four aspects, what is its genre (21 genres in total, including “Health & Fitness”, “Christianity”, “Self-Help”, “Illuminati”, “Vaccines”, “QAnon”, “Deep State”, “Covid-19”, “Flat Earth”, “General Conspiracy”, “Culture & Society”, “5G & EMF”, “Aliens”, “Fiction & Biography”, “Eastern Philosophy”, “Science & Technology”, “Obama & Democrats”, “History”, “Trump & MAGA”, “9/11” and “Miscellaneous”), whether it is conspiratorial, is it in the list of Problematic Infomedic Amazon books, and is it a conspiratorial seed book. To do this, three of our group members viewed each books’ name, summaries, category, and publication year one by one. When we define a book is conspiratorial, we are following five principles: nothing happens by accident, nothing is as it seems, everything is connected, tone/style of conspiracy theories and assumption that this is somehow going against received wisdom (Butter and Knight, 2020)

To broaden the core dataset, DMI recommendations scraper was used to scrape the recommendations from the first 12 books in the direct “COVID” query. From this scrape, one hundred eighty-nine books from the “Viewed Also Viewed” recommendation list were added and analysed.