Auditing E-Commerce Platforms for Algorithmically Curated
Vaccine Misinformation
This post summarizes a research paper that audited Amazon for algorithmically curated vaccine misinformation. This paper was accepted to CHI’21, a premier conference in Human-Computer Interaction. This work was done with Dr. Tanu Mitra.
People increasingly rely on search engines for health information. Yet, the algorithms powering search engines are not traditionally designed to take into account the credibility and trustworthiness of such information. In recent times, Amazon specifically has faced criticism from several technology critics for not regulating health-related products on its platform with critics going as far as to call it a “dystopian” store for hosting anti-vaccine books. What is the amount of health misinformation present in Amazon’s search results and recommendations? How does personalization due to user history built progressively by performing real-world user actions, such as clicking or browsing certain products, impact the amount of misinformation returned in subsequent search results and recommendations? In our paper, we dabble into these questions. We conduct 2 sets of systematic audit experiments: Unpersonalized audit and Personalized audit.
Method
- Compiling high-impact vaccine-related topics and search queries: We curated a list of 48 search queries belonging to 10 popular vaccine-related topics like ‘HPV vaccine’, ‘immunization’, ‘MMR vaccine and autism’, etc using Google Trends and Amazon’s autocomplete suggestions.
- Selecting Amazon components to audit: For the audits, we collected 3 major Amazon components namely, search results, autocomplete suggestions, and recommendations. Amazon presents several recommendations as users navigate through the platform. We collected recommendations present on three different Amazon pages: homepage, pre-purchase page, and product pages. Each page hosts several types of recommendations (see the Figure above).
- Unpersonalized Audit: To gain in-depth insights about Amazon’s searching and sorting algorithm, we collected search results of 48 vaccine-related search queries without logging in to Amazon to eliminate the influence of personalization for 15 consecutive days. The search results were collected across 5 different Amazon filters each day: “featured”, “price low to high”, “price high to low”, “average customer review” and “newest arrivals”. We also extracted the recommendations present on the product pages of the first three search results. All collected products were later annotated for their stance on health misinformation — promoting, neutral, or debunking.
- Personalized Audit: To determine the impact of personalization due to user history on the amount of health misinformation returned in search results, recommendations, and auto-complete suggestions, we created several bot accounts that built user history progressively over 7 days by performing several real-world actions, such as “search”, “search + click”, “search + click + add to cart”, “search + click + mark top-rated all positive review as helpful”, “follow contributor” and “search on third party website” ( Google.com in our case). We controlled for several extraneous factors that could lead to potential errors in our audit data collection. Every day the bot account built history by performing actions on a single Book/contributor and then collected various recommendation pages, search results, and autocomplete suggestions.
Both our audit experiments resulted in a dataset of 4,997 unique Amazon products distributed across 48 search queries, 5 search filters, 15 recommendation types, and 6 user actions, conducted over 22 (15+7) days. We have released our dataset to the public in the hope to promote research on algorithmically curated misinformation on e-commerce platforms.
Summarized findings of Unpersonalized audit
- We find a higher percentage of products promoting health misinformation (10.47%) compared to products that debunk health misinformation (8.99%) in the search results. Similarly, the percentage of misinformative product page recommendations (12.95%) is much higher than the debunking recommendations (1.95%).
- Amazon returns a high number of misinformative search results when users sort their searches by filters “featured” or “average customer reviews” and a high number of debunking results when they sort results by the filter “newest arrivals”.
- Amazon ranks misinformative results higher than debunking results especially when results are sorted by filters “average customer reviews” and “price low to high”.
- Analysis of product page recommendations suggests that recommendations of products promoting health misinformation contain more health misinformation when compared to recommendations of neutral and debunking products
Summarized findings of Personalized audit
- Search results sorted by filters “average customer review”, “price low to high” and “newest arrivals” along with auto-complete suggestions are not personalized. User actions involving clicking a search product leads to personalized homepages.
- We find evidence of the filter-bubble effect in various recommendations found in homepages, product, and pre-purchase pages.
- The amount of misinformation present in homepages of accounts building their history by performing actions “search + click” and “mark top-rated all positive review as helpful” on misinformative products was more than the amount of misinformation present in homepages of accounts that added the same misinformative products in the cart. The finding suggests that Amazon nudges users more towards misinformation once a user shows interest in a misinformative product by clicking on it but hasn’t shown any intention of purchasing it.
- Overall, our findings suggest that once users start engaging with misinformative products, they will be presented with more misinformative stuff at every point of their Amazon navigation route and at multiple places.
Combatting health misinformation
Through our findings, we recommend several short-term and long-term strategies that Amazon can adopt to combat misinformation on its platform.
- Short term strategies: design interventions. The simplest short-term solution would be to introduce design interventions. The platform can use interventions as an opportunity to communicate to users the quality of data presented to them by signaling misinformation bias. The platform could introduce a bias meter or scale that signals the amount of misinformation present in search results every time it detects a vaccine-related query in the search bar. The bias indicators could be coupled with informational interventions like showing Wikipedia and encyclopedia links. The second intervention strategy could be to recognize and signal source bias. Through our work, we also present a list of authors who have contributed to the most misinformative books. Imagine a design where users are presented with a message “The author is a known anti-vaxxer and is known to write books that might contain health information” every time they click a book written by these authors.
- Long term strategies: algorithmic modifications and policy changes. Long term interventions would include modification of search, ranking, and recommendation algorithm. Our investigations revealed that Amazon has learnt from users’ past viewing and purchasing behavior and has categorized all the anti-vaccine and other problematic health cures together. Amazon needs to “unlearn” this categorization. The platform should also stop promoting health misinformative books by sponsoring them. It should introduce some minimum quality requirements that should be met before a product is allowed to be sponsored or sold on its platform. In recent times Amazon introduced several policy and algorithmic changes including rollout of a new feature “verified purchase” to curb the fake reviews problem on its platform. Similar efforts are required to ensure product quality as well. Amazon can introduce a similar “verified quality” or “verified claims” tag with health-related products once they are evaluated by experts.
We hope our work acts as a call to action for Amazon and also inspires vaccine and health audits on other platforms.
