Eli Pariser coined the term filter bubble in his book “The Filter Bubble: What the Internet is Hiding from You.” in which he argues that web-based systems that deliver personalized content put their users in filter bubbles and limit their exposure to content more representative of the actual information space diversity[1]. Following recommendations, users are trapped in an unchanging space that limits creative thinking while reinforcing adopted views and beliefs. In the book, Pariser point to the possible harmful effects of filter bubbles on online discourse and society.
In this post, we will cover the basics of recommender systems and personalization and their possible negative effects on content consumption. We will also show how filter bubbles emerge in personalized display advertising and discuss possible effects on media planning and buying.
Personalization and recommender systems
Personalization has been the key for many companies to improve user experience or increase sales. For example, Netflix reported that 75% of its users watched came from recommendations[2], while Amazon says that 35% of its sales came from customers buying a recommended product[3]. This leaves little space for debate about the efficacy of recommender systems and personalization in commerce, but how do recommender systems function?
Task and setup
Given a set of users, items and ratings for these items, the task of a recommender system is to provide a set of items for each user that improves their user experiences. For a streaming platform, this would mean recommending the next movie or song to listen to, while in online shops, the gaol would be recommending the next product to buy.
Types of recommender systems
There three main types of recommender systems:
- Collaborative filtering – the system recommends items to the user by using ratings and preferences of all users assigned to the available items. There are two types of collaborative filtering approaches user-user and item-item.
- Content-based – the system recommends similar items based on a particular item and behavioral or demographic data about the user. The recommendation is made by analyzing available content or metadata about the item. The intuition behind these types of recommender systems is that if a user likes a particular item, he or she will also like an item that is similar to it.
- Hybrid approaches – the system makes recommendations using content-based and collaborative filtering simultaneously. The idea is to eliminate the shortcomings of the two approaches, such as the cold-start problem in collaborative filtering and data sparsity in content-based approaches.
Popular algorithm for content-based recommendation is k-nearest neighbor, while collaborative filtering relies on more complex approaches such as matrix factorization. In recent years deep learning-based approaches gained momentum. However, a recent paper questioned the progress made by deep learning approaches, stating reproducibly issues difficulties and bad experimental design, i.e., choice of the baselines[4].
The filter bubble effect has been extensively discussed in the context of collaborative filtering-based recommender systems, showing that taking recommendations lessens the risk of a filter bubble[5]. It is still an open question if content-based recommender systems have the same effect.
The filter bubble effect in display advertising
The filter bubble effect emerges in display advertising due to target advertising as systems try to serve personalized ads to increase conversion. Unfortunately, the filter bubble effect has not been examined in much detail in the context of display advertising. In the following, we will explain the main differences between targeting advertising and recommender systems and the data used to achieve effective display ad targeting.
Task and setup
Given a user and a set of display ads (banner ads), the goal is to decide which display ad to show the user. In the display advertising context, this process is often referred to as targeting. Compared to the recommender system setup, this is a different setting. Next to selecting an ad to deliver to the user, there is a cost associated with the number of times an ad is shown or the number of ad impressions. Further, this setup is highly dynamic in terms of the number of the items (display ads) recommended. For example, there can be far more and constantly changing ad creatives advertising a single product compared to a web shop’s relatively static number of products. The main difference, however, is the competitive nature of the real time bidding process behind each recommendation (delivered targeted displayed ad).
Targeted display advertising approaches
To achieve high personalization and effective targeting, algorithms rely on the following data types:
- Behavioral data – refers to how individuals or groups behave or interact with a product, service, or environment. In display advertising, behavioral data can be collected from various sources, such as website analytics, social media interactions, or surveys, and can provide valuable insights into customer behavior and preferences.
- Demographic data – refers to statistical information about the characteristics of a population and users. This can include age, gender, income, education level, occupation, and other factors that can be used to classify or describe a group of people.
- Geo-location – data in the online advertising to deliver ads to users based on their geographic location. Geo-targeting can be achieved through various means, such as IP addresses, GPS coordinates, or zip codes. Geo-targeting allows advertisers to target their ads to specific regions, cities, or neighborhoods and to tailor their ads to users’ specific needs and interests in those locations.
- Content-based – Content and contextual advertising is a type of online advertising that serves ads to users based on the context of the website or content they are viewing. In other words, the displayed ads are relevant to the content or theme of the webpage or app. Contextual advertising uses algorithms to analyze the content of a webpage or app and match it with relevant ads. This can be done by analyzing the words and phrases used on the page, the images and videos displayed, or other contextual cues.
Popular algorithms relying on these data types are typically machine learning-based approaches.
Retargeting could increase the filter bubble effect
Compared to recommender systems, the filter bubble effect in display advertising could be even higher due to retargeting activities. Media buyers and planners use retargeting to serve display ads for a specific product to users who have already seen a targeted ad, embarked on a customer journey, or even completed it. As retargeting is often realized using cookies and/or pixels, the filter bubble effect in display advertising is tightly related to user tracking and privacy.
Consequences of the filter bubbles effect for media planners and buyers
The filter bubble effect in display advertising leads to a situation in which users see the same display ads. Intuitively high exposure to the same products might increase conversion. However, using (re-)targeting extensively and putting your customer in a filter bubble can hurt your brand or product. Furthermore, the filter bubble effect can lead to banner blindness – situations in which users ignore ad creatives.
Similar to ordinary users, media planners and buyers often live in their own filter bubbles concerning content and display ads shown next to it. They are trapped in an unchanging display advertising landscape that limits their competitive edge as they are blind to the advertising strategies of competitors.
Invest in data and analytics foundations: Media buying and planning is impossible if marketers don’t have the means to understand the needs of high-value customers on an ongoing basis. For this purpose, top marketers have to monitor the advertising land space and collect and analyze structured and unstructured data to identify behavioral patterns of competitors while ensuring they are using the optimal media buying strategy.
References
[1] Pariser, E. (2011). The filter bubble: How the new personalized web is changing what we read and how we think. Penguin.
[2] Amatriain, X., & Basilico, J. (2012). Netflix recommendations: Beyond the 5 stars (part 1). Netflix Tech Blog, 6. http://techblog.netflix.com/2012/04/netflixrecommendations-beyond-5-stars.html. visited on 2023-23-02.
[3] Marshall, M. (2006). Aggregate Knowledge raises $5 M from Kleiner, on a roll. Venture Beat. http://venturebeat.com/2006/12/10/aggregateknowledge-raises-5m-from-kleiner-on-a-roll/. visited on 2023-23-02.
[4] Ferrari Dacrema, M., Cremonesi, P., & Jannach, D. (2019, September). Are we really making much progress? A worrying analysis of recent neural recommendation approaches. In Proceedings of the 13th ACM conference on recommender systems (pp. 101-109).
[5] Nguyen, T. T., Hui, P. M., Harper, F. M., Terveen, L., & Konstan, J. A. (2014, April). Exploring the filter bubble: the effect of using recommender systems on content diversity. In Proceedings of the 23rd international conference on World wide web (pp. 677-686).
Do you want to break your display ads bubble? Are you interested in defining your customized ad campaign monitoring strategy?
Our cloud-based infrastructure is highly scalable and enables customized ad campaign monitoring at all levels, e.g., profiles, websites, locations, etc.
Get in touch
