Tag Archives: recommender system

“The Filter Bubble: exploring the effects of using recommender systems on content diversity” (Nguyen et al. 2014)

The article chosen addresses the so-called bubble effect identified by Pariser (2012). This bubble effect suggests that by using recommender systems (RS), users are exposed to only a few products that they will like and miss out of many others.  The paper wants to investigate this through understanding the content diversity at an individual level provided by collaborative filtering. It suggests to be the first study observing the effects of this phenomenon on an individual level.

From the study conducted by Lee and Hosanagar (2015) we have understood that there are many opposing views existing in the literature on content diversity on an individual level. Therefore, as the current article claims to be the first one studying this phenomenon on an individual level it is interesting to see how the study has been conducted and what their conclusions are compared to the article of 2015.

The paper addresses very well the debate in regard of the bubble effect: whether recommender systems may be harmful to users. First the behavioral aspects of people who are exposed to similar content and what the effect is on their individual behavior is addressed. As they want to measure the effects on an individual level it is important to recognize what has been found in regard of individual behaviors on content exposure.

For this study, they use the long-term users of MovieLens as they need longitudinal data to draw conclusions on the user’s behavior over time. Two research questions are addressed; 1. Do recommender systems expose users to narrower content over time? 2. How does the experience of users using recommender systems differ of those who do not rely on recommender systems?

The article uses the “tag genome” developed by Vig et al. (2012) to analyze the diversity of the movies that are recommended and consumed (rated). This appears to be a strong measure as it identifies the content of the movie and identifies the similarities content-wise. Multiple articles have used the movie genres (Lee and Hosanagar, 2015) or the ratings given (Adamopoulos and Tuzhilin, 2014) to identify similarities which seems to be less generalizable as the content of the movies still can vary greatly when using these metrics.

The article describes clearly how the findings should be interpreted and addresses multiple questions that have risen based on the findings. This leads to a well-rounded study where the effect of item-item collaborative filtering is exposed. First, the article addresses whether recommender systems expose users to a narrower content over time through comparing the content diversity at the beginning with the content diversity at the end of the user’s rating history. This comparison can show the development of content consumption of the user over time. It has been found that the content diversity of both user groups (using RS or not to rate movies) becomes quite similar over time. Furthermore, it identifies whether using RS reduces the total content-diversity consumed of that user. The conclusion is that users using RS over time consume more diverse content then users ignoring RS. Finally, the experience of the two user groups are evaluated and it is observed that users using RS seem to consume more enjoyable movies based on their ratings given.

As the limitations of the article suggest itself it would be interesting to study the phenomenon in a more experimental setting where the behavior of users can be observed in more detail. This would help in understanding the reasons of the decisions made by the users based on the recommendations. The multiple studies conducted in the field of RS mostly focus on collaborative filtering as this RS is the most commonly used (Lee and Hosanagar, 2015) but research should also focus on other recommender systems to make sure that those used benefit the user the most.



Adamopoulos, P. and Tuzhilin, A., 2014, October. On over-specialization and concentration bias of recommendations: Probabilistic neighborhood selection in collaborative filtering systems. In Proceedings of the 8th ACM Conference on Recommender systems (pp. 153-160). ACM. http://dl.acm.org/citation.cfm?id=2645752

Lee, D. and Hosanagar, K., 2015. ‘People Who Liked This Study Also Liked’: An Empirical Investigation of the Impact of Recommender Systems on Sales Diversity.https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2603361

Nguyen, T.T., Hui, P.M., Harper, F.M., Terveen, L. and 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). ACM. http://dl.acm.org/citation.cfm?id=2568012

Are cross-platform social recommender systems the future?

(This blog post is based on the research article ‘A social recommender mechanism for e-commerce: Combining similarity, trust, and relationship’ by Li, Wu and Lai, 2013)

In a world where there is not only a massive selection of different products but also the internet that enables us to theoretically choose among all those offerings without the high search costs that have hindered us from an informed choice in the past, the problem of information overload is a challenge for both the consumers and the companies offering those products (Li, Kauffman, Van Heck, Vervest and Dellaert, 2014). As explained by Murray and Häubl (2009), especially in the e-commerce environment, good recommender systems (RS) that aid the consumer in finding the right product are becoming increasingly important, ensuring a higher consumer satisfaction and increased sales for the offering companies (Li, Wu and Lai, 2013). Even though good RSs exist in the market (such as Amazon’s collaborative item-to-item filtering mechanism or Netflix’s hybrid model of collaborative and content based filtering (Jones, 2013)), it seems that today’s RSs fall behind what should be technically possible. Thus, Li et al. (2013) suggest a new RS that does not only take a customer’s past behavior into account but adds a social component that would greatly increase the information available to the system and thus improve its prediction accuracy. An important drawback of today’s RSs mentioned by Li et al. (2013) is that the platforms utilizing RSs are independently operated and only use the data obtained within the boundaries of their respective platform. Real value, however, could be obtained when integrating the data of various different platforms and adding the social component of the social network of a consumer to the recommendations shown on an e-commerce platform. In real life, after all, we also tend to ask our friends for advice when shopping and especially in cliques of close friends, the shopping behavior of individuals potentially influences the shopping behavior of the others (Li et al., 2013; Shang, Hui, Kulkarni and Cuff, 2011). Continue reading Are cross-platform social recommender systems the future?