Intro
Recommendation systems are widely applied in a variety of domains in our life, such as entertainment, e-commerce and social media. By screening massive amounts of data, the system delivers value and convenience to users. Yet, have you ever had such an experience when your recommender keeps pushing ads of very similar products, or the songs in your playlist are in a similar musical style which you might have been fed up with?
More than user experience
This is not just concerned with user experience, but also with economic effect and development. As one of the side effects, popularity bias is a common type of recommendation bias. It has been observed that under some conditions popular items are more likely to be recommended leading to a rich get richer effect, exacerbating the unequal distribution of social resources. Likewise, “information cocoons” is another type of bias that may happen due to the closed-loop attribute of recommendation algorithm.
Solutions?
Combination of multiple recommendation strategies seems to be a possible solution. Take Spotify as an example. As a music streaming service provider, its recommendation system Discover Weekly and algorithm are well designed, giving Spotify an advantage over other competitors. The system is able to recommend songs that suit your taste but sound different. It is also able to avoid popularity bias and bring attention to new songs or niche musical works. Instead of simply relying on one recommendation strategy (or model), the combination of several recommendation strategies (or models) plays an important role in the success. Here are the three types of models that Spotify employs:
1.Collaborative Filtering models, which analyse both your behaviour and others’ behaviours
2.Natural Language Processing (NLP) models, which analyse text
3.Audio models, which analyse the raw audio tracks themselves
Another possible solution is to add a post-processing algorithm to adjust and re-rank the set of the recommendations. For post-processing, the key point is to eliminate bias but meantime keep the utility of recommendations as high as possible.
References:
Abdollahpouri, H., Burke, R. and Mobasher, B., 2019. Managing Popularity Bias in Recommender Systems with Personalized Re-ranking. arXiv:1901.07555 [cs]. [online] Available at: <http://arxiv.org/abs/1901.07555> [Accessed 23 Sep. 2019].
Ciocca, S., 2018. How Does Spotify Know You So Well?[online] Medium. Available at: <https://medium.com/s/story/spotifys-discover-weekly-how-machine-learning-finds-your-new-music-19a41ab76efe> [Accessed 20 Sep. 2019].
Farnadi, G., Kouki, P., Thompson, S.K., Srinivasan, S. and Getoor, L., 2018. A Fairness-aware Hybrid Recommender System. arXiv:1809.09030 [cs, stat]. [online] Available at: <http://arxiv.org/abs/1809.09030> [Accessed 23 Sep. 2019].
Tsintzou, V., Pitoura, E. and Tsaparas, P., 2018. Bias Disparity in Recommendation Systems. arXiv:1811.01461 [cs]. [online] Available at: <http://arxiv.org/abs/1811.01461> [Accessed 20 Sep. 2019].
