Spanish version / Versión en castellano: Descarga
Recommender Systems Problems and Strategies
The fundamental purpose of any recommendation system is to filter information according to user preferences. In the context of the project eTUR2020: Tourism & Retail, it is about offering the most relevant products and services to the user according to their needs and contextual circumstances.
One of the great problems of any recommendation system is the lack of data to work effectively and generate relevant recommendations to the user:
- The problem of cold start and first raters. When the system doesn’t have enough users, when it boots for the first time, when a new user signs into the system, when a new product is introduced, etc. there isn’t enough data to generate relevant recommendations.
- The dispersion of the data. This problem doesn’t only exist during the previous problem, but even when having enough users they value very few of the products available.
To place ourselves, there are several approaches to recommender systems and the type of user feedback they need. Although there are other, the more usual ones that require end-user collaboration are (Ricci, Rokach, & Shapira, 2011):
- Strategies based on collaborative filtering, focusing on ratings of products by users. While there are implicit ways to collect user feedback,the problem in which we are interested is primarily to encourage the user to explicitly rate catalog products using likes, numerical scales or other social mechanics and then match users to start recommending.
- Strategies based on content, focusing on product descriptions and ratings of each user to generate a classification system of items interesting or not for that user. In this case the problem is to perform a preliminary process of description using a taxonomy, or encourage users to communally do this exercise, labeling each product and creating a folksonomy which subsequently serves to make recommendations. Then it is also necessary that the user rate the items to start producing recommendations.
- Strategies based on demographics, focused on knowing the characteristics of demographic groups, i.e. nationality, age, gender, income, etc. In this case the strategy is to motivate users to complete their profile data.
- Hybrid strategies, combining in various ways the best features of each of the above strategies.
However, recently, thanks to the widespread use of mobile devices, contextual factors have gained importance in recommender systems. To the usual two-dimensions (user and item) these adds the context of the user: place, time, weather, number of companions, attempted purchases, user relevant dates, etc. (Adomavicius & Tuzhilin, 2008). The project eTUR2020: Tourism & Retail combines classical approaches introducing this innovation as part of the recommendation system.
The involvement of end users is one of the most important factors in the effectiveness of recommender algorithms. That is why gamification techniques represent interesting solutions for users to transparently and voluntarily collaborate for recommender systems to deliver more relevant results.
We are facing a case where engineering achieves a high degree of development but runs into a social barrier: the user does not cooperate, for whatever reason, and he has to be motivated to do so. No matter how good the algorithm is, without user feedback, it hardly operates effectively.
There is considerable literature on user feedback motivation and although not explicitly using the term gamification, in some cases the mechanisms resonate a lot. As an example, the work of Rashid et al. (2006), prior to the appearance of the term gamification, or the later work of Farzan & Brusilovsky (2011).
Strangely, we have identified only two indexed papers that include in their title the terms gamification and recommendation systems. In general, in those articles we observe:
- A significant and rapid increase in user participation which translates into more i.e. product ratings and reduction of the cold start. However, there has to be more experiments to understand in isolation how each mechanic affects the overall effect with the ultimate goal of determining useful patterns in each context (Feil et al., 2016).
- There is evidence about the risks of poor design and not taking into account collateral psychological effects that affect the extrinsic and intrinsic motivation. For example the use of incentives based on personal needs can motivate individual self-deception and cause positive bias in the ratings (Farzan & Brusilovsky, 2011).
- Users are often motivated to rate items when they do not meet their expectations (positive or negative). This is problematic because the recommendation system must operate with extreme ratings (Ziesemer et al., 2014).
- By using tangible rewards (extrinsic), users are more likely to return to the portal and further rate items. However, users can also do it intrinsically motivated when they understand that it is important to do and that it is “right” for the benefit of the group (Rashid et al., 2006; Ziesemer et al, 2014.).
This small study represents the beginning of a task in project eTUR2020: Tourism & Retail in which new strategies in collecting feedback will be studied. As you can see, the strategies are not always about improving algorithms, but to get the involvement and collaboration of end users to improve the system. Another example of innovation and co-creation in the era of open innovation.
Adomavicius, G., & Tuzhilin, A. (2008). Context-aware recommender systems. In Proceedings of the 2008 ACM conference on Recommender systems – RecSys ’08 (p. 335). New York, New York, USA: ACM Press. doi:10.1145/1454008.1454068
Farzan, R., & Brusilovsky, P. (2011). Encouraging user participation in a course recommender system: An impact on user behavior. Computers in Human Behavior, 27(1), 276–284. doi:10.1016/j.chb.2010.08.005
Feil, S., Kretzer, M. E., Werder, K., & Maedche, A. (2016). Using Gamification to Tackle the Cold-Start Problem in Recommender Systems. In Proceedings of the 19th ACM Conference on Computer Supported Cooperative Work and Social Computing Companion – CSCW ’16 Companion (pp. 253–256). New York, New York, USA: ACM Press. doi:10.1145/2818052.2869079
Rashid, A. M., Ling, K., Tassone, R. D., Resnick, P., Kraut, R., & Riedl, J. (2006). Motivating participation by displaying the value of contribution. In Proceedings of the SIGCHI conference on Human Factors in computing systems – CHI ’06 (p. 955). New York, New York, USA: ACM Press. doi:10.1145/1124772.1124915
Ricci, F., Rokach, L., & Shapira, B. (2011). Introduction to Recommender Systems Handbook. In Recommender Systems Handbook (pp. 1–35). Boston, MA: Springer US. doi:10.1007/978-0-387-85820-3_1
Ziesemer, A. de C. A., Müller, L., & Silveira, M. S. (2014). Just Rate It! Gamification as Part of Recommendation. In M. Kurosu (Ed.), 16th International Conference, HCI International 2014 (pp. 786–796). Heraklion, Crete, Greece. doi:10.1007/978-3-319-07227-2_75