In the article "Inspectability and Control in Social Recommenders. Knijnenburg, B., Bostandjiev, S., O'Donovan, J., and Kobsa, A. (2012)" the authors study the influence of inspectability and control over subjective system aspects and user experience of social recommenders. In order to study this phenomena, TasteWeights system was designed and implemented on an online user experiment with 267 test subjects.
TasteWeights is a Facebook music recommender system that gives users control over the recommendations, and explains how they came about. The algorithm assigns weights to the user's friends by the overlap in music 'likes' between them calculated with Pearson's correlation coefficient. Afterwards, the algorithm assigns weights to the each friends' music items by the sum of the weights of all friends that liked that particular item. The recommendations are displayed in decreasing order of recommendation weight. TasteWeights has controlability due to the fact that it allows users to adjust the weights of the items and friends as they desire in order to get the best recommendations. The system has inspectability by implementing visualizations of the recommendations giving the user information about how the algorithm came up with those results. There are two forms of visualization: the most complete one is a full-graph that connects specific friends to the items that were recommended so that the user knows which friend influenced which recommendations. The second one is a simpler and it corresponds to only a list of the recommendations in which the score of each item is encoded in the length of a bar. This allows the user to compare scores between items. This last feature is contained in the full-graph as the list-only view is a portion of the full-graph.
Some of the results of the online experiment showed that the inspectability and control manipulations have a positive effect on understandability. Understandability has a positive effect on perceived control and the latter on satisfaction with the system. Inspectability positively associates with a higher number of known recommendations. The number of recommendations has a smaller direct negative effect over the satisfaction with the system and a higher undirect positive effect over it through higher perceived recommendation quality and control.
Overall the article is solid. First it presents the topic of recommender systems focusing rightaway on inspectability and control using as social recommenders as the domain for the examples. Then in the Related Work section, the authors focus on those two characteristics, user experience and personal characteristics. The purpose of recalling previous work on user experience is related to the fact that the study focuses on analyzing the impact of insepectability and control over user experience. Personal characteristics are also included as they aknowledge the influence of it on the results obtained in their framework. The algorithm could have been skipped by the authors but it was explained focused on how the weights are assigned to friends and items. This aids comprehension on the study of effects of item-control vs. friend-control vs. no-control.
TasteWeights is a good tool for the experiments as it only has two features that independently correspond to each study variable (inspectability and control). As it only recommends, shows visualizations and lets users tune it, any effect of the system on the user's experience will come from those three sources.
The best element of the article is the structural equation model for the data of the experiment. The authors managed to summarize the results of the experiments in one figure that is easy to understand. The attention to detail is remarkable as they were able to calculate how each aspect of the system, personal characteristics of the user, interacion and experience relates directly or undirectly between each other. The recollection of qualitative data via the questionnaire was analyzed quantitatively. The statistical rigourosity also is a key point on the results as they found out that even though there is a negative effect of known items over user satisfaction, the total effect of inspectability on it is statistically significant on a positive association.