Tailored to fit? Implicit and explicit user evaluations of algorithm-based mobile news

Since the up rise of mobile news consumption, audiences are offered abundant updates on current events, wherever and however they want. Still, this endless stream of information tends to become overwhelming, hence welcoming automatically learning recommendation algorithms to filter what is relevant for each individual user. In this study, we elaborate on the process and outcomes of a media innovation project, inquiring the value of such recommendations as assessed by a panel of 105 test users. In collaboration with a team of creative research engineers, a test environment was designed, logging each individual action with the mobile application. The designed app was continuously filled with branded news items, provided in real time by both commercial broadcasters’ and publishers’ newsrooms. Our experiment was based on three test conditions, with news updates based on (a) self-reported news category preferences, (b) a self-learning algorithm based on individual content consumption, (c) a self-learning algorithm that supplements the content consumption with contextual information (i.e. time of day, type of device). Per set of three item consumptions, each user was automatically prompted to assess whether the previous recommend item was either interesting of not (thumbs up or down). The ratio of these logged ratings functioned as dependent variable. After two weeks, the results indicate the content-based algorithm to outperform self-reported preferences and the context-based version, although preliminary data analysis suggests the latter to improve significantly over time. In conclusion this study, combining various types of implicit and explicit user data, offers a glimpse in the value of generating personalized news diets