TY - CONF
T1 - Improving On-line Genre-based Viewer Profiling
T2 - TVX2017 Workshop on In-Programme Personalisation for Broadcast (IPP4B)
AU - Veloso, Bruno
AU - Malheiro, B.
AU - Burguillo, Juan Carlos
AU - Foss, Jeremy
PY - 2017
Y1 - 2017
N2 - Typically, recommendation algorithms are unable to make recommendations for new users due to the inherent lack of information, i.e., the cold start problem. To overcome this problem, this work addresses builds new viewer profiles by combining general and personal feature-based profiles using both the frequency and the rating of each feature. For each newly arrived viewer, we create a dynamic profile by combining the corresponding demographic stereotypical profile with the individual profile and, then, as the number of the viewergenerated events increases, we gradually fade the general component and strengthen the individual component. Specifically, we combine the genre frequency & rating of the viewer personal and demographic stereotype profiles. This novel viewer profiling algorithm was evaluated with the MovieLens 100k and 1M data sets, using content-based and collaborative stream mining recommendation techniques. When compared with the standard average user stereotype, the results with the demographic stereotypes show a significant improvement in terms of classification accuracy, identical prediction accuracy and an increase in run time.
AB - Typically, recommendation algorithms are unable to make recommendations for new users due to the inherent lack of information, i.e., the cold start problem. To overcome this problem, this work addresses builds new viewer profiles by combining general and personal feature-based profiles using both the frequency and the rating of each feature. For each newly arrived viewer, we create a dynamic profile by combining the corresponding demographic stereotypical profile with the individual profile and, then, as the number of the viewergenerated events increases, we gradually fade the general component and strengthen the individual component. Specifically, we combine the genre frequency & rating of the viewer personal and demographic stereotype profiles. This novel viewer profiling algorithm was evaluated with the MovieLens 100k and 1M data sets, using content-based and collaborative stream mining recommendation techniques. When compared with the standard average user stereotype, the results with the demographic stereotypes show a significant improvement in terms of classification accuracy, identical prediction accuracy and an increase in run time.
M3 - Paper
ER -