The E-AT model assumes that each group member has an equal probability to be a representative (see Eqn. (2)). In this section, we argue that group members have dier-ent impacts in group item selection decisions. Hence, we introduce the notion of personal impact parameter to model this dierence. Accordingly, we propose a personal impact topic (PIT) model by incorporating the personal impacts to control the generation of group representatives. We also pro-vide the model learning algorithm to infer personal impacts, topic distribution of users and item distributions of topics. Finally, to over-come the over-tting problem when input dataset is sparse, we extend the PIT model by exploring additional social features.