The performance of TV recommender system which uses machine learning techniques is degraded due to imbalanced distribution of collected viewing preferences. As users have a tendency to provide positive feedbacks much more than the negative ones, the system that does not use methods to deal with class imbalance provides poor prediction of the contents that the user does not like to watch. Thus undesirable contents can often be recommended, which is perceived by users as bad recommendation. The probability of bad recommendations can be significantly decreased if the information about class imbalance is incorporated into the machine learning algorithm. However, this improvement comes at expense of degraded prediction of contents that user likes to watch; thus the quality of recommendations is decreased. In addition to learning algorithm, the choice of performance metric that is maximized influences user perception of the system performance. In this paper, it is shown that using the adjusted G-mean instead of G-mean metric can increase the quality of recommendations provided by the system based on neural network, without significant increase in the probability of bad recommendations. This further results in the increase of the recommendation diversity and, consequently, in the increase of user satisfaction.