In the recent years, the wide spreading of smart-phones that are daily carried by humans and fit with tens of sensors triggered an intense research activity in human activity recognition (HAR). HAR in smartphones is seen as essential not only to better understand human behavior in daily life but also for context provision to other applications in the smartphone. Many statistical and logical based models for on-line or off-line HAR have been designed, however, the current trend is to use deep-learning with neural network. These models need a high amount of data and, as most discriminative models, they are very sensitive to the imbalanced class problem. In this paper, we study different ways to deal with imbalanced data sets to improve accuracy of HAR with neural networks and introduce a new over-sampling method, called Border Limited Link SMOTE (BLL SMOTE) that improves the classification accuracy of Multi-Layer Perceptron (MLP) performances.