Abstract-In this paper, a noise injection method to improve personal recognition models is presented. The idea of the method is to build more general recognition models for eHealth using a small original data set and by expanding the area covered by training data using noise injection. This way, it is possible to train models that are less vulnerable to changing conditions, and thus more accurate, but still the data gathering phase can be non-burdensome. The proposed method was tested using two classifiers (linear discriminant analysis and quadratic discriminant analysis) and three human activity recognition data sets collected using inertial sensors of a smartphone. Two of these data sets are open data sets. The results show that noise injection improves the true positive recognition rates. With first data set the improvement varies from 1.3 to 2.0 percentage units, with second from 1.4 to 4.5 percentage units, and with third the highest improvement was 2.5 percentage units. Moreover, the results show that the method improves precision and reduces false positive rates. In addition, experiments were made using different training set sizes to show that the improvement in true positive rate is bigger if the original training data set is small. In this study, the method is experimented using human activity data sets but it is not limited to this application area and can be used with any time series data.Index Terms-Activity recognition; eHealth; noise injection; smartphone; inertial sensors; accelerometer; personal models;
I. PROBLEM STATEMENT AND RELATED WORKIn eHealth it is often important to recognize what a person is doing or a condition of a person. However, people are different, and therefore, a user-independent recognition model that provides accurate recognition results for one person does not necessary provide as high results for other person. There are many possible reasons for this phenomena. For instance, person's can be different: physical characteristics, health state or gender can have an effect to the recognition results. On the other hand, external factor such as weather, terrain and location can cause problems to recognition models. In addition, in the real-life many other unseen contingencies can happen and the training data set used to train the recognition models cannot include all of these. This means that model that seem to work really well when tested with testing data do not work as well when it is used online, in real-time, real-life applications [1].In the case of human activity recognition, often the recognition is done using user-independent models and good recognition rates have been achieved (for instance [2], [3], [4], [5]). However, it has been shown that user-independent models do not work accurately for instance if trained with healthy study subjects and tested with subjects who have difficulties to move [6]. Personal models could solve this problem. In fact, it has been shown that user-dependent models are more accurate than user-independent ones [7]. However, in [8] it was shown...