Most conventional learning algorithms require both positive and negative training data for achieving accurate classification results. However, the problem of learning classifiers from only positive data arises in many applications where negative data are too costly, difficult to obtain, or not available at all. This paper describes a new machine learning approach, called ILoNDF (Incremental data-driven Learning of Novelty Detector Filter). The approach is inspired by novelty detection theory and its learning method, which typically requires only examples from one class to learn a model. One advantage of ILoNDF is the ability of its generative learning to capture the intrinsic characteristics of the training data by continuously integrating the information relating to the relative frequencies of the features of training data and their co-occurrence dependencies. This makes ILoNDF rather stable and less sensitive to noisy features which may be present in the representation of the positive data. In addition, ILoNDF does not require extensive computational resources since it operates on-line without repeated training, and no parameters need to be tuned. In this study we mainly focus on the robustness of ILoNDF in dealing with high-dimensional noisy data and we investigate the variation of its performance depending on the amount of data available for training. To make our study comparable to previous studies, we investigate four common methods: PCA residuals, Hotelling's T2 test, an auto-associative neural network, and a one-class version of the SVM classifier (lately a favored method for one-class classification). Experiments are conducted on two real-world text corpora: Reuters and WebKB. Results show that ILoNDF tends to be more robust, is less affected by initial settings, and consistently outperforms the other methods.