| Summary: | The increasing number of small, cheap devices, full of sensing capabilities lead to an untapped source of data that can be explored to improve and optimize multiple systems, from small-scale home automation to large-scale applications such as agriculture monitoring, traffic flow and industrial maintenance prediction. Yet, hand in hand with this growth, goes the increasing difficulty to collect, store and organize all these new data. The lack of standard context representation schemes is one of the main struggles in this area. Furthermore, conventional methods for extracting knowledge from data rely on standard representations or a priori relations. These a priori relations add latent information to the underlying model, in the form of context representation schemes, table relations, or even ontologies. Nonetheless, these relations are created and maintained by human users. While feasible for small-scale scenarios or specific areas, this becomes increasingly difficult to maintain when considering the potential dimension of IoT and M2M scenarios. This thesis addresses the problem of storing and organizing context information from IoT/M2M scenarios in a meaningful way, without imposing a representation scheme or requiring a priori relations. This work proposes a d-dimension organization model, which was optimized for IoT/M2M data. The model relies on machine learning features to identify similar context sources. These features are then used to learn relations between data sources automatically, providing the foundations for automatic knowledge extraction, where machine learning, or even conventional methods, can rely upon to extract knowledge on a potentially relevant dataset. During this work, two different machine learning techniques were tackled: semantic and stream similarity. Semantic similarity estimates the similarity between concepts (in textual form). This thesis proposes an unsupervised learning method for semantic features based on distributional profiles, without requiring any specific corpus. This allows the organizational model to organize data based on concept similarity instead of string matching. Another advantage is that the learning method does not require input from users, making it ideal for massive IoT/M2M scenarios. Stream similarity metrics estimate the similarity between two streams of data. Although these methods have been extensively researched for DNA sequencing, they commonly rely on variants of the longest common sub-sequence. This PhD proposes a generative model for stream characterization, specially optimized for IoT/M2M data. The model can be used to generate statistically significant data’s streams and estimate the similarity between streams. This is then used by the context organization model to identify context sources with similar stream patterns. The work proposed in this thesis was extensively discussed, developed and published in several international publications. The multiple contributions in projects and collaborations with fellow colleagues, where parts of the work developed were used successfully, support the claim that although the context organization model (and subsequent similarity features) were optimized for IoT/M2M data, they can potentially be extended to deal with any kind of context information in a wide array of applications.
|
|---|