The Maastricht Study is an extensive study of type 2 diabetes (10.000 participants). During the testing period of participants are, among other tests, equipped with an activity monitor for one whole week. This device will track whether participants are walking, sedentary or standing. This thesis is about discovering whether there are different ways in which people are active when they have type 2 diabetes, taking into account the duration, intensity, frequency and time of the day and week, which is currently done with summarized daily data at the Maastricht Study. Because there is more detailed data available, for example ‘raw’ data with the sensors’ X, Y & Z coordinates sampled at 20Hz and lists of every new behavioural state of a participants, there is still a lot of potential for new insights in activity patterns with the current dataset.

Analysis will be performed using pattern recognition techniques on our data set. Firstly, the data will be restructured using barcoding and symbolic analysis. Barcoding is an algorithm which enables you to expand our current three states to more detailed states, which also takes into account if they have a long or short duration. Symbolic analysis is an algorithm that can be used to find main patterns in data. A combination of barcoding and symbolic analysis could provide insights in more intrinsic patterns than symbolic analysis without barcoding because of the higher level of detail. This level of detail is mainly wanted in this research because one of the things to look into is the patterns in spread of activity across the day, which can be visualized more easily with the abovementioned approach. With the end product it will be possible to express activity patterns in (multiple) variables to give an accurate indication of different movement pattern types per group of people based on the values of these variables.

Finding semantically rich and computer-understandable representations for textual dialogues, utterances and words is crucial for dialogue systems or conversational agents, as their performance is mostly depending on the understanding of the previous conversation. The path the research is currently taking is finding distributed vector representations (embeddings) with one ultimate goal: the more semantically similar two representations are, the closer they shall be in the vector-space. Encoding the "meaning" of text into vectors is a current trend and it continues from words, phrases and documents to actual human-to-human conversations. This form of representation allows to apply many machine learning techniques that previously performed poorly when encountering textual data.

In this thesis the utilization of embeddings for answer retrieval is explored and a novel approach is proposed of how to incorporate vector representations and several answer proposing models into Monte-Carlo Tree Search (MCTS) to represent states and actions of the search tree and simulate entire conversations.

The goal of the thesis is to develop a recommender system for e-commerce data. Most research activity bases on explicit feedback data, i.e. user intentionally give feedback in the form of ratings to express their preferences for different products. Online shops in contrast only provide implicit feedback. Users do not actively express their preferences, but they have to be inferred from their purchase history or browsing behaviour. Such kind of data is more noisy than explicit feedback data and it lacks negative feedback. Not purchased or viewed items can be either unknown or disliked by the user. Different collaborative filtering and content-based models are investigated and evaluated to find a model that is suitable to compute accurate recommendations in real time. Moreover different hybrid recommender systems are examined to improve prediction performance and ensure that the recommender system can be adapted to different online shops.

Visual attention mechanisms in humans account for the limited capacity of the brain to process sensory information. Through selection of regions of interest within the visual field, these processes reduce task complexity to enable rapid scene analysis. A two-fold approach towards attention models has been proposed that incorporates intrinsic features of the input data (bottom-up) and volitional task-dependent control (top-down) to guide future eye movements. This thesis attempts to integrate computational models of bottom-up and top-down attention into a common deep learning framework based on recurrent neural networks. Unlike previous work in machine learning literature, the algorithm will be designed to match experimental eye tracking data for scene analysis in order to gain potential insights into brain functions and narrow the performance gap between artificial and human visual systems.