Emotion Recognition is one of the vastly studied areas of affective computing. Attempts have been made to design emotion recognition systems for everyday settings. The ubiquitous nature of Intelligent voice assistants (IVAs) in households, make them a great anchor for the introduction of emotion recognition technology to consumers. The existing systems lack such pipelines and rely on dictionary-based architectures in their design. Further, these systems lack conversational properties and are merely an extension of information retrieval engines.

In this setting, we propose to introduce and develop emotion recognition pipelines that are suited to the interactions, common with these IVAs. To augment the existing emotion recognition pipelines which rely on audio information, we look at physiological information derived from wearables. Our proposed model uses multimodal embeddings with a Siamese Network to achieve the task of emotion recognition from a few samples. Physiological signals of blood volume pulse (BVP) and electrodermal activity (EDA) are used as additional input embeddings to two audio embeddings arising from the speech samples. We employ the state-of-the-art training schedules for Siamese Networks, which use a very limited amount of training on support datasets via sample pair comparisons. The performance of the model is evaluated using weighted binary accuracy and f1 scores.

The proposed model is applied on two datasets that denote two unique experimental settings - the K-EmoCon dataset and RECOLA dataset. We demonstrate an improvement in the state-of-the-art accuracy with the K-EmoCon dataset with accuracies of 63.97% and 66.91% on arousal and valence dimensions respectively. Further, on the RECOLA dataset, the model performs moderately well with 53.81% and 53.87% respectively for arousal and valence dimensions. In addition to this, we present a study of the effects of variation of available support set for training from the dataset. We make some salient observations for these experiments across individual participants and also identify how the label distributions affect the performance of the model. Further, we investigate the impact of real-world noise samples from the DEMAND dataset on the two datasets. We observe that the proposed model is robust and performs sustainingly well even in the presence of imputed noise.

Communication is one of the main challenges in Human-Agent Teams (HATs). An important aspect of communication in HATs is the use of explanation styles. This thesis examines the influence of an explainable agent adapting its explanation style to a supervising human team leader on team performance, trust, situation awareness, collaborative fluency, explanation satisfaction, understandability, and user-awareness. To perform a simulated Search And Rescue (SAR) task, a HAT is designed. With this design, a pre-study is then conducted using a questionnaire to discover the best-ranked explanation styles in the most important situations of the SAR task. Next, the user-study is carried out with 46 participants, using the HAT design and analysed data from the pre-study. There are two conditions: the agent adapting the explanation style to the human team leader and the baseline condition where the explanation style is randomised. The results show that the subjective measurements of trust, understandability, explanation satisfaction, and perceived user-awareness are significantly higher in the adaptive agent condition group. The same cannot be concluded for objective measurements such as team performance and situation awareness.

Social VR is the application of virtual reality that supports remote social interaction in virtual spaces. Users communicate and interact with others in the social VR environment through avatars, which are virtual anthropomorphic characters that aim to represent humans in virtual worlds. In addition, the development of the HMD and commercially available motion capture systems enable the avatars in the virtual environment to detect and reflect the real-time motions, even facial expressions of people. However, the avatars still lack an indication of biofeedback - e.g., body temperature, breathing, heart rate, muscle contraction -, which serves as social cues for communication in reality. While some features, for example, emojis, supports users to express their feeling or emotions for richer communication, the missing information often results in miscommunication in the virtual space. It remains a barrier to a fully immersed experience in the social VR space. This project proposes a concept of visualizing biosignals of the avatars in the social virtual reality space for a richer-level interaction in virtual reality. With the technologies available to capture and reflect accurate biofeedback in real-time, we would like to explore ways and possibilities to map the bio states of the users in reality to avatars in the virtual world. The project starts with conducting user researches to understand the current user behaviors in the social VR spaces and their perspectives on sharing biosignals. Based on the requirements gathered from the user study, the scope of the project is narrowed down to a ‘watching entertainment’ scenario, and the ways to visualize biosignals on avatars were explored through a co-design session with designers. After that, four biosignal visualization techniques in two biosignals - heart rate and breathing rate - are prototyped under the VR jazz bar setting. Finally, the user study is conducted with 16 pairs (32 participants in total) to test and compare the effects of each biosignal visualization technique in watching entertainment scenarios with a companion. As a result, the embodied visualizations are the most understandable and least distracting visualization method among the four methods. Furthermore, the limitations of the research, recommendations on biosignal visualizations, and recommendations on conducting design research are provided.

Recommender Systems assist the user by suggesting items to be consumed based on the user's history. The topic of diversity in recommendation gained momentum in recent years as additional criterion besides recommendation accuracy, to improve user satisfaction. Accuracy and diversity in recommender systems coexist in a delicate trade-off due to the complexity in capturing user tastes through a limited amount of interactions. Graphs have been employed for recommendation, given their ability to efficiently represent user-item interactions. Graph convolutions, as learning over graphs tools, have reached state-of-the-art accuracy on recommender system benchmarks. However, the potential of graph convolutions to improve the accuracy-diversity trade-off is unexplored. Here, we develop a model that learns from a nearest neighbor and a furthest neighbor graph via a joint convolutional model to establish a novel accuracy-diversity trade-off for recommender systems. In detail, the nearest neighbor graph connects entities (users or items) based on their similarities and is responsible for improving accuracy, while the furthest neighbor graph connects entities based on their dissimilarities and is responsible for diversifying recommendations. The information between the two convolutional modules is balanced already in the training phase through a regularizer inspired by multi-kernel learning. Numerical experiments on three benchmark datasets showed the joint convolutional model can improve substantially the catalog coverage or the diversity among recommended items; or boost both by a lesser amount. We compared our model against state-of-the-art accuracy-oriented algorithms, showing diversity gains up to seven times by trading as little as 1\% in accuracy. We also compared the joint model against algorithms proposing a different accuracy-diversity trade-off, evidencing our model achieves better accuracy while preserving a wide diversity range. Our findings highlight that the joint convolutional model offers a balance in each setting that is difficult to be achieved with a single model.

Small and medium enterprises (SMEs) bring a significant contribution to each country’s economy, ensuring both a high employment rate and financial prosperity. Despite their essential role, these type of companies presents a higher vulnerability to default than large corporates. The default event implies that the SME could not properly reimburse the money owned to its suppliers. If their default could be forecast, experts could adopt measures in order to prevent it or its impactful consequences. For this reason, developing a credit scoring model which is able to predict which SMEs are endangered of default is crucial. There are plenty of credit scoring prediction models available in the literature. However, most of them are only relying on the financial status of one company. In this work, we are presenting a novel method for credit scoring prediction which does not only take into account the SMEs’ financial situation, but also their position and role within a transactional network. A transactional network is a graph, whereby the nodes are represented by SMEs and the edges show that between two nodes there should be at least one transaction. In our work, we highlight the limitations that traditional models face and provide an alternative to overcome them. Furthermore, our findings show that combining network features with financial features could lead to a more accurate prediction and increase the robustness of the model. For this reason, we believe that the transactional network carries significant insights and could be a meaningful addition to financial based credit scoring prediction models.

Keeping a steady cash flow is one of the biggest if not the biggest problem that Small to Medium Enterprises (SMEs) deal with daily. Within the different types of cash flow, Accounts Receivable (AR) classifies the balance of money that needs to be paid by the company's customers. In the most typical case, after receiving goods or services, the customer receives an invoice with the amount that is owed to the supplier. However, this often does not happen before the aforementioned date, meaning that the invoice is often paid late. Intervention requires resources and over-intervention could cause unwanted customer dissatisfaction. Knowing whether an invoice is going to be paid late can be vital information. Current methods of late payment prediction focus only on the history between the seller and the buyer and are unusable when this history is not present. Intuitively, one's business depends on the relationships and transactions that it has with its neighbors. Suggesting that neighbor behavior could be useful when predicting the cash flow of a company. Unfortunately, this type of information is not always given and needs to be data mining from non-relational data. This work presents a method for building a relational network of SMEs using entity resolution and improving the current state of the art of late payment prediction using features extracted from the graph.

The use of interactive media applications is becoming more common. With the current advancements in technologies, applications are becoming more advanced and research is being performed into adding more modalities to these applications to enrich the experience and increase their capabilities and uses.
he performance of these applications depends heavily on the network resources available to them. Through the use of congestion control algorithms, these applications are able to get a fair share of the available bandwidth. However,
congestion control algorithms are unable to minimize the delay caused by competing traffic. By combining the capabilities of OpenFlow with a priority queueing setup, we showed that we can provide low latency to a subset of flows. To control the utilization of the network resources and the effect the prioritization has on best-effort traffic, we use OpenFlow to control the bit rates of the high-priority flows. We show that by adjusting the ECN marking ratios of flows we are able to accurately control the bitrate of each of them.

Representing geometry data as voxels allows for a massive amount of detail that can be rendered in real-time. Storing this type of data as a directed acyclic graph (DAG) has recently led to immense improvements in memory consumption, which is one of the main limitations often associated with voxel-based approaches. We present a method for further decreasing the memory consumption of voxel data in a DAG through a form of lossy compression, meaning that the data is slightly altered from its original state. Our method clusters similar nodes in the graph together in a way that minimizes the geometric error that is introduced. The amount of compression that is applied can be influenced through a set of parameters that affect which nodes are included in the clustering process and the granularity of the clusters that are detected. The memory consumption of the sparse voxel DAG is reduced from 10% to up to 50%, depending on the characteristics of the scene, while introducing errors that are only visible when viewing the scene from up close.
Our method is complementary to other state of the art compression methods for this data structure, and has a negligible effect on real-time rendering performance.

In recent years, there has been a widespread increase in the adoption of virtual reality and 3d displays which require many images as input, e.g. head mounted displays or lightfield displays. Real-time rendered imagery for such displays is commonly computed using brute-force rendering pipelines based on forward(+) or deferred rendering pipelines. This process can be sped up based on the coherence between images, and is the aim of the presented study. The result is a proof-of-concept framework meant as alternative for brute force generation of each image, whereby the developed algorithms interpolate from a set of source images based on the theory of epipolar geometry. This theory describes a relation of a surface point in 3d space between different images.

Besides advantages as better performance compared to the baseline, interpolation-based image generation also has disadvantages. Specific to the presented use case is the need for special support of commonly used view dependent properties. As such, support for specularity is demonstrated based on inclusion of the Phong reflection model in the framework.

Additionally, several techniques are included to improve the performance of the basic interpolation algorithm. Thus, further reducing the frame times of the developed interpolation-based rendering pipelines compared to the baseline rendering pipeline, which for this study is a deferred rendering pipeline. Although, the increased performance results in decreased image quality, the resulting images are of acceptable quality when compared to the baseline. All aspects considered, the results show a rendering framework based on epipolar geometry is viable, but the current implementation leaves many features used in production rendering pipelines to be ported. Moreover, the current framework implementation is built upon techniques that may have a lot of room for improvement.

Online advanced media streaming services using HTTP adaptive stream-
ing are increasingly popular. However in practice, the multi-protocol, multi-
format nature of adaptive streaming creates a lot of engineering eort and
costs for the operators, in the storage and preparation of the dierent
formats. In this work, we acknowledge these issues and we study a stream-
ing setup that can address these. Such streaming setup consist of a backend
cloud storage and a processing node that generates streaming presentations
for different devices on-the-fly. We analyse the streaming setup and its
performance by testing in cloud deployments. Through this evaluation, we
identify the performance limitations of the setup, imposed by the transfer
of data between the object storage and the processing node. We propose a
new backend storage caching scheme, based on rarely used existing feature
of dref in the specication of the MPEG-4 standard. Experimental results
show that the proposed scheme can improve the streaming performance,
such as reduced latency and increased outgoing traffic volume towards the
clients