CrowdEmotion produce software to measure a person's emotions based on analysis of microfacial expressions using a machine learning algorithm to recognize which features correspond with which emotions. The features are derived by applying a bank of Gabor filters to a set of frames. CrowdEmotion needed to improve the accuracy, processing speed and cost-efficiency of the tool. In particular they wanted to know if a subset of the bank of Gabor filters was sufficient, and whether the image filtering stage could be implemented on a GPU. A framework for choosing the optimum set of Gabor filters was established and ways of reducing the dimensionality of this were interrogated. Taking a subset of Local Binary Patterns was found to be fully justified. Meanwhile choosing a gridding pattern is open to interpretation; some suggestions were made about how this choice might be improved.

In this work we analyze traces of mobility and co-location among a group of nearly 1000 closely interacting individuals. We attempt to reconstruct the Facebook friendship graph, Facebook interaction network, as well as call and SMS networks from longitudinal records of person-to-person offline proximity. We find subtle, yet observable behavioral differences between pairs of people who communicate using each of the different channels and we show that the signal of friendship is strong enough to stand out from the noise of random and schedule-driven offline interactions between familiar strangers. Our study also provides an overview of methods for link inference based on offline behavior and proposes new features to improve the performance of the prediction task.

Peer feedback is the act of letting students give feedback to each other on submitted work. There are multiple reasons to use peer feedback, including students getting more feedback, time saving for teachers and increased learning by letting students reflect on work by others. In order for peer feedback to be effective students should give and receive useful feedback. A key challenge in peer feedback is allocating the feedback givers in a good way. It is important that reviewers are allocated to submissions such that the feedback distribution is fair - meaning that all students receive good feedback. In this paper we present a novel way to intelligently allocate reviewers for peer feedback. We train a statistical model to infer the quality of feedback based on a dataset of feedback quality evaluations. This dataset contains more than 20,000 reviews where the receiver of the feedback has indicated the quality of the feedback. Using this model together with historical data we calculate the feedback-giving skill of each student and uses that as input to an allocation algorithm that assigns submissions to reviewers, in order to optimize the feedback quality for all students. We test the performance of our allocation strategy using real data from over 600 peer feedback sessions and simulate the effects of different allocation strategies. By comparing our method with a random allocation algorithm and a "super-informed oracle" algorithm we demonstrate that we are able to allocate reviewers to submissions in such a way that all submissions receive feedback of similar quality and that we are able to significantly outperform simple random allocation of reviewers. Additionally we investigate the effect of pre-allocating reviews in comparison to allocating reviewers live during the review process and show that live-allocation leads to better results. Our method is robust to reviews not being completed and other real-life quirks and improves as more feedback data is collected.

The act of producing content - for example in forms of written reports - is one of the most used methods for teaching and learning all the way from primary school to university. It is a learning tool which helps students relate their theories to practice. Getting relevant and helpful feedback on this work is important to ensure a good learning experience for the students. Providing this feedback is often a time-consuming job for the teacher. An effective way to learn is to teach others, and similarly give feedback on work done by others. One way to approach a combined solution to the above challenges, is to use peer assessment in the classroom which as a learning method has become more and more popular. In this paper we look at data collected using the web-based peer assessment system Peergrade. The dataset consists of over 350 courses at more than 20 educational institutions and with a total of more than 10,000 students. The students have together made more than 100,000 peer-evaluations of work by other students, and these evaluations together contain more than 10,000,000 words of text feedback. A key problem when using peer assessment is to ensure high quality feedback between peers. Feedback here can be a combination of quantitative / summative feedback (numerical) and qualitative / formative feedback (text). A lot of work has been done on validating and ensuring quality of quantitative feedback. We propose a way to let students evaluate the quality of the feedback they receive to obtain a quality measure for the feedback. We investigate this measure of feedback quality, which biases are present and what trends can be observed across the dataset. Using our measure of feedback quality, we investigate how it relates to various factors like the length of the feedback text, the number of spelling mistakes, how positive it is and measures of the student's report-writing skills.

Today's societies are enveloped in an ever-growing telecommunication infrastructure. This infrastructure offers important opportunities for sensing and recording a multitude of human behaviors. Human mobility patterns are a prominent example of such a behavior which has been studied based on cell phone towers, Bluetooth beacons, and WiFi networks as proxies for location. While mobility is an important aspect of human behavior, it is also crucial to study physical interactions among individuals. Sensing proximity that enables social interactions on a large scale is a technical challenge and many commonly used approaches—including RFID badges or Bluetooth scanning—offer only limited scalability. Here we show that it is possible, in a scalable and robust way, to accurately infer person-to-person physical proximity from the lists of WiFi access points measured by smartphones carried by the two individuals. Based on a longitudinal dataset of approximately 800 participants with ground-truth Bluetooth proximity collected over a year, we show that our model performs better than the current state-of-the-art. Our results demonstrate the value of WiFi signals as a tool for social sensing and show how collections of WiFi data pose a potential threat to privacy.

Human mobility patterns are inherently complex. In terms of understanding these patterns, the process of converting raw data into series of stop-locations and transitions is an important first step which greatly reduces the volume of data, thus simplifying the subsequent analyses. Previous research into the mobility of individuals has focused on inferring ‘stop locations’ (places of stationarity) from GPS or CDR data, or on detection of state (static/active). In this paper we bridge the gap between the two approaches: we introduce methods for detecting both mobility state and stop-locations. In addition, our methods are based exclusively on WiFi data. We study two months of WiFi data collected every two minutes by a smartphone, and infer stop-locations in the form of labelled time-intervals. For this purpose, we investigate two algorithms, both of which scale to large datasets: a greedy approach to select the most important routers and one which uses a density-based clustering algorithm to detect router fingerprints. We validate our results using participants’ GPS data as well as ground truth data collected during a two month period.

Let $d\geq 3$ be a fixed integer. We give an asympotic formula for the expected number of spanning trees in a uniformly random $d$-regular graph with $n$ vertices. (The asymptotics are as $n\to\infty$, restricted to even $n$ if $d$ is odd.) We also obtain the asymptotic distribution of the number of spanning trees in a uniformly random cubic graph, and conjecture that the corresponding result holds for arbitrary (fixed) $d$. Numerical evidence is presented which supports our conjecture.

The use of data analysis competitions for selecting the most appropriate model for a problem is a recent innovation in the field of predictive machine learning. Two of the most well-known examples of this trend was the Netflix Competition and recently the competitions hosted on the online platform Kaggle. In this paper, we will state and try to verify a set of qualitative hypotheses about predictive modelling, both in general and in the scope of data analysis competitions. To verify our hypotheses we will look at previous competitions and their outcomes, use qualitative interviews with top performers from Kaggle and use previous personal experiences from competing in Kaggle competitions. The stated hypotheses about feature engineering, ensembling, overfitting, model complexity and evaluation metrics give indications and guidelines on how to select a proper model for performing well in a competition on Kaggle.

Many complex networks feature relations with weight information. Some models utilize this information while other ignore the weight information when inferring the structure. In this paper we investigate if edge-weights when modeling real networks, carry important information about the network structure. We compare five prominent models by their ability to predict links both in the presence and absence of weight information. In addition we quantify the models ability to account for the edge-weight information. We find that the complex models generally outperform simpler models when the task is to infer presence of edges, but that simpler models are better at inferring the actual weights.

We consider string matching with variable length gaps. Given a string T and a pattern P consisting of strings separated by variable length gaps (arbitrary strings of length in a specified range), the problem is to find all ending positions of substrings in T that match P. This problem is a basic primitive in computational biology applications. Let m and n be the lengths of P and T, respectively, and let k be the number of strings in P. We present a new algorithm achieving time O((n + m)logk + α) and space O(m + A), where A is the sum of the lower bounds of the lengths of the gaps in P and α is the total number of occurrences of the strings in P within T. Compared to the previous results this bound essentially achieves the best known time and space complexities simultaneously. Consequently, our algorithm obtains the best known bounds for almost all combinations of m, n, k, A, and α. Our algorithm is surprisingly simple and straightforward to implement.