For the concluding speaker in our September-talks series, I am happy to announce that Hartmut Lentz will talk about his work on spread of infectious disease on temporal networks. Hartmut works at the “Institut für Epidemiologie” at the Friedrich Loeffler Institute. He is a fantastic speaker and an authority on temporal networks. Full details below.
- Date: Thursday, September 28th
- Time: 1pm
- Location: DTU, Building 321, 1st floor lab-space
- Title: Spread of infectious diseases in temporal networks
Abstract: Many networks are treated as static objects, although they are in fact strongly time-dependent. This can have a dramatic impact on the possible spreading patterns for infectious diseases.
A static (aggregated) trade network is constructed as follows: if two nodes are connected directly to each other in a time-dependent network, the same connection is present in the static network. A fundamental difference between the static and the time dependent view however, is the consideration of paths, i.e. indirect connections over more than one edge. Concerning paths, the causality of the edges used plays an essential role. In an aggregated network, paths can seem causal, although they do not follow a time-respecting sequence of edges in the real system. This leads to a systematic overestimation of outbreak sizes, if time-dependent networks are treated as static.
We introduce a new method, which allows for the computation of the total causal path structure of a temporal network (represented by its accessibility graph) using the adjacency matrices of its snapshots. In addition, information about the timescales required for path traversal can be derived from the step-by-step derivation of the accessibility graph of the network. This procedure directly yields the distribution of shortest path durations in a temporal network. In addition, we define the new measure causal fidelity that compares the number of paths in a temporal network with its aggregated counterpart. This measure allows a quantitative assessment of how well a temporal network can be approximated by a static aggregated one.
The methods presented here require only basic knowledge linear algebra and can be implemented efficiently. Their capability is demonstrated for three examples: networks of social contacts, livestock trade, and sexual contacts.
The September talk series continues full steam ahead. This week, you have a chance to see Philipp Lorenz talk about dynamics of topics in online social media. Philipp is a PhD student at TU Berlin’s Institut für Theoretische Physik in the Nonlinear Dynamics and Control: empirical networks and neurodynamics group. Phillip’s work focuses on temporal communities of hashtags, modeling the rise and fall of online topics, threshold models with repost and recovery, and more. Details of the talk below
- Date: Tuesday, September 19th
- Time: 2pm
- Location: DTU, Building 321, 1st floor lab-space
- Title: Capturing and modeling the dynamics of online topics
Abstract: Online media have a huge impact on public opinion, economics and politics. Every day, billions of posts are created and comments are written, covering a broad range of topics. Especially the format of hashtags, as a discrete and condensed version of online content, is in our focus. Here we present a pipeline, consisting of methods from static community detection as well as novel approaches for tracing the dynamics of topics in temporal data. We build co-occurrence networks from hashtags with timestamped edges. On static snapshots we infer the community structure and solve the resulting bipartite matching problem, by taking into account higher order memory. The results are robust to temporal fluctuations and instabilities of the static community detection.
The resulting dynamics in various datasets and for different observables, such as the community sizes or the likes they gather, as a proxy for the popularity of a topic, we observe universal behavior. Despite their versatility we find that in all datasets the distributions of gains and losses in popularity are fat-tailed, indicating occasional but large and sudden changes in public interest.
We hypothesise that only a few mechanisms may govern this behavior:
- Gaining interest follows the rule of preferential attachment .
- Saturation of the limited attention span decreases its fame.
- discrete ranking leads to a competition between threads.
With these ingredients, we are able to design a class of models, which can reproduce the qualitative dynamics and the quantitative distributions of dynamical properties in the empirical observations. The model parameters and the required configuration for a given dataset is informational with respect to the sociological and psychological mechanisms that drive the dynamics of popularity in different contexts.
We continue the streak of exciting September talks. This time it’s DTU alum (now MIT) Bjarke Felbo who recently caused international press frenzy (see MIT Technology Review, BBC, Newsweek, Business Insider, The Telegraph, The Register, Huffington Post (FR), Numerama (FR) for details) with his sarcasm-savvy deep learning algorithm. Now there’s a great opportunity you can get all the technical details and ask questions, etc.
- Place: Technical University of Denmark, Building 210, room 112.
- Date: September 13th, 2017
- Time: 13:00
Title: Using millions of emoji occurrences to learn any-domain representations for detecting sentiment, emotion and sarcasm
Abstract: NLP tasks are often limited by scarcity of manually annotated data. In social media sentiment analysis and related tasks, researchers have therefore used binarized emoticons and specific hashtags as forms of distant supervision. Our paper shows that by extending the distant supervision to a more diverse set of noisy labels, the models can learn richer representations. Through emoji prediction on a dataset of 1246 million tweets containing one of 64 common emojis we obtain state-of-the-art performance on 8 benchmark datasets within sentiment, emotion and sarcasm detection using a single pretrained model. Our analyses confirm that the diversity of our emotional labels yield a performance improvement over previous distant supervision approaches.