Students projects and Master thesis available!

Title: Machine Learning for Predicting Users’ Behaviours in Virtual Reality

Academic supervisor: Dr. Laura Toni

Requirements: Good programming skills, optional prior knowledge of Deep Learning.

Background
Omnidirectional visual content can be visualized via wide field-of-view or immersive Virtual Reality (VR) displays, such as head-mounted displays (HMD), or on standard 2D displays using video players that allows interactive video navigation, such as YouTube 360 [1] or GoogleMap Viewer. In both scenarios, the viewer is assumed to be at the center of a viewing sphere, displaying the view corresponding to the direction he is facing. During the displaying, the viewer can change his viewing direction by rotating his head (for head mounted devices) or by browsing the video with a cursor (for YouTube 360). Understanding and predicting the viewers’ head movement (displayed viewports) has become essential to efficiently compress and stream omnidirectional videos.

The goal of this project is to develop a software able to capture the viewport displayed by users during streaming sessions using Google cardboards. Once the software is in place, the student will run subjective tests to gather an exhaustive data set of navigation patterns in 360 videos. 

References:
[1] https://salient360.ls2n.fr/grand-challenges/icme18/ [papers listed as winners]


Title: Graph-based Learning for Recommendation Systems      

Academic supervisor:Dr. Laura Toni

Requirements: Very good knowledge of linear algebra, and good coding skills (python). Machine learning background is a plus.

Nowadays video streaming platforms, as YouTube, usually recommends personalized sets of videos to users based on their past selected videos. More at large, every recommendation and ads system nowadays is based on machine learning algorithms that learn and predict the best video or webpage to suggest to a given final user.

Within this framework, contextual learning has been proposed as a viable learning method to keep up with the high dimensionality of the search domain (e.g., possible videos to recommend). Even if contextual learning reaches optimality (minimizing the regret bound) asymptotically, the learning process scales with the context space dimension, leading to unacceptable performance in scenarios of large context domains (as the case of recommendations). To improve the learning process, the context structure can be exploited. This project is aimed at study structural contextual learning in the case of graph-based contexts.

References:

[1] M. Valko, R. Munos, B. Kveton, & T. Kocak, “Spectral Bandits for Smooth Graph Functions”, ICML 2014.

[2] A. Tossou, C. Dimitrakakis, and D. P. Dubhashi, “Thompson Sampling for Stochastic Bandits with Graph Feedback”, AAAI. 2017.


Title:   Influence Maximization in Social Networks          

Academic supervisor:Dr. Laura Toni

Requirements: very good and solid background on linear algebra. Basic knowledge of Matlab

How does an opinion spread within social networks? For an idea/ad to become viral (viral marketing), where should we place it within the network? A social network plays a fundamental role as a medium for the spread of information, ideas, and influence among its members. An idea or innovation will appear and it can either die out quickly or make significant inroads into the population. If we want to understand the extent to which such ideas are adopted, it can be important to understand how the dynamics of adoption are likely to unfold within the underlying social network. 

This project is aimed at studying these dynamics via influence maximisation (IM) problems on large scale networks.  The key problem is the following: where do we place original ads (seed nodes) in such a way that their spreading across the network is maximised. The novelty will be to learn system dynamics and optimise seeds nodes using graph signal processing toolboxes.

References:

[1] Kempe, David, Jon Kleinberg, and Éva Tardos. “Maximizing the spread of influence through a social network.” Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2003.

[2] Li, Yuchen, et al. “Influence maximization on social graphs: A survey.” IEEE Transactions on Knowledge and Data Engineering (2018).


Title:   Representation Learning on Graphs          

Academic supervisor:Dr. Laura Toni

Requirements: very good and solid background on linear algebra. Basic knowledge of Matlab (ideally Python)

Many domains such as social networks, recommender systems and biological protein- protein networks can be modelled as graphs. Graphs do not only allow to efficiently store and access relational knowledge about interacting entities, but they are also an omnipresent data structure within machine learning. Machine learning applications aim to make predictions, or discover new patterns using graphs as feature information. In this framework, a major challenge is to find a way to represent, or encode, the graph structure such that machine learning models can easily exploit these embeddings.

The recent years have seen increasing effort in building systems that automatically learn to encode graph structure into low-dimensional embeddings. This project aims at developing more data-efficient node embedding models, such that high-quality node representations can be computed from limited amount of observations on large graphs.

References:
[1] A. Grover and J. Leskovec, “node2vec,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining – KDD ’16, pp. 855– 864, 2016.


Title:   Inferring Graph Signal for Network Science        

Academic supervisor:Dr. Laura Toni

Requirements: good understanding of linear algebra, coding skills (ideally python)

Graphs are flexible structures that represent connections between data, and model many everyday applications. Numerical examples can be found in geographical, transportation, biomedical and social networks, such as temperatures within a geographical area, traffic capacities at hubs in a transportation network, or human behaviours in a social network. The information on these networks is in constant evolution (think for example at the traffic flow) and the model driving these temporal evolution is not always known. 

This project will be focused on developing novel machine learning techniques on graphs to efficiently learn these models. In collaboration with the intelligent transport department, the learned model will be validated with realistic traffic data.

References:

[1] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega and P. Vandergheynst. The Emerging Field of Signal Processing on Graphs: Extending High-Dimensional Data Analysis to Networks and Other Irregular Domains, in IEEE Signal Processing Magazine, vol. 30, num. 3, p. 83-98, 2013.

[2] Dong, Xiaowen, et al. “Learning Laplacian matrix in smooth graph signal representations.” IEEE Transactions on Signal Processing 64.23 (2016): 6160-6173.


Title:   Graph Based Learning for Drug Design    

Supervisor:    Dr Laura Toni    

Requirements: good understanding of linear algebra, and good coding skills (ideally python)

Drug design is aimed at discovering novel molecules that are active on specific targets. Because chemical space is too large to be screened in its entirety for drugs active against a particular target, automated design and screening of selected compounds is desirable.   In the literature, a generative deep learning model based on recurrent neural networks (RNNs) for de novo drug design has been proposed [1].  The key intuition is to translate the each protein into a string and feed each set of string (collections of proteins) to a RNN to learn the most active drug design.In this project, we advance state-of-the-art techniques by modelling each protein as a graph and learning graph features that underpin active interactions. Given an available data set, the student will be able to investigate novel cutting edge learning methods for automating the drug screening in drug discovery applications.