Abstract

Satisfaction with life (SWL) is an important concept for measuring the well-being of individuals and nations, but it is not trivial to measure. While measurement scales, such at Diener SWL scale, have been used widely, they require administration of survey, which always carries challenges, including representativeness of the sample and sample size. Facebook, on the other hand, is a widely used social network and provides a wealth of data that have been successfully used for inferring psychometric results.

This thesis presents a custom made machine learning pipeline which uses Facebook statuses and Facebook page likes to predict satisfaction with life of the user. The pipeline uses a wealth of features extracted from Facebook data, using multiple feature engineering techniques, such as Like topic modelling with Latent Dirichlet Allocation, text modelling with TF-IDF, word topic modelling and text mining, all of which are described. The pipeline makes a choice of hyperparameters for Random Forest or XGBoost models and performs either classification into three SWL categories, or regression on SWL score.

The results beat theoretical and previous state of the art baselines for Pearson’s r correlation between predicted and actual scores for SWL, despite being obtained on smaller datasets with fewer available Facebook features. The best performing Facebook data model incorporates Facebook statuses only and achieves Pearson’s r = 0.30, which is a 77% improvement over the previous work. Results are discussed and a number of directions for further research, data engineering and methodology improvement are suggested.

Description

This project was done for Optimization for data science course, professor Francesco Rinaldi (Department of Mathematics, Padua).

An adversarial image is an image that tricks an image classifier, but not a human. If you, the human, see an adversarial image of a kitten, it will be obvious to you that the picture contains a kitten, but state-of-the-art image classifier with a high accuracy rate, that this image was designed to fool, will misclassify it.

This project implements a few variants of the Frank-Wolfe algorithm which can be used for adversarial attacks on images. Four variants are presented – for white box and black box attacks, with and without momentum. The model attacked is Inception V3 model, and the dataset used is Imagenet validation dataset.

Doctors face daily decisions about the best care for their patients, and their own clinical experience can be enhanced using evidence-based medicine, such as through clinical trial data. One of the most robust ways of synthesizing research evidence across healthcare trials is through a systematic review. Cochrane is a not-for-profit organization that creates, publishes and maintains systematic reviews of health care interventions.

Cochrane systematic reviews provide high-quality, relevant synthesized research evidence to decision-makers at the international, national and local level, to make sure citizens receive the medical and social care based on the best available evidence. While this is a rigorous approach, it can take up to three years to produce a major systematic review, which limits our ability to use up to date research to guide decision-making. The goal of this project was to build a system that identifies relevant studies prospectively.

This project was done during Data Science for Social Good fellowship, in partnership with Cochrane. I have managed the project, team members were Kim De Bie, Anthony Rentsch and Nishant Kishore, technical mentor was Pablo Rosado.

This project was presented at Harvard CRCS workshop on AI for Good.

Paper

West Midlands Combined Authority is a cluster of 18 local authorities and four Local Enterprise Partnerships (LEPs), working collaboratively on many projects ranging from transport, jobs growth, industrial strategy, public service reform etc. to deliver their vision of a more prosperous West Midlands. As the region grows, both economically and in terms of population, it is becoming increasingly important to ensure high quality and equal access to transportation.

The goal of the project is to make transportation access in the West Midlands region fairer. The partner, WMCA, has indicated there is a problem of areas being isolated from public transport lines, with poor access to private vehicles and have a congested highway network – this is particularly prevalent among areas with high minority and low-income populations. Although the partner has some instinct on which areas are experiencing isolation, the volume, scope, and extent of the problem has not been analysed and quantified.

The project involves finding a better way to measure equity of access to transportation services, and creating a data product which allows policy makers to analyse if and where the disadvantaged people live, if the system as it currently is does not serve any particular demographic group, and offer insights on how to make transport better and fairer.

This project was done during Data Science for Social Good fellowship, in partnership with Transport for West Midlands. I have managed the project, team members were Ece Calicus, Tammy Tseng, Renzhe Yu and James Trimarco, technical mentor was Adolfo de Unanue.

This project was done for Business, economic and financial data course, professor Omar Paccagnella (Department of Statistics, Padua). My team members were Martina Contisciani and Nicolò Ruggeri.

Different individuals, or groups of individuals, could understand the same survey question differently, depending on cultural or individual differences. This could lead to bias in measurement of variables of interest, an error which could propagate onwards, resulting in wrong analysis.

Market research heavily relies on results of customer satisfaction analysis to inform decision making and further product development. However, this measurement is prone to analytical errors, amongst other reasons due to it being subjective. One of the important dimensions often used is collection of subjective evaluation of product performance after the purchase and use, compared with the expectations the customer had of the product a priori. These evaluations are usually done with ordinal scales, and so, due to the subjective nature of the measuring concept, the results are prone to being sensitive to individual or cultural differences.

This project utilises anchoring vignettes to correct for personal and cultural differences in response style of participants. We use parametric models (Ordered Probit and Chopit) to capture demographic differences in response style for self reported satisfaction with a technological device.

This project was done for Human data analytics course, professor Michele Rossi (Department of information engineering, Padua). My team mate was Nicolò Ruggeri.

Abstract

Keyword spotting has been an increasingly studied problem in the recent years due to the growing number of applications. The use of keyword spotting in portable devices developed by major technology companies incentivised the contribution of academic and technical community. In addition, a benchmark dataset was recently open sourced by Google, making results comparison standardised through different papers and architectures.

A major difficulty in developing models appropriate for this task is the need to keep the computational and memory cost appropriately low, due to the requirement to process data in place, within a low memory setting. Even though other approaches are possible, like in-device sound detection with cloud processing after activation, the need for compact, low footprint architecture is compelling in many use cases.

In line with some of the recent advances in the field, this paper aims to study one of the architectures proposed recently, trying to improve it from a memory and computational footprint and classification performance points of view. We successfully improved in both directions.

Beyond its state-of-the-art performances, the selection of the model has been based on the versatility of it, as well as for the two main ideas that it exploits: recurrent neural networks and attention mechanism. For this reason, by methodically studying effects of architectural modifications on this model, we attempt to draw some functional conclusions about the role of layers in the network itself. This is in line with attempt of the technical community to tackle the problem of the black-box nature of some machine learning models, present within majority of deep learning approaches, which is considered one of the major drawbacks in regards to the usage of neural networks, and go beyond just prediction and towards understanding of the models.

This project was done for Biological data course, professor Silvio Tossato (Department of Biomedical Sciences, Padua). My team members were Carla Glassl, Martina Contisciani and Nicolò Ruggeri.

The project performs characterisation of a group of kinase proteins in humans by developing Hidden Markov Models of the group, performing search and alignment of protein sequences to match the model, clustering results of the search into functional families, and functional analysis of the families through utilisation of wide range of bioinformatics and data science tools and techniques.

This project was done for Cognitive, behavioural and social data course, professor Giuseppe Sartori (Department of Psychology, Padua). My team members were Matteo Migliorini, Martina Contisciani and Nicolò Ruggeri.

Development of the system which analyses video of a person talking, detect facial features, including eyes and detects and measures blink rate of the person, to consequently be used for lie detection.
The project included development of the software which uses deterministic approach (Python, OpenCV) as well as exploration of machine learning techniques for the same purpose (SVM, Decision Tree, Random Forrest).

The system was developed in order to be used by the researchers at Human Inspired Technology Research Centre.