Classical Quantile Regression and One-Dimensional Vector Quantile Regression

Advised by Prof. Alfred Galichon
Presentation Report Code

In OLS, we are often constrained to the following assumptions: linearity; strict exogeneity; no multicollinearity; spherical errors; normality. In the context of quantile regression, the usage of conditional quantile allows one to deal with heteroskedastic data. We may also be able to obtain a more complete picture of the distribution by computing multiple regression curves of different percentiles. Under Prof. Alfred Galichon’s guidance, I explored both classical quantile regression and the one-dimensional version of vector quantile regression, originally proposed by Professor Galichon and his colleagues. My work confirmed their equivalence, not just in computational outcomes but also in their linear programming formulation. A significant focus was on the computational aspect, especially the manipulation of matrices to solve quantiles as an optimization problem.

Analyzing Gender Distribution and Trends of Sexism in French Advertising: A Natural Language Processing Approach

Supervised by Prof. Eric Brousseau and Bruno Chaves Ferreira
Report

When I interned in at ACSS Institute - PSL, we worked on a research project facing the Audiovisual and Digital Communication Regulatory Authority (Arcom). I designed algorithms of gender prediction based on scripts and applied Natural Language Processing models such as KeyBert and zero-shot classification for keyword extraction and specific theme detection. The algorithm is capable of calculating the percentage of men and women speakers in each Ad sector to examine the gender distribution in French advertisements across various industries, with the results plotted clearly with Pyplot. An interpretation focused on the potential trends of sexism in French Ads.