Hi, I am Yu.

Research Engineer & Data Scientist in Machine Learning, Recommender Systems and Knowledge Engineering.

About me

After having graduated from University of Montpellier (Master degree in computer science, specialty: "Data, Knowledge & Natural Language") in 2017, I worked for Cirad during 12 months as a research engineer. After having applied what I've learnt during the Master courses to develop a semantic tool for the company, I realised that I need to enrich my knowledge and improve my working/thinking skills, even though my developed tool seemed to pleasure firm agents.

To this end, I chose to continue with a thesis project for obtaining a Ph.D, which is, for me, a difficult, challenging but worthing task. Currenty at the last year of my thesis in IMT Mines Alès, I focus on recommendation systems, a typical application of machine learning, which is widely applied in the real-world setting to alleviate the information overload for end-users. The particular aim of my thesis project is to improve recommendation performances from a data-to-knowledge perspective. The "data" perspective consists in improving the recommendation accuracy (i.e. rating prediction and ranking) by leveraging statistic models while the "knowledge" perspective takes a step more further, aiming at improving the quality of recommendations, in terms of more user-centered aspects such as the recommendation diversity and explicability, by leveraging knowledge engineering notions such as the semantic web and knowlege graphs.

Training-wise, I have a comprehensive background in machine learning and in knowledge engineering. My thesis project makes me improve technique skills (e.g. more familiar with the main machine learning frameworks, the containerization notion etc.). In addition, which is important for me, my Ph.D project also allows me to improve common working skills such as presentation, writing, step-back thinking, team working etc.

When not working, I spend my time making delicious food, travaling, thinking, sharing and discussing life experiences with friends (notably my wife 😉 ), self-taught and self-learning of new technologies.

Domains of interests

  • Machine learning (supervised/unsupervised learning)
  • Recommender systems (collaborative filtering, content-based filtering, recommendation diversity and explanation)
  • Knowledge engineering (semantic web, knowledge graph, ontology, reasoning)
  • Machine learning in graphs (node classification, link prediction, graph embedding, etc.)
  • Information retrieval (query-answering)
  • Data visualisation
Permanent Jan. 2022
Data scientist, R&D engineer for Appvizer, Montpellier

Activities: Personnalization of user recommendations, CTR improvement by leveraging M/DL, NLP models and/or knowledge graphs.

3 Years Oct. 2018
Ph.D, researcher at CERIS (ex LGI2P) Lab@IMT Mines Alès

Research aim & project: From data to knowledge: Towards more accurate, diversified and transparent recommendations.

1 year & 2 Months Feb. 2017
R&D Engineer for DGDRS@Cirad, Montpellier

Project: Design & development of a knowledge-based system leveraging heterogenous data within different information systems with cross-domain ontologies (e.g. AGROVOC). Knowledge extraction and visualisation.

3 Months Jun. 2016
Internship for Pradeo Security System, Montpellier

Project: Study & development of web services for the intergration of Salesforce CRM with the firm's back-end.

2 Years 2015
Master in Data, Knowledge & Natural Language Processing at University of Montpellier

Enriched knowledge: data management, knowledge representation & reasoning, machine learning and natural language processing.

Skills

Language

Python, Java, R, Javascript, Latex

Exploitation environment

Linux, MacOS, Windows

Framework & Tools

Pycharm, Rstudio, Jupyter Lab, Google CoLab, Github, Docker, JavaEE

Thin Laptop

Models

Neuron Networks, Nearest Neighbours, Linear Regression, Logistic Regression, Singular Value Decomposition, Graph Embeddings, KMeans, Optimisation, Hyperparameter tuning

Recommendation & Information retrieval

Collaborative filtering, Matrix factorization, Content-based filtering, Search engine, Word2vec, Greedy selection, Recommendation explanation

Framework & Tools

Numpy, Pandas, Scikit-learn, Scipy, Surprise Lib, PyTorch, Bayesian-optimization

Thin Laptop

Chart types

Scatterplot, Barplot, Relational plot, Density plot, Network, WordCloud, Histogram, Map

Framework & Tools

ggplot2, Seaborn, Matplotlib, GraphViz, D3.js, Vis.js, Tableau, Google Chart

Thin Laptop

Concept & Techniques

Ontology, Knowledge graph, Knowledge representation & reasoning, Semantic web, Linked data

Framework & Tools

GraphDB Ontotext, DBpedia Endpoint, SPARQLWrapper

Thin Laptop

Concept & Techniques

Model-View-Controler (MVC), Object-Relational Mapping (ORM), Front-end (Javascript, HTML5, CSS3, JQuery), Back-end (Apache)

Framework & Tools

Django, Vaadin, Bootstrap

Thin Laptop

Presentation

KeyNote, Powerpoint

Databases

Postgres, SQLite, GraphDB

Remote working (Due to COVID)

Zoom, Skype, Microsoft Teams, StarLeaf

Thin Laptop

Recommendation explanation based on DBpedia category tree

The aim of the project is to explain items recommended for the end-users, independently to the used recommender systems. The explanations in our approach are based on the overlap item properties between user liked items and recommended ones. We take advantage of the publicly available DBpedia knowledge graph and proposed an algorithm which dynamically explores the entire knowledge graph (with 1.6M nodes and 4M edges) to retrieve relevant properties for the item explanations task. In addition, we proposed also an explanation model based on user-interactions. It is to evaluate the underlying algorithm-independence of the recommender system.

To evaluate the model, I developed a movie recommendation system based on the MovieTweetings dataset. The application is developed with Django 3 and deployed using Docker compose. Based on the developed application, an online user-study is carried out, involving engineer students as participants.

Diversity study and analyse

This project aims at comparing different families of recommender systems, in terms of diversity since recommender systems have been mainly evaluated based on the accuracy and few study focused on their comparaison, in terms of other aspects, such as the diversity.

I implemented, for this purpose, 7 different recommendation approaches on three benchmark datasets (with hyperparameter tuning for each RS and each dataset), including classical collaborative filtering approaches, matrix factorization, knowledge graph embedding based models and recent proposed deep neuron network-based model.

EBCR

EBCR means for Empirical Bayes Concordance Ratio, an approach for adjusting user-user or item-item similarities during the memory-based collaborative filtering algorithm. The proposed method allows to find a more reliable neighbourhood, for a given user, on which the user rating predictions may based.

The model learns a prior distribution of pair-wise user concordance ratios and use the prior to adjust similarities a posteriori. The results showed a signficant improvement of rating predictions, in terms of RMSE and MAE errors.

Eco-FiL

Eco-FiL means for "Economy & Filières", which represents a knowledge base that constructed by multi-domains experts including economy, agronomy, biology etc. The aim of the project was to facilitate knowledge extraction and visualisation for different types of agents (researcher, HR, manager). To this end, I developed a tool based on an ontology composing of the main concepts and instances of multi domains of interest. Based on semantic web techniques, the tool transfers automatically the firm's heterogenous data (within different informations systems) to structured knowledge. The approach is able to infer new knowledge (facts) from the existing data, thanks to the defined ontology.

The second key feature of the developed tool was the design of the visualisation of the extracted knowledge. As most of the users are not familiar with how AI allows to do this job, it is thus cricial to provide with them an appropriate representation of the obtain results. For this purpose, I chose Networks as visualisation styles, which allows users to quickly understand the extracted information, such as who co-worked with who on which topic or which concept. In addition, the density of the resulting network allowed managers to identify the key concept to consier, the key agent for every given project.




Contact

Feel free to contact me for any need. Send me an email at [email protected] or reach me on social networks