Kirill Yakunin

NLPMonitor is a mass-media monitoring information system, which scrapes news publications from mass media and social networks in order to perform informational trend analysis as well as document classification using multiple evaluation criteria.

I’ve built the system from scratch, including:

1) Containerization (Docker), server administration
2) Storage – a hybrid of ElasticSearch and PostgreSQL
3) Scraping using Scrapy
4) Apache Airflow
5) Django web application
6) Text preprocessing, topic modeling, and classification pipeline

I proposed a method for creating topic-based text representation, which allows for building computationally efficient text classification models. A method for high-level labeling was proposed, which allows for classifiers with minimal manual labeling. A number of experiments were conducted including:

1) Sentiment analysis based on expert labeling of 100-200 topics
2) Propaganda identification based on expert labeling of several dozen media-sources
3) News popularity identification based on automatic labeling based on objective user engagement data

The project was integrated into the Ministry of Science and Education of Kazakhstan, the results were published in a Q1 Scopus-indexed article.

Magnum is the largest FMCG retailer of Kazakhstan, which recently started developing digital products, such as eCommerce (delivery) and loyalty programs. The recommendation system is a vital part of such infrastructure. In this project, I built a recommendation system from scratch, including ETL and model training pipelines, an analytical web tool, and an API for integration. The recommendation algorithm is based on novel Item2Vec approach, which employs ideas of distributed semantics in order to attempt to find contextually similar objects.

However, this approach on its own is not flexible enough and doesn’t intrinsically create value. For example, it can recommend just some commonly bought items like bread, milk and bananas. In order to solve this problem, a number of additional algorithms and business rules were applied in order to create three modes:

1) Complimentary items recommender
2) Upsell recommender
3) Category recommender

The system was integrated with eCommerce and loyalty applications and showed a 50 to 70% increase in user engagement compared to a popular but simple item recommender.

This was a scientific project that required building a geoinformation system for decision-making support for the installation of renewable energy sources (RES) generators. A number of factors need to be taken into account:

1) Technological limitations, such as slope, available area, etc.
2) Economic feasibility – the cost of infrastructure, demand for energy in nearby areas, etc.
3) Energy potential – some integrated indicator of solar/wind/hydro/biomass energy potential

It’s also possible to install different types of RES generators with different characteristics, capacity and prices.

I helped develop the Bayesian fuzzy AHP (BaFAHP) MCDM method, which allows for the calculation of fitness for a given piece of land based on spatially-distributed data (both vector and raster) and expert assessments of factor importance. I’ve also performed a number of computational experiments, performed preparation of data from open sources (NASA SSE, OpenStreetMap, etc.), and integrated the developed algorithm in a Java Spring web application.

I also performed research with colleagues from Moscow State University, including comprehensive model result validation.

The results were published in Q1 Scopus-indexed IEEE Access journal.