Projects

To achieve full understanding of the use and application of ML algorithms, our participants will work on a real-life industry project, translating theoretical knowledge to practical process and overcoming realistic challenges.

Duration:

~6 months

Scope:

~400 work hours total

Team:

2-3 students

Data:

Real data provided by company

Guidance:

Experienced mentors provided by Y-DATA

Support:

Weekly meetings with company data-owner

Projects 2018-2019

Playbuzz

Content to context matching

Playbuzz is a storytelling platform offering real-time analytics tools to enable partners to engage users, boost reach, raise brand awareness, improve monetization capabilities, and optimize content for maximum social interaction. Provided with a dataset of the publishers’ pages and the relevant Playbuzz tagged units as well as previous users’ behavior on those pages, the company is looking to increase the users engagement with Playbuzz content units. Previous research has shown that users tend to engage better with a content unit that relates to the page content and context (subject, entities etc.). The aim of this project is to analyze the publishers web pages’ context (especially news and content sites) using NLP and try to find the most suitable matching between relevant content pieces via ML, which will be tested on site with A/B testing to verify its effectiveness.

Full project cycle

The process of working on the project follows popular industry standards and methodologies and incorporates a growing set of tools the students possess to methodically understand and solve a real-world problem. Our students have a full-cycle data science project in their portfolio upon graduation, covering all industry-standard stages: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation.

A customer operates a forum where programmers ask each other questions, provide answers and rate questions giving them \"ups\" and \"downs\". The forum has a core expert community that provides good answers and valuable insights. However, they often waste their time handling questions of little to no value: marking questions as duplicates and redirecting them, closing topics with incoherent or irrelevant questions etc. Because of this, the overall efficiency of the system suffers.

The customer desires to improve the system efficiency as measured by the mean time between a question being posted and the first accepted (upvoted) answer given. How do we translate this request into ML terms? Should this be treated as a classification (good question or not) or regression (predicted number of up/down votes) problem? Which metric should be used? Accuracy, precision, recall, AUC etc. – which is most relevant to the situation? Is the problem symmetrical? We’re more concerned with losing a valuable question then missing several low-value ones. What are our resource constraints? We probably need a solution that runs immediately once a question is posted, so can’t use resource-heavy algorithms.

Downloading the questions and additional information, followed by a decision whether to use date-time stamps, user ID and other technical information. Cleaning the data by removing intrusive elements such as empty or corrupted questions, questions in Chinese asked in an English forum or embedded as text in a picture. Are more low-value questions asked by first-time users? Maybe we want to construct a separate model for first-time and veteran user questions. Are most questions plain text or do they use embedded code which should be parsed and analyzed separately? Do we have enough questions with up/down votes to construct a metric, or are there many old questions predating the up/down vote system which can’t be used and should be removed?

Recognizing which features provide statistically significant contribution to the problem at hand, and removing those that don’t: Is our users’ geolocation relevant to question-quality? Is the vast majority from the same region? Extracting useful features for learning and constructing new ones when needed: parsing date-time stamps, text vectorization, one-hot encoding keyword tags, etc. If we’re splitting the data into training and test sets, how is the split made? By time? By user ID?

The heart of the learning process: selecting and building the model. We need to choose an algorithm: be it a simple logistic regression or XGBoost or an ensemble of neural networks, it must be chosen based on the resources available and the peculiarities of the problem. Then we train it, making sure we avoid overfitting, and tune hyperparameters to maximize its performance.

Finally, our model is ready, it marks all low-value questions with astounding precision and leaves all relevant questions intact. But will it help to boost consumer's metric? To check it we should probably design and conduct an A/B-test and to determine statistical significance of the results. If A/B-testing is impossible or undesired, we may probably use some sort of causal impact inference.

Projects

Projects 2018-2019

Playbuzz

Full project cycle

Example Project