Tatev Karen Aslanyan Portfolio
Research-grade GitHub portfolio with published papers, recommender systems, and deep learning projects — shows how to blend academic credentials with industry relevance.
Location
Didn't say. We're respecting the mystery.
Skills
What makes it work
A breakdown of the choices that make this portfolio stand out.
GitHub README as a structured portfolio page
A well-crafted README with sections, badges, project summaries, and navigation turns a GitHub profile into a landing page. This format is native to the hiring workflow for data science roles — recruiters and technical leads go to GitHub first.
Published papers as third-party validation
Peer-reviewed papers confirm that domain experts evaluated and approved the methodology. For data science roles, especially at research-adjacent companies, publications carry weight that no self-reported project description can match.
Specialization in recommender systems
Recommender systems power every major platform (Netflix, Spotify, Amazon). Claiming deep expertise in a commercially valuable specialization is more hirable than broad ML generalism at every career stage beyond entry level.
What Data Scientists can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Invest seriously in your GitHub profile README. It is your data science portfolio — format it like a webpage, not a code comment.
Each project repository should have its own detailed README: business problem, dataset, methodology, evaluation metric, and result. Not just a notebook.
Kaggle competition placements and published papers are objective metrics. Display them prominently — they distinguish you in a field full of people who "know Python and scikit-learn."
Specialize. "Recommendation systems at scale" is more valuable than "I know supervised and unsupervised learning."
Document not just what model you chose, but why you chose it over alternatives. Hiring managers look for decision-making evidence, not just execution.
Case studies
Why, then How. Tatev Karen Aslanyan puts every case study through that two-line template, and most come with a written case-study paper — from a recommender algorithm presented at ACM SAC 2021 down to a cats-versus-dogs CNN, so even the small ones read as deliberate.
Case Study: Price Prediction with Recurrent Neural Networks
Why: exact price prediction is nearly impossible, but approximate trend estimation can inform an investor's buying strategy. How: an RNN with 5 stacked LSTM layers, combined with time-series analysis on 5 years of historical stock price data, predicts upward and downward trends in future prices. The code trains on a Google stock price train/test split.
The Why/How framing plus a written case-study paper shows how she communicates modelling decisions, not just code.
5 stacked LSTM layers of 50 units each, trained for 120 epochs
Case-study paper accompanies the code
Frames the goal honestly: trend direction, not exact price
Case Study: Customer Churn Rate Analysis with Artificial Neural Networks
Why: estimate which bank customers are likely to leave, so that marketing can try to retain them. How: bank customer behaviour data trains an artificial neural network (two 6-unit ReLU hidden layers, a sigmoid output, the Adam optimizer, 100 epochs) to predict each customer's probability of leaving, evaluated with a confusion matrix and accuracy.
A compact, fully documented classification case study tied to a clear business action — a good template for a first portfolio piece.
Business motivation stated before the model
Categorical encoding, scaling and train/test split all shown in the code
Case-study paper written up alongside the code
Case Study in Multivariate Statistics: Ranking US Cities
Why: identify the US cities with the highest capital, urban growth and development, for anyone looking to relocate. How: Places Rated Almanac data is used to rank cities on a single combination of 9 rating variables using Principal Components Analysis and Factor Analysis, with Canonical Correlation Analysis to examine correlation between two sets of rating variables.
Classical multivariate statistics in R rounds out a deep-learning-heavy portfolio and shows the statistical grounding behind the models.
PCA, FA and CCA applied to one real ranking problem
Written in R, showing range beyond Python
Same Why/How template as the ML case studies
Projects
LDA-LFM: A New Hybrid Recommender Algorithm
Why: most recommenders use only ratings and ignore textual reviews or user and item characteristics, and many cannot handle large datasets. How: a hybrid algorithm that combines a Latent Factor Model for ratings with LDA topic modelling on review text, extended to accept extra user- and item-specific features. It was evaluated on Amazon.com datasets against baselines, including standard matrix factorization, an offset model and an LDA-first model.
A peer-reviewed algorithm with public Python code and baselines is external validation that a self-described project cannot match — and recommender systems are commercially central.
Two ACM publications: the SAC 2021 paper and an extended ACR version
Code includes the baselines (standard MF, offset, LDA-first) as well as the proposed model
Fuses text (topic modelling) with ratings instead of using ratings alone
More Data Scientist portfolios
Ready to build your portfolio?
Follow the Data Scientist roadmap — skills, projects, and timeline to get hired.