Abdul Rafey Tahir

Foot Locker USA has an extensive loyalty program that rewards customers based on their spending. They have three customer classes based on heuristics:
• X1: VIP customers
• X2: average customers (with reasonable spending)
• X3: low-spending customers

The project involved building a machine learning model to predict which customers would and wouldn't change their category in the next quarter based on data from the past eight quarters.

Using pandas, I cleaned and prepared the dataset from 2019 Q1 to 2020 Q4 for feature engineering. The target variable class_change came from 2021 Q1 data. It was set to 1 for customers whose classes had changed during that time (either upgraded or downgraded) and 0 for those who hadn't. The test set target variable came from 2021 Q2. After generating quarterly features from data—including the number of website visits, orders placed, items checked out, items viewed, the amount spent, and other session data and shopping history—a random forest classifier was trained using scikit-learn. The model did fairly well on the test set, with a recall of 0.62 and a precision of 0.87. It was then deployed on Azure.

The idea was to create an application for YouTube content creators to filter out requests for new content from their viewers in the comment section. This would help them create content that the audience demanded to watch.

There project involved:

• Data collection. I developed a scraper using Python and Selenium, crawled many channels, and collected the top 100 comments on every video. The clients had a team of annotators who annotated about 100,000 comments, of which there were roughly 10,000 requested comments.

• Model training. The dataset was quite imbalanced, so I downsampled negative class data points to 30,000. I then applied basic text preprocessing techniques using Pandas and NLTK, such as special character removal, lowercase alphabets, text tokenization, and stemming. Also, I generated simple frequency-based features using TF-IDF vectorization in scikit-learn and trained a support vector machine (SVM) model. It achieved 83% recall and 91% precision on the validation set.

• Inference and optimization. I used my scraper to scrape a few more websites and ran an inference for the scraped comments. Since the goal was high precision, we got the events classified as requests, annotated them again, and retrained the model.

My team and I developed an LLM-based meeting minutes generation model, which was trained on 1,000 meeting videos scraped from YouTube. The first part involved audio/video-to-text transcription, for which we used Meta's Whisper model. Once the transcript was generated, we used Llama 2 to create a summary of the transcript and then used the summary to generate meeting minutes in the form of bullet points and a flow chart.

We used GCP for training and a Tesla A100 for running training jobs. For inference, we used P100 GPU, which worked pretty well for the scope of our application. We also built a Flask app that is containerized and deployed on GCP. The models are deployed as API endpoints and consumed by the app.

Used "eCommerce Behaviour Data from Multi-category Store" with 285 million users. Performed RFMV analysis on the data to segregate more frequent and high-spending customers from others. I then used the LightFM library to train a deep learning recommendation model based on purchased data. Trained an SRGNN deep learning model from the RecBole library on session data. Combined the predictive capability of the two models for a better recommendation. In addition to the recommendation system, the focus was on churn reporting, especially for more frequent and high-spending customers. The goal was to retain these customers by suggesting personalized marketing strategies.