Projects

Overview:

High-entropy alloys (HEAs) are a new class of materials that contain five or more principal elements mixed in near-equal atomic ratios. These alloys exhibit exceptional mechanical properties such as strength, toughness, and thermal stability. However, experimental trial-and-error methods to find optimal compositions are time-consuming and costly.

This project leverages machine learning (ML) to predict key mechanical properties of HEAs — such as yield strength, ultimate tensile strength, and hardness — based on their elemental composition and processing conditions.

Project Goals:

1. Data Collection & Cleaning: Aggregate a dataset of experimentally characterized HEAs.
   
   

2. Feature Engineering: Compute descriptors like atomic radius difference, mixing entropy, valence electron concentration, etc.
   
   

3. Model Training: Train regression models (e.g., Random Forest, XGBoost, Neural Network) to predict mechanical properties.
   
   

4. Model Interpretation: Use SHAP values to identify key factors affecting property predictions.
   
   

5. Materials Discovery: Suggest new HEA compositions with optimized mechanical properties.

This project focused on forecasting the end-of-day closing price of the S&P 500 index (SPX) over short-term (7-day) and medium-term (28-day) time horizons using a deep learning approach. The primary goal was to build a model capable of capturing the temporal dependencies and nonlinear patterns inherent in historical financial time series data.

To achieve this, I employed Long Short-Term Memory (LSTM) networks — a specialized type of Recurrent Neural Network (RNN) known for its ability to retain long-range dependencies and memory over time. This makes LSTMs particularly well-suited for time-series forecasting tasks like stock market prediction, where past price patterns often influence future trends.

Overview:

In contemporary healthcare, early cancer detection is of paramount importance. Identifying cancerous conditions at an early stage significantly improves patient outcomes and enables more effective treatment strategies. This project aims to leverage the potential of machine learning and deep learning methodologies to predict three prominent types of cancers: oral, cervical, and brain tumors.

Objectives:

• Develop and fine-tune machine learning models for the early detection of oral, cervical, and brain tumors.

• Utilize comprehensive datasets, comprising both image and risk factor data, to train and validate the predictive models.

• Evaluate the performance of the models using metrics such as accuracy, sensitivity, and specificity.

• Establish a foundation for the seamless integration of predictive models into clinical practice, enhancing diagnostic capabilities in cancer detection.

Four kind of analysis is performed in this big dataset and they are Descriptive and Diagnostic analysis,Predictive analysis and Prescriptive analysis. Descriptive and Diagnostic analysis is done to identify hidden insights from data; Predictive analysis is done to predict severity illness determination from machine learning algorithm and prescriptive analysis is done to show how this machine learning algorithm going to help identifying severity illness.

This dataset is imbalanced and has a lot of missing values. As we know working with missing value is challenging and thats why data cleaning and feature engineering is one of the core part of this project.

This is a classification problem and I selected Gradient Boosting Classifiers model for predicting severity illness.Gradient boosting classifiers are a group of machine learning algorithms that combine many weak learning models together to create a strong predictive model. Decision trees are usually used when doing gradient boosting. Gradient boosting models are becoming popular because of their effectiveness at classifying complex datasets.

To handle imbalanced data BorderlineSMOTE was used. I over sampled the minority class by BorderlineSMOTE and then again trained the model to see how it performs. Difference between severity illness prediction with imbalanced data and balanced data results were compared .

For prescriptive analysis LIME(Local Interpret-able Model-agnostic Explanations) was used to explain the model’s result. Prescriptive model describes how this model can be applied in the real life scenario.

Built a machine learning pipeline to detect anomalous performance patterns in digital advertising campaigns across CPA and CPC revenue models.

Key Features:

• Real-time data processing from Elasticsearch with 30-day rolling window

• Isolation Forest models trained separately for CPA/CPC revenue structures

• Advanced feature engineering (profit margin, cost per conversion, conversion rate deviations)

• Automated cleanup maintaining optimal dataset size

• Streamlit dashboard for anomaly visualization and investigation

Technical Stack:
Python, Pandas, Scikit-learn, Elasticsearch, Streamlit, Parquet

Impact:

• Reduced false positives by 40% through program-specific modeling

• Enabled proactive campaign optimization saving estimated 15% in ad spend

• Processed 10M+ daily records with automated data lifecycle management

In this project, I used population statistics from the US Census Bureau to determine and understand thingss that was our main objective of the projects. what i found in this projects are given below:

• Correlation (R^2 value) between sales and income which is 0.78

• Correlation (R^2 value) between customer ratings and product return rate which is 0.50

• Prdicted the incomes of the customers from purchases and sales data

• Which products should we advertise most for most profits

• Which market should we target

A neural network that can recognize images of articles of clothing. Fashion MNIST data was used. The MNIST dataset contains images of handwritten digits (0, 1, 2, etc) in an identical format to the articles of clothing I used here. The images are 28x28 arrays, with pixel values in the range [0,255].

On behalf of my company, I performed many data science related projects which helped the company finding out the targeted customers and business stategies. This is one of the those projects related to email marketing.

On behalf of my company, I performed many data science related projects which helped the company finding out the targeted customers and business stategies. This is one of the those projects related to email marketing.

Mobile apps are everywhere. They are easy to create and can be lucrative. Because of these two factors, more and more apps are being developed. In this notebook, we will do a comprehensive analysis of the Android app market by comparing over ten thousand apps in Google Play across different categories. We'll look for insights in the data to devise strategies to drive growth and retention. 

Barack Obama is left-handed. So are Bill Gates and Oprah Winfrey; so were Babe Ruth and Marie Curie. A 1991 study reported that left-handed people die on average nine years earlier than right-handed people. Nine years! Could this really be true? 

Applying Bayes' rule

The probability of dying at a certain age given that you're left-handed is not equal to the probability of being left-handed given that you died at a certain age. This inequality is why we need Bayes' theorem, a statement about conditional probability which allows us to update our beliefs after seeing evidence.

We want to calculate the probability of dying at age A given that you're left-handed. Let's write this in shorthand as P(A | LH). We also want the same quantity for right-handers: P(A | RH).

Here's Bayes' theorem for the two events we care about: left-handedness (LH) and dying at age A.

P(A|LH)=P(LH|A)P(A)P(LH)

P(LH | A) is the probability that you are left-handed given that you died at age A. P(A) is the overall probability of dying at age A, and P(LH) is the overall probability of being left-handed. We will now calculate each of these three quantities, beginning with P(LH | A).

To calculate P(LH | A) for ages that might fall outside the original data, we will need to extrapolate the data to earlier and later years. Since the rates flatten out in the early 1900s and late 1900s, we'll use a few points at each end and take the mean to extrapolate the rates on each end. The number of points used for this is arbitrary, but we'll pick 10 since the data looks flat-ish until about 1910.

Final comments

We got a pretty big age gap between left-handed and right-handed people purely as a result of the changing rates of left-handedness in the population, which is good news for left-handers: you probably won't die young because of your sinisterness. The reported rates of left-handedness have increased from just 3% in the early 1900s to about 11% today, which means that older people are much more likely to be reported as right-handed than left-handed, and so looking at a sample of recently deceased people will have more old right-handers.

Neural Information Processing Systems (NIPS) is one of the top machine learning conferences in the world where groundbreaking work is published. In this Project, I analyzed a large collection of NIPS research papers from the past decade to discover the latest trends in machine learning. The techniques used here to handle large amounts of data can be applied to other text datasets as well. Here i found out the top 10 machine learning topics from 1987 until 2017.