README2.md

ml-projects

Title	Objective	Approach
Flipkart Object Detection DataSet Image >	Bounding Box prediction	Designed a visual feature pipeline with attention on the object in image Data Augmentation Technique along with its bounding box Used Single Stage Detector Approach Focal Loss with YOLO and SSD
Amazon Product Review classification DataSet Text >	Classification	Data Cleaning/feature enginnering Linear/Non-Linear Model Deep Learning Attention Model Pretrained Bert Model Ensemble
Hike Friend Recommendation Dataset Very big Dataset(45M observation, graph edge-representation) Relational Feature Category Numerical	Link Prediction	Graph Based features such as (adamic-adar, common-resource-allocation,...) SVD feature for each user Comunity-clustering Subsemble(I did this after competition is over, to understand more about sampling and model building) neighbour-based feature(Removed highly cardinal feature) Also tried Deep learning approach (Graph Embedding), but couldn't handle at that time properly
HDFC Risk Prediction Dataset 2500 unknown predictors	Classification	Feature Understanding(EDA) feature engineering designed feature interaction tools ensemble model using xgboost/lighgbm/catboost and linear/non-linear simple model statistical model to understand the feature importance using p-values
Club Mahindra Hotel Room Price Prediction Dataset Category Numerical Relational Dataset	Regression	date-time based feature Aggregation based feature Relational Features Ensemble using different set of tranformed target space
Cifar-10 Classification using Conditional Feature Dataset Image	Comparison between ResNet and my modified feature pipeline Classification	Developed a weighted feature pipeline using global and local feature. Global feature put constrained on local feature, to specifically focused on features of object in image Better attention map around object, which reflect its learned feature. Improved score by 1.37% over Resnet
Facenet Dataset Image	Face Verification	Matching Network Approach Build a Student-Attentdance hardware using arduino Hard Mining Approach(generate all permutation between classes to handle small dataset) network-in-network approach to handle overfitting as i have very small dataset. Achieved 93% accuracy
Few Shot Learning(Prototype Network) Dataset Image	Classification (training on very small dataset)	Prototype Algorithm implementation There is more to this(will update in future)
Hackerearth Platform Recommendation System Dataset Text	User-Problem Rating Prediction	My main concerns was to handle following question carefully: What is the strongest and weakest area of user? What is the level of problem? What problem user have just solved? If user gets stuck at current problem, what problem should help him(to gain confidence and to improve skill in that area)? Exploration and explotation strategy in recommending problem And many more?
JP.Morgan House Price Prediction Dataset Category Numerical	Regression	Date based feature and Dummy feature Interaction based feature Bayesian optimization out of fold prediction to generate Meta feature for ensemble
Stock Prediction Dataset Time-Series stock prices	Future price prediction Regression	Deep learning approach using RNN and LSTM