Yash Kanojia

Cryptocurrency price prediction

Objective: To predict price of cryptocurrencies based on historical data and improve previously built model using LSTM model.

Approach used:

• Calculate important technical indicators(Ichimoku cloud,Relative Strength Index,Moving average etc) for cryptocurrency using Ta-Lib python library(used in technical analysis of financial market).
• Create Keras-LSTM(Long Short Term Memory) model including technical indicator and compared its results with previous model(given by client).
• Used sentiment analysis to understand trend using tweets mentioning particular cryptocurrency.
• Included market emotions such as optimism,trust,fear obtained from news API to build another model and check results. This time I had better results then model with only technical indicator.

Alzheimer Stage Prediction

Objective: To predict stage of Alzheimer Disease with the help CNN image classifier model.

Approach used:

• The data used for model building is from Alzheimer's Disease Neuroimaging Initiative(ADNI) which is open-source data for Alzheimer disease.
• Data collected is in nested format with each folder representing the data of various stages of a single patient.
• Data is segregated into three class(AD,MCI,CN) for all patients.
• The slices of brain images(~30) is grouped together to form a single image for a patient.
• CNN classifier model is trained on image data formed above.

Recommendation Engine for Gift website

Objective: To recommend products to user based on their interactions and demographic information using Hybrid recommendation model.

Approach used:

• The data is generated in real time from Mobile IOS app and stored in Mix-Panel.
• Data is downloaded periodically from Mix-Panel along with product data from MongoDB.
• Interactions is generated from Mix-Panel data. Demographic information is provided such as country,region,city,age,gender etc
• For Base model product impressions(views) is converted into ratings(More views means more rating). After that Matrix Factorization is used to predict rating for users and products. This model does not take any features into consideration.
• The best model is created by LightFM which can handle implicit data for recommendation.
• Feature engineering is done to ensure suitable number of features are present in model training to learn more about user.