Data Science Portfolio

Human Activity Recognition

• Predicted Activity based on Accelerometer and Gyroscope readings from smart phone

• Fitted Logistic Regression | Linear SVC | rbf SVM classifier | DecisionTree | Random Forest | GradientBoosting DT on featured data
• LSTM model trained on Raw data
• t-sne Visualization

Amazon Fashion Prediction System

• Predict similar apparel and recommend those apparell based on which apparel(query product) the user is watching.
• Data acquired in policy compliant manner using Amazon API.
• Text based product recommendation, i.e using product title, brand, description, color and price.
• Results with Text Featurized using BOW, TF-IDF, IDF, average word2vec, IDF weighted word2vec are subjectively compared.
• Product Image based model using VGG-16 CNN also trained.
• Final model build as a weighted Nearest neighbor model using Image, Title, Brand and Color
• Final model uses title: Idf-Word2vec, brand: one hot encoding, colour: one hot encoding, image: VGG-16 CNN

Query product

Suggested product

Taxi Demand Prediction

• Predicted number of pickups, given location cordinates(latitude and longitude) and time of the day.
• Time-series forecasting and Regression.
• New York citymap is divided into several regions based on region radius and density of passenger.
• K-means clustering used for geometric division of the city map
• Time series data is divided into 10 minutes interval.
• Frequency domain features are also used along with time domain data to improve the model.
• Linear Regression, Random Forest and Gradient Boosted Decision trees(GBDT) models were trained and tested.
• GBDT model performed better than other models
• Plots shows how the map of the city is divided into different regions, each color is a region

Quora question similarity

• Identify which questions asked on Quora are duplicates of questions that have already been asked.
• Modelled as classification problem with class as duplicate and not duplicate.
• 15 features are extracted, some of them are:
  • word_common: Number of common unique words in question1 and question2.
  • word_total: total number of words in qs1 + total number of words in qs2.
  • word_share: (word_common) / (word_total)
  • common_word_count / min( len(q1 word), len(q2 word))
• Along with the above extracted features question text is featured using tf-idf.
• Fitted Logistic Regression, Linear SVM and XGboost gradient boosted decision trees.
• Question appearance count.

Question appearance count

Word Cloud for Duplicate Question pairs

Word Cloud for non-Duplicate Question pairs

Netflix Movie Recommendation System

• Predicted the rating that a user would give to a movie that they have not yet rated.
• Minimized the difference between predicted and actual rating (RMSE and MAPE).
• Generated user-user and movie-movie similarity matrix.
• Featurized the data further into 13 features.
• Models trained: Surprise-baseline, Surprise-knn-baseline, Surprise-SVD++, XGboost
• Importance in Gradient Boosted Decision Trees

Personalized Cancer Diagnosis

• Classified the given genetic variations/mutations based on evidence from text-based clinical literature into 9 classes.
• Dataset has 3 important attributes 'Gene' 'Variation' and 'Clinical text'.
• Dataset is imbalanced, modeled with and with out balancing to compare results.
• Clinical text featurized using tf-idf. Gene and variation are featured using both One-hot encoding and response coding.
• Dimensionality reduction of text vector using truncated SVD.
• Fitted and compared various models including Logistic Regression, Linear SVC , Naive Bayes and Random forest
• Experimented with stacking the models and also to use max-voting on the results of the models.
• Interpretability of results is important here so a calibrated classifier is stacked with the models
• Ploted Confusion, precision and recall matrices to have more Interpretability.
• Random-forest-precision matrix