-
Toxic Comments Classification - Build a multi-headed model that predicts probabilities of different types of toxicity such as threats, obscenity, insults, and identity-based in the comments.
-
Metrics: AUC ROC
-
Models used: Logistic Regression and RNN (LSTM/GRU)
-
Approach:
-
Used Matplotlib and Seaborn to visualize distribution of target classes, correlation of categorical variables using confusion matrix and frequently used words using WordCloud.
-
Created more features and visualized those features to understand whether longer comments are more toxic, etc
-
Vectorized the words using TF-IDF and used LogisticRegression with additional features created from step 2 and without those features.
-
Corrected spelling mistakes using FastText embeddings and trained a RNN network with LSTM and GRU cells using Keras.
-
-
-
StackOverFlow Tag Prediction - Suggest tags based on the title and question text (multi-class classification).
-
Metrics: Mean F1 Score
-
Models used: Logistic Regression with OneVsRest Classifier, SGDClassifier with hinge loss
-
Approach:
-
Used Matplotlib and Seaborn to visualize distribution of tags, tags per question and most frequent tags.
-
Preprocessed the data using NLTK to remove stopwords, special characters, HTML tags and SnowballStemmer to stem the words. Saved the cleaned data to SQLite DB.
-
Vectorized the words using TF-IDF and used LogisticRegression with One Vs Rest Classifier and SGDClassifier with hinge loss. Used GridSearch to tune hyperparameters.
-
-
-
Quora Question Pair Similarity – Identify whether a given pair of questions is similar (binary classification).
-
Metrics: LogLoss
-
Models used: Random Model, Logistic Regression, Linear SVM and XGBoost.
-
-
Santander Product Recommendation – Build a recommendation system so Santander can cross sell products to their existing customers (multi-class classification).
-
Metrics: LogLoss
-
Models used: XGBoost, Matrix Factorization Techniques – SVD
-
Approach:
-
Explained the approach in a blog.
-
Used Matplotlib and Seaborn to visualize products distribution by various features.
-
Converted the dataset to be used as a multi-class classification problem.
-
Created new feature lags and statistics metrics such as min, max and standard deviations.
-
Modeled using XGBoost and Matrix Factorization techniques such as SVD.
-
-