SortingHat

Harry Potter Classification Model using Screenplay and Book data

Data

The features consists of Book, House, Character, and Text

• Book - title of the harry potter book with the medium (book/movie)
• House - five levels, 4 houses and non entry which are characters who are not in the houses
• Character - characters that are also separated by medium
• Text - unprocessed sentences

Feature Engineering

Bag of Words

• A name list of all unique words from the dataset will be created as a set of features for bag of words
• Bag of words will be counted and joined to each sentence_id that corresponds with the character and its sentence

Bigrams

• A bigram list of bigrams that have been observed more than 1 will be created as a set of features for bigrams
• Bigrams will be counted and joined to each sentence_id that corresponds with the character and its sentence

Trigrams

• A trigram list of trigrams that have been observed more than 1 will be created as a set of features for trigrams
• Trigrams will be counted and joined to each sentence_id that corresponds with the character and its sentence

TF-IDF

• TF-IDF will be used in conjunction with bag of words and ngrams to find unique words/ngrams for each house
• This will act as a feature reduction
• Top_N will be used to find the top n bag of words/ngrams for each house
• N will be set arbitrarily however will try to balance word diversity and how well it generalizes the data

Sentiment Lexicon

• Sentiment will be measured for each character
• Multiple lexicons will be used (Huliu, Jockers Rinker, NRC, Senticnet, Sentiword, SlangSD, Socal Google, and NRC emotions)
• The sentiment will be measured by setence level however, will be averaged for each character
• The averaging is done through simple averaging instead of weighted average

Pre-Processing for Model Training

• Once features are generated, the feature lists will need to be saved to create feature constraints for future data
• Ex: if "dog" was never said and the new data contains the word "dog", "dog" will be ignored and only words in the list will be used
• The data will then be condensed/grouped by each character,
• Bow/Ngrams may need to be centered due to character dialogue variance

Modeling

• The model will be a multiclass classification problem
• Multiple models will be created and once tuned, will be ensembled
• Models will range from simple logistic regression model to Xgboost and SVM models
• Model ensembling will be done through stacking with Xgboost

Model List

Multinomial Logistic Regression (Completed)

• method = lda

Naive Bayes Classifier (Completed)

• method = naive_bayes

Regularized Logistic Regression (L1) (Completed)

• method = glmnet (Alpha = 1)

Penalized Logistic Regression (L2) (Completed)

• method = glmnet (Alpha = 0)

Elastic Net (Completed)

• method = glmnet

Multivariate Adaptive Regression Spline (Completed)

• method = earth

K-Nearest Neighbor (Completed)

• method = knn

Random Forest (Completed)

• method = rf

Support Vector Machine (Completed)

• method = svmLinear

Extreme Gradient Boosting (Completed)

• method = XGBoostLinear