Implement in Java or C++ or Python (or other language that the tutors will allow you) two or three (depending on whether your team has two or three members) by the following learning algorithms so that they can be used to rank texts into two (unrelated) categories (e.g. positive/negative opinion)
• Naïve Bayes classifier, multivariate Bernoulli form or polynomial form.
• ID3, optionally with sawing or early extension termination of every tree.
• Random Forest, using ID3 or its variant (eg where will produce trees that will not exceed a maximum depth, which will be given as a hyper-parameter) to produce the trees.
Each text should be represented by a vector of properties with values 0 or 1, which will indicate which vocabulary words the text contains. The vocabulary should include the m most frequent words of of training data, first skipping the n most frequent and the k most rare words of the training data, where m, n, k will be hyperparameters. Optionally you can add a selection of properties through information gain calculation (or through other means) to naive Bayes classifier. Others algorithms already incorporate feature selection methods.
Demonstrate the learning capabilities of your implementations using the 'Large Movie Review Dataset', which is known and as "IMDB dataset" (see https://ai.stanford.edu/~amaas/data/sentiment/, https://keras.io/api/datasets/imdb/). You should include in the report your results of the experiments you will perform with your implementations in this data set, showing (at least): • learning curves and corresponding tables showing the percentage of correctness (accuracy) in the training data (training data, which have been used each time) and control (test data) function of the number of training examples that are used in each repetition of the experiment • corresponding curves and tables with accuracy results (precision), recall, F1 for one of the two categories, as a function of the number of training examples.
Compare the performance of your implementations with theirs performance of other available implementations (e.g. Weka or Scikit-learn) of the same learning algorithms that you implemented in part A' or others learning algorithms (e.g. MLPs implementations of Scikit-learn), constructing the same curves and tables as in Part A'. I will need you to compare with at least two or three available implementations of learning algorithms. You should use the same representations of the texts as in Part A (feature vectors with values of 0 or 1 that will indicate which words of vocabulary appear in the text or not, with the same vocabulary as in Part A'). Here, too, you can use ready-made implementations of text preprocessing and property selection, as well as ready-made libraries for constructing curve charts.
Compare the results of Parts A' and B' with the results of a rolling window MLP and/or an RNN, which will implement in Tensorflow/Keras, representing the words with nests word embeddings. Construct the same curves here as well tables, as in parts A and B. Also construct curves that to show the change in error (loss) in the training examples and growth, depending on the number of seasons.