decision_tree.py

import pandas as pd
import numpy as np
import random as rand
import scipy.stats as stats
import sys
from multiprocessing import Process
from multiprocessing import Queue

import utilities as util
import constants as cnst
import measures

from node import TreeNode

def decision_tree_parallel(examples, attributes, bin_targets, queue):
    root = decision_tree(examples, attributes, bin_targets)
    queue.put(root)

'''
    Decision tree learning
    Binary data   - binary matrix with N rows and 45 cols
                  - each row is a list of AUs that describe
                  - a certain emotion
    Attributes    - the list of Action Units (AU) that are candidates
                  - for the best attribute at a certain point
    Target vector - emotions vector with 1 for a certain emotion
                  - and 0 otherwise
'''
def decision_tree(examples, attributes, bin_targets):
    all_same = check_all_same(bin_targets)

    if all_same:
        return TreeNode(None, True, bin_targets.iloc[0].iloc[0])
    elif not attributes:
        # Majority Value
        return TreeNode(None, True, majority_value(bin_targets))
    else:
        best_attribute = choose_best_decision_attr(examples, attributes, bin_targets)
        tree = TreeNode(best_attribute)
        for vi in range(0, 2):
            examples_i = examples.loc[examples[best_attribute] == vi]
            indices = examples_i.index.values
            bin_targets_i = bin_targets.ix[indices]

            if examples_i.empty:
                # Majority Value
                return TreeNode(None, True, majority_value(bin_targets))
            else:
                attr = set(attributes)
                attr.remove(best_attribute)
                tree.set_child(vi, decision_tree(examples_i, attr, bin_targets_i))

        return tree

'''
    Helper functions
'''
def check_all_same(df):
    return df.apply(lambda x: len(x[-x.isnull()].unique()) == 1 , axis = 0).all()

def majority_value(bin_targets):
    res = stats.mode(bin_targets[0].values)[0][0]
    return res

def choose_best_decision_attr(examples, attributes, bin_targets):
    max_gain = -sys.maxsize - 1
    index_gain = -1

    # p and n: training data has p positive and n negative examples
    p = len(bin_targets.loc[bin_targets[0] == 1].index)
    n = len(bin_targets.loc[bin_targets[0] == 0].index)

    for attribute in attributes:
        examples_pos = examples.loc[examples[attribute] == 1]
        examples_neg = examples.loc[examples[attribute] == 0]
        index_pos = examples_pos.index.values
        index_neg = examples_neg.index.values

        p0 = n0 = p1 = n1 = 0

        for index in index_pos:
            if bin_targets[0][index] == 1:
                p1 = p1 + 1
            else:
                n1 = n1 + 1

        for index in index_neg:
            if bin_targets[0][index] == 1:
                p0 = p0 + 1
            else:
                n0 = n0 + 1

        curr_gain = gain(p, n, p0, n0, p1, n1)
        if curr_gain > max_gain:
            index_gain = attribute
            max_gain = curr_gain

    if max_gain == -sys.maxsize - 1:
        raise ValueError('Index gain is original value...')

    return index_gain


# Gain(attribute) = I(p, n) – Remainder(attribute)
def gain(p, n, p0, n0, p1, n1):
    return get_info_gain(p, n) - get_remainder(p, n, p0, n0, p1, n1)

# Information Gain I
# I(p, n) = − p+n log 2 ( p+n ) − p+n log 2 ( p+n ) and
def get_info_gain(p, n):
    if p + n == 0:
        return 0
    term_1 = float(p / (p + n))
    term_2 = float(n / (p + n))
    return stats.entropy([term_1, term_2], base=2)

# Remainder(attribute) = (p0 + n0)/(p + n) * I(p0, n0) + (p1 + n1)/(p + n) * I(p1, n1)
def get_remainder(p, n, p0, n0, p1, n1):
    return ((p0 + n0)/(p + n)) * get_info_gain(p0, n0) + ((p1 + n1)/(p + n)) * get_info_gain(p1, n1) if p+n != 0 else 0


'''
    predictions  - DataFrame column with predicted emotions for each test_data_df,
                 - indexes from 1 to 6
    expectations - DataFrame column wtih expected emotions, basically test_data_labels

    Computes confusion matrix by incrementing conf_matrix[expectation[i], prediction[i]]
'''
def compare_pred_expect(predictions, expectations):
    confusion_matrix = pd.DataFrame(0, index=cnst.EMOTIONS_INDICES, columns=cnst.EMOTIONS_INDICES)
    predictions, expectations = predictions.reset_index(drop=True), expectations.reset_index(drop=True)

    for index in predictions.index.values:
        e = expectations.iloc[index] - 1
        p = predictions.iloc[index] - 1
        confusion_matrix.loc[p, e] += 1

    return confusion_matrix
"""
    Input: List with length = 6 of tuples of the form (prediction, depth, percentage)
    prediction : Each tree's prediction for one specific example
    depth: The depth at which the perticular prediction was found on the tree
    percentage: The accuracy of that specific tree based on it's performance on the validation data

    Output: The most accurate prediction
    Three cases: 1. One tree recognized this emotion(a single "1" value in predictions)
                    => return the index of the tree
                 2. Zero trees recognized this emotion =>
                  First Criteria: Choose tree which decided to not recognize it furthest
                                    away from root (highest depth)
                  Second Criteria: Choose tree with lowest accuracy
                 3. Multiple trees recognized this emotion =>
                   First Criterion: Choose tree which recognized it closest to the root
                                    reason: more generality
                   Second Criterion: Choose tree with highest accuracy
"""
def choose_prediction_random(pred_proc_depth):
    predictions, proc, depths = zip(*pred_proc_depth)
    occurrences = [index for index, value in enumerate(predictions) if value == 1]
    if len(occurrences) == 1:
        return occurrences[0]
    elif len(occurrences) == 0:
        return rand.randint(0, 5)
    else:
        return rand.choice(occurrences)


def choose_prediction_optimised(pred_proc_depth):
    predictions, proc, depths  = zip(*pred_proc_depth)
    indexes = [index for index, value in enumerate(predictions) if value == 1]

    if len(indexes) == 1:
        return indexes[0]
    elif len(indexes) == 0:
        res = 0
        MAX = 0
        max_depth_indexes = []
        for i in range(0, len(depths)):
            if depths[i] > MAX:
                MAX = depths[i]
                del max_depth_indexes[:]
                max_depth_indexes.append(i)
            elif depths[i] == MAX:
                max_depth_indexes.append(i)
        if len(max_depth_indexes) == 1:
            res = max_depth_indexes[0]
        else:
            min_proc = 100
            min_proc_index = 0
            for i in max_depth_indexes:
                if (proc[i] < min_proc):
                    min_proc = proc[i]
                    min_proc_index = i
            res = min_proc_index
        return res
    else:
        res = 0
        MIN = 10000
        max_depth_indexes = []
        for i in indexes:
            if depths[i] < MIN:
                MIN = depths[i]
                del max_depth_indexes[:]
                max_depth_indexes.append(i)
            elif depths[i] == MIN:
                max_depth_indexes.append(i)
        if len(max_depth_indexes) == 1:
            res = max_depth_indexes[0]
        else:
            max_proc = 0
            max_proc_index = 0
            for i in max_depth_indexes:
                if (proc[i] > max_proc):
                    max_proc = proc[i]
                    max_proc_index = i
            res = max_proc_index
        return res

'''
    Takes your trained trees (all six) T and their Precision and the features x2 and
    produces a vector of label predictions
'''
def test_trees(T_P, x2):

    T, P = zip(*T_P)

    predictions = []

    for i in x2.index.values:
        example = x2.loc[i]
        T_P_D = []
        for j in range(0, len(T_P)):
            prediction, depth = TreeNode.dfs_with_depth(T[j], example)
            T_P_D.append([prediction, P[j], depth])

        prediction_choice = choose_prediction_optimised(T_P_D)
        predictions.append(prediction_choice + 1)

    return pd.DataFrame(predictions)

def visualise(df_labels, df_data, N):
    for e in cnst.EMOTIONS_LIST:
        root = decision_tree(df_data, set(cnst.AU_INDICES), util.filter_for_emotion(df_labels, cnst.EMOTIONS_DICT[e]))
        TreeNode.plot_tree(root, e)

def apply_d_tree_parallel(df_labels, df_data, N):
    print(">> Running decision tree algorithm on multiple processes.\n")

    res = pd.DataFrame(0, index=cnst.EMOTIONS_INDICES, columns=cnst.EMOTIONS_INDICES)

    segments = util.preprocess_for_cross_validation(N)

    for test_seg in segments:
        print(">> Starting fold... from:", test_seg)
        print()

        T = []
        # Split data into 90% Training and 10% Testing
        test_df_data, test_df_targets, train_df_data, train_df_targets = util.divide_data(test_seg, N, df_data, df_labels)

        # Further split trainig data into 90% Training and 10% Validation data
        K = train_df_data.shape[0]
        segs = util.preprocess_for_cross_validation(K)
        validation_data, validation_targets, train_data, train_targets = util.divide_data(segs[-1], K, train_df_data, train_df_targets)

        processes = []
        queue_list = []

        for e in cnst.EMOTIONS_LIST:
            print("Building decision tree for emotion...", e)
            train_binary_targets = util.filter_for_emotion(train_df_targets, cnst.EMOTIONS_DICT[e])

            q = Queue()
            queue_list.append(q)

            process = Process(target=decision_tree_parallel, args=(train_df_data, set(cnst.AU_INDICES), train_binary_targets, q))
            processes.append(process)
            process.start()

        for p in processes:
            p.join()

        for q in queue_list:
            T.append(q.get())

        # Use validation data to set a priority to each tree based on which is more accurate
        percentage = []
        T_P = []
        for e in cnst.EMOTIONS_LIST:
            print("\nValidation phase for emotion: ", e)
            validation_binary_targets = util.filter_for_emotion(validation_targets, cnst.EMOTIONS_DICT[e])
            results = []
            # Calculate how accurate each tree is when predicting emotions
            for i in validation_data.index.values:
                results.append(TreeNode.dfs2(T[cnst.EMOTIONS_DICT[e]- 1], validation_data.loc[i], validation_binary_targets.loc[i].at[0]))
            ones = results.count(1)
            percentage.append(ones/len(results))
            print("Validation phase ended. Priority levels have been set.")

        print("All decision trees built.\n")

        T_P = list(zip(T, percentage))

        predictions = test_trees(T_P, test_df_data)
        confusion_matrix = compare_pred_expect(predictions, test_df_targets)

        print(confusion_matrix)
        # Print accuracy for each fold
        diag = sum(pd.Series(np.diag(confusion_matrix),
                            index=[confusion_matrix.index, confusion_matrix.columns]))
        sum_all = confusion_matrix.values.sum()
        accuracy = (diag/sum_all) * 100
        print("Accuracy:", accuracy)

        res = res.add(confusion_matrix)

    # res = res.div(10)
    res = res.div(res.sum(axis=1), axis=0)

    for e in cnst.EMOTIONS_LIST:
        print("----------------------------------- MEASUREMENTS -----------------------------------")
        print(measures.compute_binary_confusion_matrix(res, cnst.EMOTIONS_DICT[e]))

    return res


'''
    Computes a confusion matrix using decison trees only.
    Does N-folds, for each of them the following algo been applied:
        - take N - 1 training data/training targets
        - make decision trees
        - gets the best prediction based on decision trees
        - compare predictions with expectations (df_test_labels)
'''
def apply_d_tree(df_labels, df_data, N):
    print(">> Running decision tree algorithm on a single process.\n")

    res = pd.DataFrame(0, index=cnst.EMOTIONS_INDICES, columns=cnst.EMOTIONS_INDICES)

    segments = util.preprocess_for_cross_validation(N)
    total_accuracy = 0
    for test_seg in segments:
        print(">> Starting fold... from:", test_seg)
        print()

        T = []
        # Split data into 90% Training and 10% Testing
        test_df_data, test_df_targets, train_df_data, train_df_targets = util.divide_data(test_seg, N, df_data, df_labels)

        # Further split trainig data into 90% Training and 10% Validation data
        K = train_df_data.shape[0]
        segs = util.preprocess_for_cross_validation(K)
        validation_data, validation_targets, train_data, train_targets = util.divide_data(segs[-1], K, train_df_data, train_df_targets)

        # Train Trees
        for e in cnst.EMOTIONS_LIST:
            print("Building decision tree for emotion: ", e)
            train_binary_targets = util.filter_for_emotion(train_df_targets, cnst.EMOTIONS_DICT[e])
            root = decision_tree(train_data, set(cnst.AU_INDICES), train_binary_targets)
            print("Decision tree built. Now appending...")
            T.append(root)

        # Use validation data to set a priority to each tree based on which is more accurate
        percentage = []
        T_P = []
        for e in cnst.EMOTIONS_LIST:
            print("\nValidation phase for emotion: ", e)
            validation_binary_targets = util.filter_for_emotion(validation_targets, cnst.EMOTIONS_DICT[e])
            results = []
            # Calculate how accurate each tree is when predicting emotions
            for i in validation_data.index.values:
                results.append(TreeNode.dfs2(T[cnst.EMOTIONS_DICT[e]- 1], validation_data.loc[i], validation_binary_targets.loc[i].at[0]))
            ones = results.count(1)
            percentage.append(ones/len(results))
            print("Validation phase ended. Priority levels have been set.")

        print("All decision trees built.\n")

        # List containing (Tree, Percentage) tuples
        T_P = list(zip(T, percentage))

        predictions = test_trees(T_P, test_df_data)
        confusion_matrix = compare_pred_expect(predictions, test_df_targets)

        print(confusion_matrix)
        # Print accuracy for each fold
        diag = sum(pd.Series(np.diag(confusion_matrix),
                            index=[confusion_matrix.index, confusion_matrix.columns]))
        sum_all = confusion_matrix.values.sum()
        accuracy = (diag/sum_all) * 100
        total_accuracy += accuracy
        print("Accuracy:", accuracy)

        res = res.add(confusion_matrix)
        print("Folding ended.\n")
        print()
    print("Total accuracy:", accuracy)
    res = res.div(res.sum(axis=1), axis=0)
    print(res)
    return res


    res = res.div(res.sum(axis=1), axis=0)

    for e in cnst.EMOTIONS_LIST:
        print("----------------------------------- MEASUREMENTS -----------------------------------")
        print(measures.compute_binary_confusion_matrix(res, cnst.EMOTIONS_DICT[e]))

    return res