p@15: # -*- coding: utf-8 -*-
p@15: """
p@15: Created on Wed Jul 22 17:42:09 2015
p@15: 
p@15: @author: paulochiliguano
p@15: """
p@15: 
p@16: 
p@25: from math import log, sqrt
p@15: import numpy as np
p@24: import pandas as pd
p@25: import cPickle as pickle
p@26: import time
p@15: 
p@25: # Item-vector dictionary
p@25: f = file('/Users/paulochiliguano/Documents/msc-project/dataset/\
p@25: genre_classification/genre_prob.pkl', 'rb')
p@25: song_library = pickle.load(f)
p@25: f.close()
p@24: 
p@25: # Load training and test data
p@25: f = file('/Users/paulochiliguano/Documents/msc-project/dataset/\
p@25: cross_validation.pkl', 'rb')
p@25: users_train, users_test = pickle.load(f)
p@25: f.close()
p@25: 
p@25: # Cosine Similarity
p@25: def cosine_similarity(vector1, vector2):
p@25:     dot_product = sum(map(lambda x, y: x * y, vector1, vector2))
p@25:     length_x = sqrt(sum(map(lambda x: x ** 2, vector1)))
p@25:     length_y = sqrt(sum(map(lambda y: y ** 2, vector2)))
p@25:     return dot_product / (length_x * length_y)
p@25: 
p@25: # Adjusted Cosine Similarity
p@25: def adj_cos_sim(vector_i, vector_j):
p@25:     avrg_w_i = (float(sum(vector_i)) / len(vector_i))
p@25:     avrg_w_j = (float(sum(vector_j)) / len(vector_j))
p@25:     num = sum(map(
p@25:         lambda w_i, w_j: (w_i - avrg_w_i) * (w_j - avrg_w_j),
p@25:         vector_i,
p@25:         vector_j)
p@25:     )
p@25:     dem1 = sum(map(lambda w_i: (w_i - avrg_w_i) ** 2, vector_i))
p@25:     dem2 = sum(map(lambda w_j: (w_j - avrg_w_j) ** 2, vector_j))
p@25:     return num / (sqrt(dem1) * sqrt(dem2))
p@25: 
p@25: # Fitness function for EDA
p@25: def Fitness(profile_u, user_subset):
p@25:     fitness_value = 0
p@25:     for songID, score in user_subset.iteritems():
p@25:         #print cosine_similarity(profile, song_library[songID])
p@25:         sim = cosine_similarity(profile_u, song_library[songID])
p@25:         if sim <= 0:
p@25:             fitness_value += -708
p@25:             #math.log(sys.float_info.min)
p@25:         else:
p@25:             fitness_value += log(score * sim)
p@25:         #fitness_value += log(score * manhattan(profile, song_library[songID]))
p@25:         #fitness_value += score * cosine_similarity(profile, song_library[songID])
p@25:     return fitness_value
p@25: 
p@26: def users_likes_subset(users, rating_threshold=2):
p@25:     # Subset of most-liked items
p@25:     users_subset = {}
p@25:     for userID, songs in users.iteritems():
p@25:         scores_above_threshold = {
p@25:             songID: score for songID, score in songs.iteritems() if score > rating_threshold
p@15:         }
p@25:         users_subset[userID]= scores_above_threshold
p@25:         
p@25:     #for songID, score in songs.iteritems():
p@25:         #print score >0
p@25:         #if score > 0:
p@25:             #print {userID: {songID: score}}
p@15: 
p@25:     #{k: v for k, v in users.iteritems() for i,j in v.iteritems() if j > 0}
p@25:         
p@25:     return users_subset
p@15: 
p@26: def eda_train(users_subset, max_gen=250):
p@25:     # TRAINING
p@25:     num_features = len(song_library.values()[0])
p@25:     # Given parameters for EDA
p@25:     population_size = len(users_subset)
p@25:     fraction_of_population = int(round(0.5 * population_size))
p@23: 
p@25:     # Generation of M individuals uniformly
p@25:     np.random.seed(12345)
p@25:     M = np.random.uniform(
p@25:         0,
p@25:         1,
p@25:         population_size * num_features
p@25:     )
p@25:     M.shape = (-1, num_features)
p@25:     profile_u = {}
p@25:     i = 0
p@25:     for userID in users_subset:
p@25:         profile_u[userID] = M.tolist()[i]
p@25:         i += 1
p@23: 
p@26:     fitnesses = []
p@25:     generation = 0
p@25:     while generation < max_gen:
p@25:         # Compute fitness values
p@25:         users_fitness = {}
p@25:         for userID in profile_u:
p@25:             users_fitness[userID] = Fitness(
p@25:                 profile_u[userID],
p@25:                 users_subset[userID]
p@25:             )
p@25:         users_fitness_df = pd.DataFrame(
p@25:             users_fitness.items(),
p@25:             columns=["userID", "fitness"]
p@25:         )
p@26:         fitnesses.append(users_fitness_df.fitness.values.tolist())
p@26:         
p@25:         # Selection of best individuals based on fitness values
p@25:         best_individuals = {}
p@25:         users_fitness_df = users_fitness_df.sort(columns='fitness')
p@25:         M_sel = users_fitness_df.tail(fraction_of_population)
p@25:         M_sel_dict = M_sel.set_index('userID')['fitness'].to_dict()
p@25:         for userID in M_sel_dict:
p@25:             best_individuals[userID] = profile_u[userID]
p@25:         
p@25:         # Calculate sample mean and standard deviation
p@25:         D = np.array([])
p@25:         for userID, features in best_individuals.iteritems():
p@25:             D = np.append(D, features, axis=0)
p@25:         D.shape = (-1, num_features)    
p@25:         D_mu = np.mean(D, axis=0)
p@25:         D_sigma = np.std(D, axis=0, ddof=1)
p@25:     
p@25:         # Sample M individuals
p@25:         M = np.random.normal(
p@25:             D_mu,
p@25:             D_sigma,
p@25:             (population_size, num_features)
p@25:         )
p@25:         #M = 1 / (D_sigma * np.sqrt(2 * np.pi)) * np.exp(- (M_range - D_mu) ** 2 / (2 * D_sigma ** 2))
p@25:     
p@25:         #M.shape = (-1, len(items.values()[0]))
p@25:         #M = D_sigma * np.random.normal(
p@25:             #population_size,
p@25:             #len(items.values()[0])
p@25:         #) + D_mu
p@25:         profile_u = {}
p@25:         i = 0
p@25:         for userID in users_subset:
p@25:             profile_u[userID] = M.tolist()[i]
p@25:             i += 1
p@25:         generation += 1
p@25:     
p@26:     return profile_u, D, np.array(fitnesses)
p@25: 
p@25: # Similarity matrix
p@25: def cb_similarity(profileID, profile_data, test_data, N):
p@26:     
p@26:     a = []
p@26:     for user, info in test_data.iteritems():
p@26:         a.extend([i for i in info])
p@26:     songIDs = list(set(a))
p@26:     
p@25:     ''' Content-based: Similarity matrix '''
p@25:     similarity = []
p@26:     for songID in songIDs:
p@25:         sim = adj_cos_sim(profile_data, song_library[songID])
p@25:         similarity.append((sim, songID))
p@26:     
p@26:     ''' Top-N recommendation '''
p@26:     similarity.sort(reverse=True)
p@26:     if len(similarity) > N:
p@26:         similarity = similarity[0:N]
p@25:         
p@25:     #sim_matrix[userID] = {t[1]: t[0] for t in similarity}
p@25:     return {t[1]: t[0] for t in similarity}
p@25: 
p@25: def evaluate_eda(
p@25:     profiles,
p@25:     test_data,
p@25:     N=10,
p@26:     rating_threshold=2,
p@26:     EDA_treshold=0.5):    
p@25:     
p@25:     ''' Evaluation '''
p@25:     
p@25:     sim_matrix = {}
p@25:     for userID, features in profiles.iteritems():
p@25:         sim_matrix[userID] = cb_similarity(userID, features, test_data, N)
p@25:     
p@25:     # Content-Based: Evaluation
p@25:     tp = 0.
p@25:     fp = 0.
p@25:     fn = 0.
p@25:     tn = 0.
p@25:     
p@26:     for user, song_rating in test_data.iteritems():
p@26:         entries = sim_matrix[user]
p@26:         for song, rating in song_rating.iteritems():
p@26:             if song in entries:
p@26:                 if rating > rating_threshold:
p@26:                     tp += 1
p@26:                 elif rating <= rating_threshold:
p@26:                     fp += 1   
p@26:             else:
p@26:                 if rating > rating_threshold:
p@26:                     fn += 1
p@26:                 elif rating <= rating_threshold:
p@26:                     tn += 1
p@26:     
p@26:     
p@26: #    for userID, songID_sim in sim_matrix.iteritems():
p@26: #        for songID, sim_value in songID_sim.iteritems():
p@26: #            score = test_data[userID][songID]
p@26: #            if score > rating_threshold and sim_value >= EDA_treshold:
p@26: #                tp += 1
p@26: #            elif score <= rating_threshold and sim_value >= EDA_treshold:
p@26: #                fp += 1
p@26: #            elif score > rating_threshold and sim_value < EDA_treshold:
p@26: #                fn += 1
p@26: #            elif score <= rating_threshold and sim_value < EDA_treshold:
p@26: #                tn += 1
p@27:     #print tp, fp, fn, tn
p@26:     if tp != 0:
p@26:         precision = tp / (tp + fp)
p@26:         recall = tp / (tp + fn)
p@26:         F1 = 2 * precision * recall / (precision + recall)
p@26:     else:
p@26:         precision = 0
p@26:         recall = 0
p@26:         F1 = 0
p@26:     
p@25:     accuracy = (tp + tn) / (tp + fp + tn + fn)
p@25:     
p@25:     return precision, recall, F1, accuracy
p@25:  
p@25:         #keys_a = set(users[userID].keys())
p@25:         #keys_b = set(test_data.keys())
p@25:         #intersection = keys_a & keys_b
p@25:         #if len(intersection) != 0:
p@25:             #similarity = {}  
p@25:         #print {k: v for k,v in song_library_fold[0].iteritems() if k in songs}
p@25:             #for songID in intersection:
p@25:             #if songID == k:
p@25:                 #similarity[songID] = adj_cos_sim(
p@25:                     #profile[userID],
p@25:                     #test_data[songID]
p@25:                 #)
p@25:                 #max_sim = max(similarity, key=similarity.get)
p@25:                 #if max_sim >= EDA_treshold:
p@25:                     #sim_matrix[userID] = {max_sim: similarity[max_sim]}
p@25:             #sim_matrix[userID] = similarity
p@25:             #sim_matrix[userID] = {max_sim: similarity[max_sim]}
p@25: 
p@25: #print len(sim_matrix)
p@25: p = np.array([])
p@25: f = np.array([])
p@25: r = np.array([])
p@25: a = np.array([])
p@25: 
p@25: for i in range(len(users_train)):
p@26:     start_time = time.time()
p@26:     profile_u, prob, fffitness = eda_train(users_likes_subset(users_train[i]))
p@26:     elapsed_time = time.time() - start_time
p@26:     print 'Training execution time: %.3f seconds' % elapsed_time
p@25:     
p@27:     pi, ri, fi, ai = evaluate_eda(profile_u, users_test[i], N=20)
p@25:     p = np.append(p, pi)
p@25:     r = np.append(r, ri)
p@25:     f = np.append(f, fi)
p@25:     a = np.append(a, ai)
p@25: 
p@25: #precision = np.array(p)
p@25: #rec = np.array(r)
p@25: #F1 = np.array(f)
p@25: #accuracy = np.array(a)
p@25: 
p@25: print "Precision = %f3 ± %f3" % (p.mean(), p.std())
p@25: print "Recall = %f3 ± %f3" % (r.mean(), r.std())
p@25: print "F1 = %f3 ± %f3" % (f.mean(), f.std())
p@25: print "Accuracy = %f3 ± %f3" % (a.mean(), a.std())