|
| 1 | +import os |
| 2 | +import pandas as pd |
| 3 | +import joblib |
| 4 | +from datetime import datetime |
| 5 | +from tqdm import tqdm |
| 6 | +from collections import defaultdict, Counter |
| 7 | +import math |
| 8 | +import numpy as np |
| 9 | +import random |
| 10 | +import copy |
| 11 | +import gc |
| 12 | +from gensim.models import Word2Vec |
| 13 | +from sklearn.neighbors import NearestNeighbors |
| 14 | + |
| 15 | +def load_datasets_and_mappings(): |
| 16 | + """Load training data and id-to-type mappings.""" |
| 17 | + training_data = pd.read_parquet('../input/otto-full-optimized-memory-footprint/train.parquet') |
| 18 | + id_to_type_mapping = joblib.load('../input/otto-full-optimized-memory-footprint/id2type.pkl') |
| 19 | + type_to_id_mapping = joblib.load('../input/otto-full-optimized-memory-footprint/type2id.pkl') |
| 20 | + |
| 21 | + return training_data, id_to_type_mapping, type_to_id_mapping |
| 22 | + |
| 23 | + |
| 24 | +def preprocess_training_data(training_data, config): |
| 25 | + """Preprocess the training data.""" |
| 26 | + training_data['aid'] = training_data['aid'].astype('int32').astype('str') |
| 27 | + |
| 28 | + # Randomly sample sessions for training |
| 29 | + sampled_sessions = random.sample(list(training_data['session'].unique()), config['train_session_num']) |
| 30 | + training_data = training_data.query('session in @sampled_sessions').reset_index(drop=True) |
| 31 | + |
| 32 | + training_data['time_stamp'] = pd.to_datetime(training_data['ts'], unit='s').dt.strftime('%Y-%m-%d') |
| 33 | + |
| 34 | + return training_data |
| 35 | + |
| 36 | + |
| 37 | +def generate_word2vec_embeddings(data): |
| 38 | + """Generate Word2Vec embeddings for session sequences.""" |
| 39 | + session_sequences = data.groupby('session')['aid'].apply(list).tolist() |
| 40 | + |
| 41 | + # Train Word2Vec model |
| 42 | + model = Word2Vec(session_sequences, min_count=1, sg=1) |
| 43 | + word_vectors = model.wv |
| 44 | + |
| 45 | + return word_vectors |
| 46 | + |
| 47 | + |
| 48 | +def recommend_items(session_items, word_vectors, nearest_neighbors, popular_items): |
| 49 | + """Recommend items based on the given session items using Word2Vec and nearest neighbors.""" |
| 50 | + item_embeddings = [] |
| 51 | + for item in session_items: |
| 52 | + if item in word_vectors: |
| 53 | + item_embeddings.append(word_vectors[item]) |
| 54 | + |
| 55 | + if len(item_embeddings) > 0: |
| 56 | + session_embedding = np.mean(item_embeddings, axis=0) |
| 57 | + _, indices = nearest_neighbors.kneighbors([session_embedding]) |
| 58 | + similar_items = nearest_neighbors._fit_X[indices.flatten()] |
| 59 | + recommended_items = [item for item in similar_items[0] if item not in session_items] |
| 60 | + recommended_items = recommended_items[:20] # Limit to 20 recommendations |
| 61 | + else: |
| 62 | + recommended_items = [] |
| 63 | + |
| 64 | + if len(recommended_items) < 20: |
| 65 | + return recommended_items + popular_items[:20 - len(recommended_items)] |
| 66 | + else: |
| 67 | + return recommended_items |
| 68 | + |
| 69 | + |
| 70 | +def load_and_preprocess_test_data(): |
| 71 | + """Load and preprocess test data.""" |
| 72 | + test_data = pd.read_parquet('../input/otto-full-optimized-memory-footprint/test.parquet') |
| 73 | + test_data['aid'] = test_data['aid'].astype('int32').astype('str') |
| 74 | + test_data['time_stamp'] = pd.to_datetime(test_data['ts'], unit='s').dt.strftime('%Y-%m-%d') |
| 75 | + test_data = test_data.sort_values(["session", "type", "ts"]) |
| 76 | + session_to_item_ids = test_data.groupby('session')['aid'].agg(list).to_dict() |
| 77 | + |
| 78 | + return session_to_item_ids |
| 79 | + |
| 80 | + |
| 81 | +def generate_recommendations(session_to_item_ids, word_vectors, nearest_neighbors, popular_items): |
| 82 | + """Generate item recommendations for each session.""" |
| 83 | + session_ids = [] |
| 84 | + recommended_item_lists = [] |
| 85 | + for session_id, session_items in tqdm(session_to_item_ids.items()): |
| 86 | + recommended_items = recommend_items(session_items, word_vectors, nearest_neighbors, popular_items) |
| 87 | + session_ids.append(session_id) |
| 88 | + recommended_item_lists.append(recommended_items) |
| 89 | + |
| 90 | + return session_ids, recommended_item_lists |
| 91 | + |
| 92 | + |
| 93 | +def create_submission_file(session_ids, recommended_item_lists, id_to_type_mapping): |
| 94 | + """Create a submission file with the recommended items for each session type.""" |
| 95 | + submission_df = pd.DataFrame() |
| 96 | + submission_df['session_type'] = session_ids |
| 97 | + submission_df['labels'] = [' '.join([str(item) for item in item_list]) for item_list in recommended_item_lists] |
| 98 | + |
| 99 | + submission_list = [] |
| 100 | + for type_ in [0, 1, 2]: |
| 101 | + type_specific_df = submission_df.copy() |
| 102 | + type_specific_df['session_type'] = type_specific_df['session_type'].apply(lambda x: f'{x}_{id_to_type_mapping[type_]}') |
| 103 | + submission_list.append(type_specific_df) |
| 104 | + submission_df = pd.concat(submission_list, axis=0) |
| 105 | + |
| 106 | + submission_df.to_csv('submission.csv', index=False) |
| 107 | + |
| 108 | + |
| 109 | +def main(): |
| 110 | + config = {'train_session_num': 12899779} |
| 111 | + training_data, id_to_type_mapping, _ = load_datasets_and_mappings() |
| 112 | + training_data = preprocess_training_data(training_data, config) |
| 113 | + word_vectors = generate_word2vec_embeddings(training_data) |
| 114 | + session_sequences = training_data.groupby('session')['aid'].apply(list).tolist() |
| 115 | + nearest_neighbors = NearestNeighbors(metric='cosine') |
| 116 | + nearest_neighbors.fit(word_vectors[session_sequences]) |
| 117 | + del training_data, session_sequences |
| 118 | + gc.collect() |
| 119 | + |
| 120 | + session_to_item_ids = load_and_preprocess_test_data() |
| 121 | + popular_items = list(training_data['aid'].value_counts().index) |
| 122 | + session_ids, recommended_item_lists = generate_recommendations(session_to_item_ids, word_vectors, nearest_neighbors, popular_items) |
| 123 | + create_submission_file(session_ids, recommended_item_lists, id_to_type_mapping) |
| 124 | + |
| 125 | + |
| 126 | +if __name__ == "__main__": |
| 127 | + main() |
0 commit comments