def main(): # Data loading root_data_path = os.path.join(get_project_root_path(), "data/") data_reader = RecSys2019Reader(root_data_path) data_reader = New_DataSplitter_leave_k_out( data_reader, k_out_value=K_OUT, use_validation_set=False, allow_cold_users=ALLOW_COLD_USERS, force_new_split=True, seed=get_split_seed()) data_reader.load_data() URM_train, URM_test = data_reader.get_holdout_split() ICM_all, _ = get_ICM_train_new(data_reader) UCM_all, _ = get_UCM_train_new(data_reader) # Ignoring users ignore_users = get_ignore_users( URM_train, data_reader.get_original_user_id_to_index_mapper(), lower_threshold=LOWER_THRESHOLD, upper_threshold=UPPER_THRESHOLD, ignore_non_target_users=IGNORE_NON_TARGET_USERS) evaluator = EvaluatorHoldout(URM_test, cutoff_list=[CUTOFF], ignore_users=ignore_users) # Model evaluation model = get_model(URM_train, ICM_all, UCM_all) print(evaluator.evaluateRecommender(model))
def get_ICM_with_fields(reader: New_DataSplitter_leave_k_out): """ It returns all the ICM_train_all after applying feature engineering :param reader: data splitter :return: return ICM_train_all and feature fields """ URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() ICM_all_dict = reader.get_loaded_ICM_dict() ICM_all_dict.pop("ICM_all") ICM_all_dict = apply_feature_engineering_ICM(ICM_all_dict, URM_train, UCM_all_dict, ICM_names_to_count=[], UCM_names_to_list=[]) ICM_all_dict = apply_filtering_ICM( ICM_all_dict, ICM_name_to_filter_mapper={ "ICM_asset": lambda x: x < np.quantile(x, q=0.75) + 0.72 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)), "ICM_price": lambda x: x < np.quantile(x, q=0.75) + 4 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)) }) ICM_all_dict = apply_transformation_ICM(ICM_all_dict, ICM_name_to_transform_mapper={ "ICM_asset": lambda x: np.log1p(1 / x), "ICM_price": lambda x: np.log1p(1 / x), "ICM_item_pop": np.log1p }) ICM_all_dict = apply_discretization_ICM(ICM_all_dict, ICM_name_to_bins_mapper={ "ICM_asset": 200, "ICM_price": 200, "ICM_item_pop": 50 }) ICM_all = None item_feature_fields = None for idx, ICM_key_value in enumerate(ICM_all_dict.items()): ICM_name, ICM_object = ICM_key_value if idx == 0: ICM_all = ICM_object item_feature_fields = np.full(shape=ICM_object.shape[1], fill_value=idx) else: ICM_all = sps.hstack([ICM_all, ICM_object], format="csr") item_feature_fields = np.concatenate([ item_feature_fields, np.full(shape=ICM_object.shape[1], fill_value=idx) ]) return ICM_all, item_feature_fields
def get_ICM_train(reader: New_DataSplitter_leave_k_out): """ It returns all the ICM_train_all after applying feature engineering. This preprocessing is used on new_best_models file :param reader: data splitter :return: return ICM_train_all """ URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() ICM_all_dict = reader.get_loaded_ICM_dict() ICM_all_dict.pop("ICM_all") ICM_all_dict = apply_feature_engineering_ICM( ICM_all_dict, URM_train, UCM_all_dict, ICM_names_to_count=["ICM_sub_class"], UCM_names_to_list=["UCM_age"]) ICM_all_dict = apply_filtering_ICM( ICM_all_dict, ICM_name_to_filter_mapper={ "ICM_asset": lambda x: x < np.quantile(x, q=0.75) + 0.72 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)), "ICM_price": lambda x: x < np.quantile(x, q=0.75) + 4 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)) }) ICM_all_dict = apply_transformation_ICM(ICM_all_dict, ICM_name_to_transform_mapper={ "ICM_asset": lambda x: np.log1p(1 / x), "ICM_price": lambda x: np.log1p(1 / x), "ICM_item_pop": np.log1p, "ICM_sub_class_count": np.log1p, "ICM_age": lambda x: x**(1 / 2.5) }) ICM_all_dict = apply_discretization_ICM(ICM_all_dict, ICM_name_to_bins_mapper={ "ICM_asset": 200, "ICM_price": 200, "ICM_item_pop": 50, "ICM_sub_class_count": 50 }) ICM_all = build_ICM_all_from_dict(ICM_all_dict) return ICM_all
def get_UCM_train_cold(reader: New_DataSplitter_leave_k_out): """ It returns all the UCM_all after applying feature engineering. This preprocessing is used on new_best_models file :param reader: data splitter :return: return UCM_all """ URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() UCM_all_dict = apply_transformation_UCM( UCM_all_dict, UCM_name_to_transform_mapper={"UCM_user_act": np.log1p}) UCM_all_dict = apply_discretization_UCM( UCM_all_dict, UCM_name_to_bins_mapper={"UCM_user_act": 50}) UCM_all = build_UCM_all_from_dict(UCM_all_dict) return UCM_all
def get_UCM_with_fields(reader: New_DataSplitter_leave_k_out): """ It returns all the UCM_all after applying feature engineering :param reader: data splitter :return: return UCM_all """ URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() ICM_dict = reader.get_loaded_ICM_dict() UCM_all_dict = apply_feature_engineering_UCM( UCM_all_dict, URM_train, ICM_dict, ICM_names_to_UCM=["ICM_sub_class"]) # These are useful feature weighting for UserCBF_CF_Warm UCM_all_dict = apply_transformation_UCM(UCM_all_dict, UCM_name_to_transform_mapper={ "UCM_sub_class": lambda x: x / 2, "UCM_user_act": np.log1p }) UCM_all_dict = apply_discretization_UCM( UCM_all_dict, UCM_name_to_bins_mapper={"UCM_user_act": 50}) UCM_all = None user_feature_fields = None for idx, UCM_key_value in enumerate(UCM_all_dict.items()): UCM_name, UCM_object = UCM_key_value if idx == 0: UCM_all = UCM_object user_feature_fields = np.full(shape=UCM_object.shape[1], fill_value=idx) else: UCM_all = sps.hstack([UCM_all, UCM_object], format="csr") user_feature_fields = np.concatenate([ user_feature_fields, np.full(shape=UCM_object.shape[1], fill_value=idx) ]) return UCM_all, user_feature_fields
def get_UCM_train(reader: New_DataSplitter_leave_k_out): """ It returns all the UCM_all after applying feature engineering. This preprocessing is used on new_best_models file :param reader: data splitter :return: return UCM_all """ URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() ICM_dict = reader.get_loaded_ICM_dict() UCM_all_dict = apply_feature_engineering_UCM( UCM_all_dict, URM_train, ICM_dict, ICM_names_to_UCM=["ICM_sub_class"]) # These are useful feature weighting for UserCBF_CF_Warm UCM_all_dict = apply_transformation_UCM(UCM_all_dict, UCM_name_to_transform_mapper={ "UCM_sub_class": lambda x: x / 2, "UCM_user_act": np.log1p }) UCM_all_dict = apply_discretization_UCM( UCM_all_dict, UCM_name_to_bins_mapper={"UCM_user_act": 50}) UCM_all = build_UCM_all_from_dict(UCM_all_dict) return UCM_all
from src.data_management.New_DataSplitter_leave_k_out import New_DataSplitter_leave_k_out from src.data_management.RecSys2019Reader import RecSys2019Reader from src.data_management.dataframe_preprocesser import get_preprocessed_dataframe from src.tuning.holdout_validation.run_parameter_search_advanced_top_pop import run_parameter_search_advanced_top_pop from src.utils.general_utility_functions import get_split_seed if __name__ == '__main__': # Data loading data_reader = RecSys2019Reader("../../data/") data_reader = New_DataSplitter_leave_k_out(data_reader, k_out_value=3, use_validation_set=False, force_new_split=True, seed=get_split_seed()) data_reader.load_data() URM_train, URM_test = data_reader.get_holdout_split() mapper = data_reader.get_original_user_id_to_index_mapper() df = get_preprocessed_dataframe("../../data/", keep_warm_only=True) # Setting evaluator # warm_users_mask = np.ediff1d(URM_train.tocsr().indptr) > 0 # warm_users = np.arange(URM_train.shape[0])[warm_users_mask] # ignore_users = warm_users cutoff_list = [10] evaluator = EvaluatorHoldout(URM_test, cutoff_list=cutoff_list) # HP tuning print("Start tuning...") version_path = "../../report/hp_tuning/advanced_top_pop/" now = datetime.now().strftime('%b%d_%H-%M-%S') now = now + "_k_out_value_3_eval/"
import numpy as np import xlearn as xl from src.data_management.New_DataSplitter_leave_k_out import New_DataSplitter_leave_k_out from src.data_management.RecSys2019Reader import RecSys2019Reader from src.data_management.data_preprocessing_fm import format_URM_positive_non_compressed, \ uniform_sampling_strategy, format_URM_positive_user_compressed, \ format_URM_negative_sampling_user_compressed, mix_URM, add_ICM_info from src.utils.general_utility_functions import get_split_seed, get_project_root_path if __name__ == '__main__': dataset = RecSys2019Reader("../../data/") dataset = New_DataSplitter_leave_k_out(dataset, k_out_value=3, use_validation_set=False, force_new_split=True, seed=get_split_seed()) dataset.load_data() URM_train, URM_test = dataset.get_holdout_split() ICM_all = dataset.get_ICM_from_name("ICM_all") URM_positive_FM_matrix = format_URM_positive_user_compressed(URM_train) URM_negative_FM_matrix = format_URM_negative_sampling_user_compressed(URM_train, negative_rate=1, sampling_function=uniform_sampling_strategy, check_replacement=True) URM_FM_matrix = mix_URM(URM_positive_FM_matrix, URM_negative_FM_matrix)[:, :-1] URM_FM_matrix = add_ICM_info(URM_FM_matrix, ICM_all, URM_train.shape[0]) root_path = get_project_root_path() fm_data_path = os.path.join(root_path, "resources", "fm_data") # Prepare train sparse matrix and labels for dumping to file FM_sps_matrix = URM_FM_matrix.copy() FM_sps_matrix[FM_sps_matrix == -1] = 1
def get_UCM_train_new(reader: New_DataSplitter_leave_k_out): URM_train, _ = reader.get_holdout_split() UCM_all_dict = reader.get_loaded_UCM_dict() ICM_dict = reader.get_loaded_ICM_dict() # Preprocess ICM ICM_dict.pop("ICM_all") ICM_dict = apply_feature_engineering_ICM( ICM_dict, URM_train, UCM_all_dict, ICM_names_to_count=["ICM_sub_class"], UCM_names_to_list=["UCM_age"]) ICM_dict = apply_filtering_ICM( ICM_dict, ICM_name_to_filter_mapper={ "ICM_asset": lambda x: x < np.quantile(x, q=0.75) + 0.72 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)), "ICM_price": lambda x: x < np.quantile(x, q=0.75) + 4 * (np.quantile(x, q=0.75) - np.quantile(x, q=0.25)) }) ICM_dict = apply_transformation_ICM(ICM_dict, ICM_name_to_transform_mapper={ "ICM_asset": lambda x: np.log1p(1 / x), "ICM_price": lambda x: np.log1p(1 / x), "ICM_item_pop": np.log1p, "ICM_sub_class_count": np.log1p, "ICM_age": lambda x: x**(1 / 2.5) }) ICM_dict = apply_discretization_ICM(ICM_dict, ICM_name_to_bins_mapper={ "ICM_asset": 200, "ICM_price": 200, "ICM_item_pop": 50, "ICM_sub_class_count": 50 }) # Preprocess UCM UCM_all_dict = apply_feature_engineering_UCM( UCM_all_dict, URM_train, ICM_dict, ICM_names_to_UCM=["ICM_sub_class", "ICM_item_pop"]) UCM_all_dict = apply_feature_entropy_UCM( UCM_all_dict, UCM_names_to_entropy=["UCM_sub_class"]) # Apply useful transformation UCM_all_dict = apply_transformation_UCM( UCM_all_dict, UCM_name_to_transform_mapper={"UCM_user_act": np.log1p}) UCM_all_dict = apply_discretization_UCM(UCM_all_dict, UCM_name_to_bins_mapper={ "UCM_user_act": 50, "UCM_sub_class_entropy": 20 }) UCM_all = None user_feature_to_range_mapper = {} last_range = 0 for idx, UCM_key_value in enumerate(UCM_all_dict.items()): UCM_name, UCM_object = UCM_key_value if idx == 0: UCM_all = UCM_object else: UCM_all = sps.hstack([UCM_all, UCM_object], format="csr") user_feature_to_range_mapper[UCM_name] = (last_range, last_range + UCM_object.shape[1]) last_range = last_range + UCM_object.shape[1] return UCM_all, user_feature_to_range_mapper