import json import os import pickle import time import ConfigSpace import pandas as pd import py_entitymatching as em import torch from ConfigSpace import Configuration from ConfigSpace.read_and_write import json as csj import py_entitymatching.catalog.catalog_manager as cm from tqdm import tqdm from md_discovery.md_mining import mining from settings import * def blocking_mining(): start = time.time() ltable = pd.read_csv(ltable_path, encoding='ISO-8859-1') cm.set_key(ltable, ltable_id) rtable = pd.read_csv(rtable_path, encoding='ISO-8859-1') cm.set_key(rtable, rtable_id) mappings = pd.read_csv(mapping_path, encoding='ISO-8859-1') matching_number = len(mappings) # if ltable_id == rtable_id: # tables_id = rtable_id attributes = ltable.columns.values.tolist() # lattributes = ['ltable_' + i for i in attributes] # rattributes = ['rtable_' + i for i in attributes] cm.set_key(ltable, ltable_id) cm.set_key(rtable, rtable_id) blocker = em.OverlapBlocker() candidate = blocker.block_tables(ltable, rtable, ltable_block_attr, rtable_block_attr, allow_missing=True, l_output_attrs=attributes, r_output_attrs=attributes, n_jobs=1, overlap_size=1, show_progress=False) candidate['gold'] = 0 candidate = candidate.reset_index(drop=True) block_time = time.time() print(f'Block Time: {block_time - start}') # 根据mapping表标注数据 candidate_match_rows = [] for t in tqdm(mappings.itertuples()): mask = ((candidate['ltable_' + ltable_id].isin([getattr(t, mapping_lid)])) & (candidate['rtable_' + rtable_id].isin([getattr(t, mapping_rid)]))) matching_indices = candidate[mask].index candidate_match_rows.extend(matching_indices.tolist()) match_rows_mask = candidate.index.isin(candidate_match_rows) candidate.loc[match_rows_mask, 'gold'] = 1 candidate.fillna(value="", inplace=True) # negative样本太多, 采样三倍于positive样本量 candidate_mismatch = candidate[candidate['gold'] == 0] candidate_match = candidate[candidate['gold'] == 1] candidate_mismatch = candidate_mismatch.sample(n=3*len(candidate_match)) candidate_for_train_test = pd.concat([candidate_mismatch, candidate_match]) # 如果拼接后不重设索引可能导致索引重复 candidate_for_train_test = candidate_for_train_test.reset_index(drop=True) cm.set_key(candidate_for_train_test, '_id') cm.set_fk_ltable(candidate_for_train_test, 'ltable_' + ltable_id) cm.set_fk_rtable(candidate_for_train_test, 'rtable_' + rtable_id) cm.set_ltable(candidate_for_train_test, ltable) cm.set_rtable(candidate_for_train_test, rtable) block_recall = len(candidate_match) / matching_number # 分为训练测试集 train_proportion = 0.5 sets = em.split_train_test(candidate_for_train_test, train_proportion=train_proportion, random_state=0) train_set = sets['train'] test_set = sets['test'] label_and_split_time = time.time() print(f'Label and Split Time: {label_and_split_time - block_time}') # 挖掘MD并保存本地 md_list = mining(train_set) mining_time = time.time() print(f'Mining Time: {mining_time - label_and_split_time}') blocking_results = (ltable, rtable, train_set, test_set, md_list, block_recall) # 将blocking结果保存到本地 with open(er_output_dir + "blocking_result.pickle", "wb") as file_: pickle.dump(blocking_results, file_) return blocking_results def matching(config: Configuration, blocking_result_): print(f'\033[33mConfig: {config}\033[0m') start = time.time() ltable = blocking_result_[0] rtable = blocking_result_[1] train_set = blocking_result_[2] test_set = blocking_result_[3] md_list = blocking_result_[4] block_recall = blocking_result_[5] ml_matcher = config["ml_matcher"] match ml_matcher: case "dt": matcher = em.DTMatcher(name='DecisionTree', random_state=0, criterion=config['tree_criterion'], max_depth=config['tree_max_depth'], splitter=config['dt_splitter'], max_features=config['dt_max_features']) case "svm": matcher = em.SVMMatcher(name='SVM', random_state=0, kernel=config['svm_kernel'], degree=config['svm_degree'], gamma=config['svm_gamma'], C=config['svm_C'], coef0=config['svm_constant']) case "rf": matcher = em.RFMatcher(name='RandomForest', random_state=0, criterion=config['tree_criterion'], max_depth=config['tree_max_depth'], n_estimators=config['number_of_tree'], max_features=config['rf_max_features']) cm.set_key(train_set, '_id') cm.set_fk_ltable(train_set, 'ltable_' + ltable_id) cm.set_fk_rtable(train_set, 'rtable_' + rtable_id) cm.set_ltable(train_set, ltable) cm.set_rtable(train_set, rtable) cm.set_key(ltable, ltable_id) cm.set_key(rtable, rtable_id) cm.set_key(test_set, '_id') cm.set_fk_ltable(test_set, 'ltable_' + ltable_id) cm.set_fk_rtable(test_set, 'rtable_' + rtable_id) cm.set_ltable(test_set, ltable) cm.set_rtable(test_set, rtable) feature_table = em.get_features_for_matching(ltable, rtable, validate_inferred_attr_types=False) train_feature_vecs = em.extract_feature_vecs(train_set, feature_table=feature_table, attrs_after=['gold'], show_progress=False) train_feature_vecs.fillna(value=0, inplace=True) test_feature_after = ['ltable_' + i for i in ltable.columns.values.tolist()] for _ in test_feature_after[:]: test_feature_after.append(_.replace('ltable_', 'rtable_')) for _ in test_feature_after: if _.endswith(ltable_id) or _.endswith(rtable_id): test_feature_after.remove(_) test_feature_after.append('gold') test_feature_vecs = em.extract_feature_vecs(test_set, feature_table=feature_table, attrs_after=test_feature_after, show_progress=False) test_feature_vecs.fillna(value=0, inplace=True) fit_exclude = ['_id', 'ltable_' + ltable_id, 'rtable_' + rtable_id, 'gold'] matcher.fit(table=train_feature_vecs, exclude_attrs=fit_exclude, target_attr='gold') test_feature_after.extend(['_id', 'ltable_' + ltable_id, 'rtable_' + rtable_id]) predictions = matcher.predict(table=test_feature_vecs, exclude_attrs=test_feature_after, append=True, target_attr='predicted', inplace=False) eval_result = em.eval_matches(predictions, 'gold', 'predicted') em.print_eval_summary(eval_result) indicators = evaluate_prediction(predictions, 'gold', 'predicted') indicators['block_recall'] = block_recall test_feature_after.remove('_id') test_feature_after.append('predicted') predictions = predictions[test_feature_after] predictions = predictions.reset_index(drop=True) predictions = predictions.astype(str) sim_tensor_dict = build_col_pairs_sim_tensor_dict(predictions) predictions['confidence'] = 0 epl_match = 0 # 可解释，预测match if len(md_list) > 0: for row in tqdm(predictions.itertuples()): x = is_explicable(row, md_list, sim_tensor_dict) if x > 0 and str(getattr(row, 'predicted')) == str(1): predictions.loc[row[0], 'confidence'] = x epl_match += 1 df = predictions[predictions['predicted'] == str(1)] interpretability = epl_match / len(df) # 可解释性 indicators['interpretability'] = interpretability # note 既然不调block参数, 不妨假设block_recall很高, 不必考虑 # if indicators["block_recall"] < indicators["recall"]: # f1 = (2.0 * indicators["precision"] * indicators["block_recall"]) / ( # indicators["precision"] + indicators["block_recall"]) # else: # f1 = indicators["F1"] performance = interpre_weight * interpretability + (1 - interpre_weight) * indicators["F1"] indicators['performance'] = performance print(f'ER Indicators: {indicators}') predictions.to_csv(er_output_dir + 'predictions.csv', sep=',', index=False, header=True) print(f'\033[33mTime consumed by matching in seconds: {time.time() - start}\033[0m') return indicators def evaluate_prediction(prediction_: pd.DataFrame, labeled_attr: str, predicted_attr: str) -> dict: new_df = prediction_.reset_index(drop=False, inplace=False) gold = new_df[labeled_attr] predicted = new_df[predicted_attr] gold_negative = gold[gold == 0].index.values gold_positive = gold[gold == 1].index.values predicted_negative = predicted[predicted == 0].index.values predicted_positive = predicted[predicted == 1].index.values false_positive_indices = list(set(gold_negative).intersection(predicted_positive)) true_positive_indices = list(set(gold_positive).intersection(predicted_positive)) false_negative_indices = list(set(gold_positive).intersection(predicted_negative)) num_true_positives = float(len(true_positive_indices)) num_false_positives = float(len(false_positive_indices)) num_false_negatives = float(len(false_negative_indices)) precision_denominator = num_true_positives + num_false_positives recall_denominator = num_true_positives + num_false_negatives precision = 0.0 if precision_denominator == 0.0 else num_true_positives / precision_denominator recall = 0.0 if recall_denominator == 0.0 else num_true_positives / recall_denominator F1 = 0.0 if precision == 0.0 and recall == 0.0 else (2.0 * precision * recall) / (precision + recall) return {"precision": precision, "recall": recall, "F1": F1} def build_col_pairs_sim_tensor_dict(predictions: pd.DataFrame): predictions_attrs = predictions.columns.values.tolist() col_tuple_list = [] for _ in predictions_attrs: if _.startswith('ltable'): left_index = predictions_attrs.index(_) right_index = predictions_attrs.index(_.replace('ltable_', 'rtable_')) col_tuple_list.append((left_index, right_index)) length = predictions.shape[0] width = predictions.shape[1] predictions = predictions.reset_index(drop=True) sentences = predictions.values.flatten(order='F').tolist() embedding = model.encode(sentences, convert_to_tensor=True, device="cuda", batch_size=256, show_progress_bar=True) split_embedding = torch.split(embedding, length, dim=0) table_tensor = torch.stack(split_embedding, dim=0, out=None) # prediction的归一化嵌入张量 norm_table_tensor = torch.nn.functional.normalize(table_tensor, dim=2) sim_tensor_dict = {} for col_tuple in col_tuple_list: lattr_tensor = norm_table_tensor[col_tuple[0]] rattr_tensor = norm_table_tensor[col_tuple[1]] mul_tensor = lattr_tensor * rattr_tensor sim_tensor = torch.sum(mul_tensor, 1) sim_tensor = torch.round(sim_tensor, decimals=4) sim_tensor_dict[predictions_attrs[col_tuple[0]].replace('ltable_', '')] = sim_tensor return sim_tensor_dict def is_explicable(row, all_mds: list, st_dict): attrs = all_mds[0][0].keys() # 从第一条md_tuple中的md字典中读取所有字段 for md_tuple in all_mds: explicable = True # 假设这条md能解释当前元组 for a in attrs: if st_dict[a][row[0]].item() < md_tuple[0][a]: explicable = False # 任意一个字段的相似度达不到阈值，这条md就不能解释当前元组 break # 不再与当前md的其他相似度阈值比较，跳转到下一条md if explicable: return md_tuple[2] # 任意一条md能解释，直接返回 return -1.0 # 遍历结束，不能解释 def ml_er(config: Configuration, blocking_result_): indicators = matching(config, blocking_result_) output_path = er_output_dir + "eval_result.txt" with open(output_path, 'w') as _f: _f.write('Precision:' + str(indicators["precision"]) + '\n') _f.write('Recall:' + str(indicators["recall"]) + '\n') _f.write('F1:' + str(indicators["F1"]) + '\n') _f.write('block_recall:' + str(indicators["block_recall"]) + '\n') _f.write('interpretability:' + str(indicators['interpretability']) + '\n') _f.write('performance:' + str(indicators['performance']) + '\n') if __name__ == '__main__': if os.path.isfile(hpo_output_dir + "incumbent.json"): with open(hpo_output_dir + "configspace.json", 'r') as f: dict_configspace = json.load(f) str_configspace = json.dumps(dict_configspace) configspace = csj.read(str_configspace) with open(hpo_output_dir + "incumbent.json", 'r') as f: dic = json.load(f) configuration = ConfigSpace.Configuration(configspace, values=dic) with open(er_output_dir + "blocking_result.pickle", "rb") as file: blocking_result = pickle.load(file) ml_er(configuration, blocking_result)