matching_dependency/ml_er/magellan_new.py

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)