address_matcher / tools /recordlinkage_funcs.py
seanpedrickcase's picture
Allowed for custom output folder. Upgraded Gradio version
8c90944
raw
history blame
18.8 kB
import pandas as pd
from typing import Type, Dict, List, Tuple
import recordlinkage
from datetime import datetime
PandasDataFrame = Type[pd.DataFrame]
PandasSeries = Type[pd.Series]
MatchedResults = Dict[str,Tuple[str,int]]
array = List[str]
today = datetime.now().strftime("%d%m%Y")
today_rev = datetime.now().strftime("%Y%m%d")
from tools.constants import score_cut_off_nnet_street
# ## Recordlinkage matching functions
def compute_match(predict_df_search, ref_search, orig_search_df, matching_variables,
text_columns, blocker_column, weights, fuzzy_method):
# Use the merge command to match group1 and group2
predict_df_search[matching_variables] = predict_df_search[matching_variables].astype(str)
ref_search[matching_variables] = ref_search[matching_variables].astype(str).replace("-999","")
# SaoText needs to be exactly the same to get a 'full' match. So I moved that to the exact match group
exact_columns = list(set(matching_variables) - set(text_columns))
# Replace all blanks with a space, so they can be included in the fuzzy match searches
for column in text_columns:
predict_df_search.loc[predict_df_search[column] == '', column] = ' '
ref_search.loc[ref_search[column] == '', column] = ' '
# Score based match functions
# Create an index of all pairs
indexer = recordlinkage.Index()
# Block on selected blocker column
## Remove all NAs from predict_df blocker column
if blocker_column[0] == "PaoStartNumber":
predict_df_search = predict_df_search[~(predict_df_search[blocker_column[0]].isna()) & ~(predict_df_search[blocker_column[0]] == '')& ~(predict_df_search[blocker_column[0]].str.contains(r'^\s*$', na=False))]
indexer.block(blocker_column) #matchkey.block(["Postcode", "PaoStartNumber"])
# Generate candidate pairs
pairsSBM = indexer.index(predict_df_search,ref_search)
print('Running with ' + blocker_column[0] + ' as blocker has created', len(pairsSBM), 'pairs.')
# If no pairs are found, break
if len(pairsSBM) == 0: return pd.DataFrame()
# Call the compare class from the toolkit
compareSBM = recordlinkage.Compare()
# Assign variables to matching technique - exact
for columns in exact_columns:
compareSBM.exact(columns, columns, label = columns, missing_value = 0)
# Assign variables to matching technique - fuzzy
for columns in text_columns:
if columns == "Postcode":
compareSBM.string(columns, columns, label = columns, missing_value = 0, method = "levenshtein")
else:
compareSBM.string(columns, columns, label = columns, missing_value = 0, method = fuzzy_method)
## Run the match - compare each column within the blocks according to exact or fuzzy matching (defined in cells above)
scoresSBM = compareSBM.compute(pairs = pairsSBM, x = predict_df_search, x_link = ref_search)
return scoresSBM
def calc_final_nnet_scores(scoresSBM, weights, matching_variables):
#Modify the output scores by the weights set at the start of the code
scoresSBM_w = scoresSBM*weights
### Determine matched roles that score above a threshold
# Sum all columns
scoresSBM_r = scoresSBM_w
scoresSBM_r['score'] = scoresSBM_r[matching_variables].sum(axis = 1)
scoresSBM_r['score_max'] = sum(weights.values()) # + 2 for the additional scoring from the weighted variables a couple of cells above
scoresSBM_r['score_perc'] = (scoresSBM_r['score'] / scoresSBM_r['score_max'])*100
scoresSBM_r = scoresSBM_r.reset_index()
# Rename the index if misnamed
scoresSBM_r = scoresSBM_r.rename(columns={"index":"level_1"}, errors = "ignore")
# Sort all comparisons by score in descending order
scoresSBM_r = scoresSBM_r.sort_values(by=["level_0","score_perc"], ascending = False)
# Within each search address, remove anything below the max
scoresSBM_g = scoresSBM_r.reset_index()
# Get maximum score to join on
scoresSBM_g = scoresSBM_g.groupby("level_0").max("score_perc").reset_index()[["level_0", "score_perc"]]
scoresSBM_g =scoresSBM_g.rename(columns={"score_perc":"score_perc_max"})
scoresSBM_search = scoresSBM_r.merge(scoresSBM_g, on = "level_0", how="left")
scoresSBM_search['score_perc'] = round(scoresSBM_search['score_perc'],1).astype(float)
scoresSBM_search['score_perc_max'] = round(scoresSBM_search['score_perc_max'],1).astype(float)
return scoresSBM_search
def join_on_pred_ref_details(scoresSBM_search_m, ref_search, predict_df_search):
## Join back search and ref_df address details onto matching df
scoresSBM_search_m_j = scoresSBM_search_m.merge(ref_search, left_on="level_1", right_index=True, how = "left", suffixes=("", "_ref"))
scoresSBM_search_m_j = scoresSBM_search_m_j.merge(predict_df_search, left_on="level_0", right_index=True,how="left", suffixes=("", "_pred"))
scoresSBM_search_m_j = scoresSBM_search_m_j.reindex(sorted(scoresSBM_search_m_j.columns), axis=1)
return scoresSBM_search_m_j
def rearrange_columns(scoresSBM_search_m_j, new_join_col, search_df_key_field, blocker_column, standardise):
start_columns = new_join_col.copy()
start_columns.extend(["address", "fulladdress", "level_0", "level_1","score","score_max","score_perc","score_perc_max"])
other_columns = list(set(scoresSBM_search_m_j.columns) - set(start_columns))
all_columns_order = start_columns.copy()
all_columns_order.extend(sorted(other_columns))
# Place important columns at start
scoresSBM_search_m_j = scoresSBM_search_m_j.reindex(all_columns_order, axis=1)
scoresSBM_search_m_j = scoresSBM_search_m_j.rename(columns={'address':'address_pred',
'fulladdress':'address_ref',
'level_0':'index_pred',
'level_1':'index_ref',
'score':'match_score',
'score_max':'max_possible_score',
'score_perc':'perc_weighted_columns_matched',
'score_perc_max':'perc_weighted_columns_matched_max_for_pred_address'})
scoresSBM_search_m_j = scoresSBM_search_m_j.sort_values("index_pred", ascending = True)
# ref_index is just a duplicate of index_ref, needed for outputs
scoresSBM_search_m_j["ref_index"] = scoresSBM_search_m_j["index_ref"]
#search_df_j = orig_search_df[["full_address_search", search_df_key_field]]
#scoresSBM_out = scoresSBM_search_m_j.merge(search_df_j, left_on = "address_pred", right_on = "full_address_search", how = "left")
final_cols = new_join_col.copy()
final_cols.extend([search_df_key_field, 'full_match_score_based', 'address_pred', 'address_ref',\
'match_score', 'max_possible_score', 'perc_weighted_columns_matched',\
'perc_weighted_columns_matched_max_for_pred_address',\
'Organisation', 'Organisation_ref', 'Organisation_pred',\
'SaoText', 'SaoText_ref', 'SaoText_pred',\
'SaoStartNumber', 'SaoStartNumber_ref', 'SaoStartNumber_pred',\
'SaoStartSuffix', 'SaoStartSuffix_ref', 'SaoStartSuffix_pred',\
'SaoEndNumber', 'SaoEndNumber_ref', 'SaoEndNumber_pred',\
'SaoEndSuffix', 'SaoEndSuffix_ref', 'SaoEndSuffix_pred',\
'PaoStartNumber', 'PaoStartNumber_ref', 'PaoStartNumber_pred',\
'PaoStartSuffix', 'PaoStartSuffix_ref', 'PaoStartSuffix_pred',\
'PaoEndNumber', 'PaoEndNumber_ref', 'PaoEndNumber_pred',\
'PaoEndSuffix', 'PaoEndSuffix_ref', 'PaoEndSuffix_pred',\
'PaoText', 'PaoText_ref', 'PaoText_pred',\
'Street', 'Street_ref', 'Street_pred',\
'PostTown', 'PostTown_ref', 'PostTown_pred',\
'Postcode', 'Postcode_ref', 'Postcode_pred', 'Postcode_predict',\
'index_pred', 'index_ref', 'Reference file'
])
scoresSBM_out = scoresSBM_search_m_j[final_cols]
return scoresSBM_out, start_columns
def create_matched_results_nnet(scoresSBM_best, search_df_key_field, orig_search_df, new_join_col, standardise, ref_search, blocker_column, score_cut_off):
### Make the final 'matched output' file
scoresSBM_best_pred_cols = scoresSBM_best.filter(regex='_pred$').iloc[:,1:-1]
scoresSBM_best["search_orig_address"] = (scoresSBM_best_pred_cols.agg(' '.join, axis=1)).str.strip().str.replace("\s{2,}", " ", regex=True)
scoresSBM_best_ref_cols = scoresSBM_best.filter(regex='_ref$').iloc[:,1:-1]
scoresSBM_best['reference_mod_address'] = (scoresSBM_best_ref_cols.agg(' '.join, axis=1)).str.strip().str.replace("\s{2,}", " ", regex=True)
## Create matched output df
matched_output_SBM = orig_search_df[[search_df_key_field, "full_address", "postcode", "property_number", "prop_number", "flat_number", "apart_number", "block_number", 'unit_number', "room_number", "house_court_name"]].replace(r"\bnan\b", "", regex=True).infer_objects(copy=False)
matched_output_SBM[search_df_key_field] = matched_output_SBM[search_df_key_field].astype(str)
###
matched_output_SBM = matched_output_SBM.merge(scoresSBM_best[[search_df_key_field, 'index_ref','address_ref',
'full_match_score_based', 'Reference file']], on = search_df_key_field, how = "left").\
rename(columns={"full_address":"search_orig_address"})
if 'index' not in ref_search.columns:
ref_search['ref_index'] = ref_search.index
matched_output_SBM = matched_output_SBM.merge(ref_search.drop_duplicates("fulladdress")[["ref_index", "fulladdress", "Postcode", "property_number", "prop_number", "flat_number", "apart_number", "block_number", 'unit_number', "room_number", "house_court_name", "ref_address_stand"]], left_on = "address_ref", right_on = "fulladdress", how = "left", suffixes=('_search', '_reference')).rename(columns={"fulladdress":"reference_orig_address", "ref_address_stand":"reference_list_address"})
# To replace with number check
matched_output_SBM = matched_output_SBM.rename(columns={"full_match_score_based":"full_match"})
matched_output_SBM['property_number_match'] = matched_output_SBM['full_match']
scores_SBM_best_cols = [search_df_key_field, 'full_match_score_based', 'perc_weighted_columns_matched', 'address_pred']#, "reference_mod_address"]
scores_SBM_best_cols.extend(new_join_col)
matched_output_SBM_b = scoresSBM_best[scores_SBM_best_cols]
matched_output_SBM = matched_output_SBM.merge(matched_output_SBM_b.drop_duplicates(search_df_key_field), on = search_df_key_field, how = "left")
from tools.fuzzy_match import create_diag_shortlist
matched_output_SBM = create_diag_shortlist(matched_output_SBM, "search_orig_address", score_cut_off, blocker_column, fuzzy_col='perc_weighted_columns_matched', search_mod_address="address_pred", resolve_tie_breaks=False)
matched_output_SBM['standardised_address'] = standardise
matched_output_SBM = matched_output_SBM.rename(columns={"address_pred":"search_mod_address",
'perc_weighted_columns_matched':"fuzzy_score"})
matched_output_SBM_cols = [search_df_key_field, 'search_orig_address','reference_orig_address',
'full_match',
'full_number_match',
'flat_number_match',
'room_number_match',
'block_number_match',
'property_number_match',
'close_postcode_match',
'house_court_name_match',
'fuzzy_score_match',
"fuzzy_score",
'property_number_search', 'property_number_reference',
'flat_number_search', 'flat_number_reference',
'room_number_search', 'room_number_reference',
'block_number_search', 'block_number_reference',
"unit_number_search","unit_number_reference",
'house_court_name_search', 'house_court_name_reference',
"search_mod_address", 'reference_mod_address','Postcode', 'postcode', 'ref_index', 'Reference file']
matched_output_SBM_cols.extend(new_join_col)
matched_output_SBM_cols.extend(['standardised_address'])
matched_output_SBM = matched_output_SBM[matched_output_SBM_cols]
matched_output_SBM = matched_output_SBM.sort_values(search_df_key_field, ascending=True)
return matched_output_SBM
def score_based_match(predict_df_search, ref_search, orig_search_df, matching_variables, text_columns, blocker_column, weights, fuzzy_method, score_cut_off, search_df_key_field, standardise, new_join_col, score_cut_off_nnet_street=score_cut_off_nnet_street):
scoresSBM = compute_match(predict_df_search, ref_search, orig_search_df, matching_variables, text_columns, blocker_column, weights, fuzzy_method)
if scoresSBM.empty:
# If no pairs are found, break
return pd.DataFrame(), pd.DataFrame()
scoresSBM_search = calc_final_nnet_scores(scoresSBM, weights, matching_variables)
# Filter potential matched address scores to those with highest scores only
scoresSBM_search_m = scoresSBM_search[scoresSBM_search["score_perc"] == scoresSBM_search["score_perc_max"]]
scoresSBM_search_m_j = join_on_pred_ref_details(scoresSBM_search_m, ref_search, predict_df_search)
# When blocking by street, may to have an increased threshold as this is more prone to making mistakes
if blocker_column[0] == "Street": scoresSBM_search_m_j['full_match_score_based'] = (scoresSBM_search_m_j['score_perc'] >= score_cut_off_nnet_street)
else: scoresSBM_search_m_j['full_match_score_based'] = (scoresSBM_search_m_j['score_perc'] >= score_cut_off)
### Reorder some columns
scoresSBM_out, start_columns = rearrange_columns(scoresSBM_search_m_j, new_join_col, search_df_key_field, blocker_column, standardise)
matched_output_SBM = create_matched_results_nnet(scoresSBM_out, search_df_key_field, orig_search_df, new_join_col, standardise, ref_search, blocker_column, score_cut_off)
matched_output_SBM_best = matched_output_SBM.sort_values([search_df_key_field, "full_match"], ascending = [True, False]).drop_duplicates(search_df_key_field)
scoresSBM_best = scoresSBM_out[scoresSBM_out[search_df_key_field].isin(matched_output_SBM_best[search_df_key_field])]
return scoresSBM_best, matched_output_SBM_best
def check_matches_against_fuzzy(match_results, scoresSBM, search_df_key_field):
if not match_results.empty:
if 'fuzz_full_match' not in match_results.columns:
match_results['fuzz_full_match'] = False
match_results = match_results.add_prefix("fuzz_").rename(columns={"fuzz_"+search_df_key_field:search_df_key_field})
#Merge fuzzy match full matches onto model data
scoresSBM_m = scoresSBM.merge(match_results.drop_duplicates(search_df_key_field), on = search_df_key_field, how = "left")
else:
scoresSBM_m = scoresSBM
scoresSBM_m["fuzz_full_match"] = False
scoresSBM_m['fuzz_fuzzy_score_match'] = False
scoresSBM_m['fuzz_property_number_match'] = False
scoresSBM_m['fuzz_fuzzy_score'] = 0
scoresSBM_m['fuzz_reference_orig_address'] = ""
scoresSBM_t = scoresSBM[scoresSBM["full_match_score_based"]==True]
### Create a df of matches the model finds that the fuzzy matching work did not
scoresSBM_m_model_add_matches = scoresSBM_m[(scoresSBM_m["full_match_score_based"] == True) &\
(scoresSBM_m["fuzz_full_match"] == False)]
# Drop some irrelevant columns
first_cols = ['UPRN', search_df_key_field, 'full_match_score_based', 'fuzz_full_match', 'fuzz_fuzzy_score_match', 'fuzz_property_number_match',\
'fuzz_fuzzy_score', 'match_score', 'max_possible_score', 'perc_weighted_columns_matched',\
'perc_weighted_columns_matched_max_for_pred_address', 'address_pred',\
'address_ref', 'fuzz_reference_orig_address']
last_cols = [col for col in scoresSBM_m_model_add_matches.columns if col not in first_cols]
scoresSBM_m_model_add_matches = scoresSBM_m_model_add_matches[first_cols+last_cols].drop(['fuzz_search_mod_address',
'fuzz_reference_mod_address', 'fuzz_fulladdress', 'fuzz_UPRN'], axis=1, errors="ignore")
### Create a df for matches the fuzzy matching found that the neural net model does not
if not match_results.empty:
scoresSBM_t_model_failed = match_results[(~match_results[search_df_key_field].isin(scoresSBM_t[search_df_key_field])) &\
(match_results["fuzz_full_match"] == True)]
scoresSBM_t_model_failed = scoresSBM_t_model_failed.\
merge(scoresSBM.drop_duplicates(search_df_key_field), on = search_df_key_field, how = "left")
scoresSBM_t_model_failed = scoresSBM_t_model_failed[first_cols+last_cols].drop(['fuzz_search_mod_address',
'fuzz_reference_mod_address', 'fuzz_fulladdress', 'fuzz_UPRN'], axis=1, errors="ignore")
else:
scoresSBM_t_model_failed = pd.DataFrame()
## Join back onto original results file and export
scoresSBM_new_matches_from_model = scoresSBM_m_model_add_matches.drop_duplicates(search_df_key_field)
if not match_results.empty:
match_results_out = match_results.merge(scoresSBM_new_matches_from_model[[search_df_key_field, 'full_match_score_based', 'address_pred',
'address_ref']], on = search_df_key_field, how = "left")
match_results_out.loc[match_results_out['full_match_score_based'].isna(),'full_match_score_based'] = False
#match_results_out['full_match_score_based'].value_counts()
match_results_out["full_match_fuzzy_or_score_based"] = (match_results_out["fuzz_full_match"] == True) |\
(match_results_out["full_match_score_based"] == True)
else: match_results_out = match_results
return scoresSBM_m_model_add_matches, scoresSBM_t_model_failed, match_results_out