# from loguru import logger # import pandas as pd # import json # from datetime import datetime # import ast # import numpy as np # from pymongo import MongoClient # from collections import defaultdict # from tqdm import tqdm # import time # import requests # import json # import os # import pandas as pd # import nltk # from nltk.tokenize import sent_tokenize, word_tokenize # from nltk.corpus import stopwords # from textblob import TextBlob # import re # from transformers import BertTokenizer, BertModel # from transformers import RobertaTokenizer, RobertaModel # import torch # from sklearn.metrics.pairwise import cosine_similarity # import numpy as np # # Download NLTK resources # nltk.download('punkt') # nltk.download('averaged_perceptron_tagger') # nltk.download('stopwords') # nltk.download('punkt_tab') # nltk.download('averaged_perceptron_tagger_eng') # class Preprocessor: # def __init__(self,df): # self.df=df # self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base') # self.model = RobertaModel.from_pretrained('roberta-base') # self.stop_words = set(stopwords.words('english')) # self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Add this line # def get_bert_embedding(self, text): # inputs = self.tokenizer(text, return_tensors='pt', truncation=True, padding=True, max_length=512) # with torch.no_grad(): # outputs = self.model(**inputs) # return outputs.last_hidden_state.mean(dim=1).squeeze().numpy() # def preprocess_text(self,text): # return text if pd.notna(text) else "" # def calculate_duration(self, time_range): # if not isinstance(time_range, str) or "-" not in time_range: # return None # start_str, end_str = time_range.split('-') # start_str = start_str.strip() + ':00' if len(start_str.split(':')) == 1 else start_str.strip() # end_str = end_str.strip() + ':00' if len(end_str.split(':')) == 1 else end_str.strip() # try: # start = datetime.strptime(start_str, '%H:%M') # end = datetime.strptime(end_str, '%H:%M') # duration = (end - start).total_seconds() / 3600 # return duration if duration >= 0 else duration + 24 # except ValueError: # return None # def calculate_sentiment_severity(self, text): # if pd.isna(text) or not text.strip(): # return pd.Series({"good_severity": 0.0, "bad_severity": 0.0}) # # Get sentiment polarity (-1 to 1) # blob = TextBlob(text) # polarity = blob.sentiment.polarity # # Define severity weights # good_weight = 0.7 # bad_weight = 0.3 # if polarity > 0: # good_severity = good_weight * polarity # bad_severity = 0.0 # elif polarity < 0: # good_severity = 0.0 # bad_severity = bad_weight * abs(polarity) # else: # Neutral (polarity = 0) # good_severity = 0.0 # bad_severity = 0.0 # return pd.Series({"good_severity": good_severity, "bad_severity": bad_severity}) # def get_avg_duration(self, hours_str): # if pd.isna(hours_str) or not isinstance(hours_str, str): # return pd.NA # try: # hours_dict = ast.literal_eval(hours_str) # if not hours_dict: # return pd.NA # durations = [self.calculate_duration(time_range) for time_range in hours_dict.values()] # valid_durations = [d for d in durations if d is not None] # return sum(valid_durations) / len(valid_durations) if valid_durations else pd.NA # except (ValueError, SyntaxError, ZeroDivisionError): # return pd.NA # def calculate_time_since_last_review(self): # present_date = datetime.now() # user_latest_timestamp = {} # # Convert review_date to datetime # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Calculate hours difference for each user's latest review # for user_id in self.df["user_id"].unique(): # latest_date = self.df[self.df["user_id"] == user_id]["review_date"].max() # if not isinstance(latest_date, datetime): # latest_date = latest_date.to_pydatetime() # hours_difference = (present_date - latest_date).total_seconds() / 3600 # user_latest_timestamp[user_id] = hours_difference # # Map the hours difference to a new column # self.df["time_since_last_review_user"] = self.df["user_id"].map(user_latest_timestamp) # def calculate_time_since_last_review_business(self): # present_date = datetime.now() # # Ensure review_date is in datetime format # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Initialize dictionary to store hours since last review for each business # business_latest_timestamp = {} # # Iterate over unique business_ids # for business_id in self.df["business_id"].unique(): # # Get the latest review date for this business # latest_date = self.df[self.df["business_id"] == business_id]["review_date"].max() # # Convert to datetime object if needed # if not isinstance(latest_date, datetime): # latest_date = latest_date.to_pydatetime() # # Calculate hours difference (already in hours) # hours_difference = (present_date - latest_date).total_seconds() / 3600 # business_latest_timestamp[business_id] = hours_difference # # Map the hours difference to the new column # self.df["time_since_last_review_business"] = self.df["business_id"].map(business_latest_timestamp) # def calculate_user_account_age(self): # present_date = datetime.now() # # Convert yelping_since to datetime # self.df["yelping_since"] = pd.to_datetime(self.df["yelping_since"]) # # Calculate user account age in days # self.df["user_account_age"] = (present_date - self.df["yelping_since"]).dt.days # def calculate_avg_time_between_reviews(self): # # Ensure review_date is in datetime format # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Sort the DataFrame by user_id and review_date to ensure chronological order # self.df = self.df.sort_values(["user_id", "review_date"]) # # Define helper function to calculate average time between reviews # def calculate_avg_time(group): # if len(group) == 1: # return 0 # If only one review, assign 0 # # Calculate differences in hours between consecutive reviews # diffs = group["review_date"].diff().dt.total_seconds() / 3600 # # Drop the first NaN (from diff) and compute the mean # return diffs.dropna().mean() # # Apply the function to each user_id group and create a mapping # avg_time_per_user = self.df.groupby("user_id").apply(calculate_avg_time) # # Map the average time back to the original DataFrame # self.df["average_time_between_reviews"] = self.df["user_id"].map(avg_time_per_user) # def calculate_user_degree(self): # # Calculate the number of unique businesses per user # user_business_counts = self.df.groupby("user_id")["business_id"].nunique() # # Map the counts back to the original DataFrame # self.df["user_degree"] = self.df["user_id"].map(user_business_counts) # def calculate_business_degree(self): # # Calculate the number of unique users per business # business_user_counts = self.df.groupby("business_id")["user_id"].nunique() # # Map the counts back to the original DataFrame # self.df["business_degree"] = self.df["business_id"].map(business_user_counts) # def calculate_rating_variance_user(self): # # Calculate the mode (most frequent rating) per user # user_rating_mode = self.df.groupby("user_id")["review_stars"].agg(lambda x: x.mode()[0]) # # Map the most frequent rating back to the original DataFrame # self.df["rating_variance_user"] = self.df["user_id"].map(user_rating_mode) # def calculate_user_review_burst_count(self): # # Ensure review_date is in datetime format # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Sort by user_id and review_date for chronological order # self.df = self.df.sort_values(["user_id", "review_date"]) # # Function to calculate the max number of reviews in any 20-day window # def calculate_burst_count(group): # if len(group) <= 1: # return 0 # No burst if 1 or fewer reviews # # Convert review_date to a Series for rolling window # dates = group["review_date"] # # Calculate the number of reviews within 20 days of each review # burst_counts = [] # for i, date in enumerate(dates): # # Count reviews within 20 days after this date # window_end = date + pd.Timedelta(days=20) # count = ((dates >= date) & (dates <= window_end)).sum() # burst_counts.append(count) # # Return the maximum burst count for this user # return max(burst_counts) # # Calculate the burst count per user # user_burst_counts = self.df.groupby("user_id").apply(calculate_burst_count) # # Map the burst count back to the original DataFrame # self.df["user_review_burst_count"] = self.df["user_id"].map(user_burst_counts) # def calculate_business_review_burst_count(self): # # Ensure review_date is in datetime format # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Sort by business_id and review_date for chronological order # self.df = self.df.sort_values(["business_id", "review_date"]) # # Function to calculate the max number of reviews in any 10-day window # def calculate_burst_count(group): # if len(group) <= 1: # return 0 # No burst if 1 or fewer reviews # # Convert review_date to a Series for rolling window # dates = group["review_date"] # # Calculate the number of reviews within 10 days of each review # burst_counts = [] # for i, date in enumerate(dates): # # Count reviews within 10 days after this date # window_end = date + pd.Timedelta(days=10) # count = ((dates >= date) & (dates <= window_end)).sum() # burst_counts.append(count) # # Return the maximum burst count for this business # return max(burst_counts) # # Calculate the burst count per business # business_burst_counts = self.df.groupby("business_id").apply(calculate_burst_count) # # Map the burst count back to the original DataFrame # self.df["business_review_burst_count"] = self.df["business_id"].map(business_burst_counts) # def calculate_temporal_similarity(self): # self.df["review_date"] = pd.to_datetime(self.df["review_date"]) # # Extract the day of the week (0 = Monday, 6 = Sunday) # self.df["day_of_week"] = self.df["review_date"].dt.dayofweek # # Function to calculate avg hours between reviews on frequent days # def calculate_avg_hours_on_frequent_days(group): # frequent_days = group["day_of_week"].mode().tolist() # if len(group) <= 1: # return 0 # frequent_reviews = group[group["day_of_week"].isin(frequent_days)] # if len(frequent_reviews) <= 1: # return 0 # frequent_reviews = frequent_reviews.sort_values("review_date") # diffs = frequent_reviews["review_date"].diff().dt.total_seconds() / 3600 # return diffs.dropna().mean() # # Calculate average hours for each user # avg_hours_per_user = self.df.groupby("user_id").apply(calculate_avg_hours_on_frequent_days) # # Map the average hours to the new column # self.df["temporal_similarity"] = self.df["user_id"].map(avg_hours_per_user) # # Drop temporary column # self.df = self.df.drop(columns=["day_of_week"]) # def calculate_rating_deviation_from_business_average(self): # # Calculate the average rating per business # business_avg_rating = self.df.groupby("business_id")["review_stars"].mean() # # Map the average rating to each row # self.df["business_avg_rating"] = self.df["business_id"].map(business_avg_rating) # # Calculate the deviation from the business average # self.df["rating_deviation_from_business_average"] = ( # self.df["review_stars"] - self.df["business_avg_rating"] # ) # # Drop the temporary column # self.df = self.df.drop(columns=["business_avg_rating"]) # def calculate_review_like_ratio(self): # # Create a binary column for liked reviews (stars >= 4) # self.df["is_liked"] = (self.df["review_stars"] >= 4).astype(int) # # Calculate the like ratio per user # user_like_ratio = self.df.groupby("user_id")["is_liked"].mean() # # Map the like ratio back to the DataFrame # self.df["review_like_ratio"] = self.df["user_id"].map(user_like_ratio) # # Drop the temporary column # self.df = self.df.drop(columns=["is_liked"]) # def calculate_latest_checkin_hours(self): # self.df["yelping_since"] = pd.to_datetime(self.df["yelping_since"]) # # Function to get the latest check-in date from a list of strings # def get_latest_checkin(checkin_list): # if not checkin_list or pd.isna(checkin_list): # Handle empty or NaN # return None # if isinstance(checkin_list, str): # checkin_dates = checkin_list.split(", ") # else: # checkin_dates = checkin_list # return pd.to_datetime(checkin_dates).max() # # Apply the function to get the latest check-in date per row # self.df["latest_checkin_date"] = self.df["checkin_date"].apply(get_latest_checkin) # # Calculate the hours difference between latest check-in and yelping_since # self.df["latest_checkin_hours"] = ( # (self.df["latest_checkin_date"] - self.df["yelping_since"]) # .dt.total_seconds() / 3600 # ) # # Drop the temporary column # self.df = self.df.drop(columns=["latest_checkin_date"]) # self.df["latest_checkin_hours"].fillna(0,inplace=True) # def compute_pronoun_density(self, text): # text = self.preprocess_text(text) # if not text: # return 0 # words = word_tokenize(text.lower()) # pos_tags = nltk.pos_tag(words) # pronouns = sum(1 for word, pos in pos_tags if pos in ['PRP', 'PRP$'] and word in ['i', 'we']) # return pronouns / len(words) if words else 0 # def compute_avg_sentence_length(self, text): # text = self.preprocess_text(text) # if not text: # return 0 # sentences = sent_tokenize(text) # return sum(len(word_tokenize(sent)) for sent in sentences) / len(sentences) if sentences else 0 # def compute_excessive_punctuation(self, text): # text = self.preprocess_text(text) # return len(re.findall(r'[!?.]{2,}', text)) # def compute_sentiment_polarity(self, text): # text = self.preprocess_text(text) # return TextBlob(text).sentiment.polarity if text else 0 # def compute_code_switching_flag(self, text): # text = self.preprocess_text(text) # if not text: # return 0 # tokens = self.tokenizer.tokenize(text.lower()) # if not tokens: # return 0 # english_words = self.stop_words # Use self.stop_words from __init__ # token_set = set(tokens) # english_count = sum(1 for token in tokens if token in english_words) # non_english_pattern = re.compile(r'[^\x00-\x7F]') # has_non_ascii = 1 if non_english_pattern.search(text) else 0 # english_ratio = english_count / len(tokens) if tokens else 0 # non_english_tokens = sum(1 for token in token_set if token not in english_words and "##" in token and has_non_ascii) # # Flag as code-switching if: # # 1. Mixed English presence (ratio between 0.1 and 0.9) # # 2. Non-ASCII characters present OR some non-English subword tokens # if 0.1 < english_ratio < 0.9 and (has_non_ascii or non_english_tokens > 0): # return 1 # return 0 # def batch_tokenize(self, texts, batch_size=32, max_length=512): # tokenized_outputs = [] # for i in tqdm(range(0, len(texts), batch_size), desc="Tokenizing with RoBERTa on GPU"): # batch_texts = texts[i:i + batch_size] # valid_texts = [self.preprocess_text(t) for t in batch_texts] # # Tokenize with fixed max_length to ensure consistent tensor sizes # inputs = self.tokenizer(valid_texts, return_tensors='pt', truncation=True, padding='max_length', max_length=max_length) # tokenized_outputs.append(inputs['input_ids'].to(self.device)) # Move to GPU # # Concatenate on GPU with consistent sizes # return torch.cat(tokenized_outputs, dim=0) # def compute_grammar_error_score(self, texts, tokenized_ids): # print("Computing grammar error scores...") # error_scores = np.zeros(len(texts), dtype=float) # vocab_set = set(self.tokenizer.get_vocab().keys()) # for i, input_ids in enumerate(tqdm(tokenized_ids, desc="Processing Grammar Errors")): # if input_ids.sum() == 0: # Empty input # continue # tokens = self.tokenizer.convert_ids_to_tokens(input_ids.cpu().tolist(), skip_special_tokens=True) # unknown_count = sum(1 for token in tokens if token not in vocab_set and token not in self.stop_words) # total_count = len([t for t in tokens if t not in self.stop_words]) # error_scores[i] = unknown_count / total_count if total_count > 0 else 0 # return error_scores # def compute_repetitive_words_count(self, texts, tokenized_ids): # print("Computing repetitive words counts...") # rep_counts = np.zeros(len(texts), dtype=int) # for i, input_ids in enumerate(tqdm(tokenized_ids, desc="Processing Repetition")): # if input_ids.sum() == 0: # Empty input # continue # tokens = self.tokenizer.convert_ids_to_tokens(input_ids.cpu().tolist(), skip_special_tokens=True) # valid_tokens = [t for t in tokens if t not in self.stop_words and len(t) > 2] # if valid_tokens: # token_counts = {} # for token in valid_tokens: # token_counts[token] = token_counts.get(token, 0) + 1 # rep_counts[i] = sum(1 for count in token_counts.values() if count > 1) # return rep_counts # def preprocess_text_for_similarity(self, text): # if pd.isna(text) or not text.strip(): # return [] # return [w for w in word_tokenize(str(text).lower()) if w not in self.stop_words] # def batch_encode_words(self, texts, batch_size=32, max_length=512): # word_lists = [self.preprocess_text_for_similarity(t) for t in tqdm(texts, desc="Tokenizing Texts")] # vocab = {word: idx + 1 for idx, word in enumerate(set.union(*[set(w) for w in word_lists if w]))} # encoded_batches = [] # for i in tqdm(range(0, len(word_lists), batch_size), desc="Encoding Words on GPU"): # batch_words = word_lists[i:i + batch_size] # encoded = np.zeros((len(batch_words), max_length), dtype=np.int64) # for j, words in enumerate(batch_words): # if words: # word_ids = [vocab.get(w, 0) for w in words][:max_length] # encoded[j, :len(word_ids)] = word_ids # encoded_tensor = torch.tensor(encoded, dtype=torch.int64).to(self.device) # encoded_batches.append(encoded_tensor) # return torch.cat(encoded_batches, dim=0), vocab # def compute_similarity_to_other_reviews(self, batch_size=32, max_length=512): # all_texts = self.df["review_text"].tolist() # all_users = self.df["user_id"].tolist() # all_review_ids = self.df["review_id"].tolist() # encoded_words, vocab = self.batch_encode_words(all_texts, batch_size, max_length) # similarity_scores = {rid: 0.0 for rid in all_review_ids} # Default scores # for i, (review_id, user_id) in enumerate(tqdm(zip(all_review_ids, all_users), desc="Computing Similarities on GPU")): # if pd.isna(review_id) or pd.isna(user_id): # continue # current_words = encoded_words[i] # if current_words.sum() == 0: # continue # other_indices = torch.tensor([j for j, u in enumerate(all_users) if u != user_id and pd.notna(u)], # dtype=torch.long).to(self.device) # if not other_indices.numel(): # continue # other_words = encoded_words[other_indices] # current_set = torch.unique(current_words[current_words > 0]) # other_flat = other_words[other_words > 0] # if other_flat.numel() == 0: # continue # other_set = torch.unique(other_flat) # intersection = torch.sum(torch.isin(current_set, other_set)).float() # union = torch.unique(torch.cat([current_set, other_set])).numel() # similarity = intersection / union if union > 0 else 0.0 # similarity_scores[review_id] = similarity.item() # return pd.Series(similarity_scores, index=all_review_ids) # def calculate_friend_count(self): # friends = [] # for v in self.df["friends"]: # if isinstance(v, str): # friends.append(len(v.split(","))) # elif type(v)==int or type(v)==float: # friends.append(0) # self.df["friends"] = friends # def count_elite_years(self, elite): # if pd.isna(elite): # return 0 # return len(str(elite).split(",")) # def transform_elite_status(self): # self.df["elite"] = self.df["elite"].apply(lambda x: True if self.count_elite_years(x) > 1 else False) # self.df["elite"] = self.df["elite"].astype(int) # def calculate_review_useful_funny_cool(self): # self.df["review_useful"] = pd.to_numeric(self.df["review_useful"], errors='coerce').fillna(0) # self.df["review_funny"] = pd.to_numeric(self.df["review_funny"], errors='coerce').fillna(0) # self.df["review_cool"] = pd.to_numeric(self.df["review_cool"], errors='coerce').fillna(0) # self.df["review_useful_funny_cool"] = ( # self.df["review_useful"] + # self.df["review_funny"] + # self.df["review_cool"] # ) # self.df["review_useful_funny_cool"] = self.df["review_useful_funny_cool"].fillna(0).astype(int) # def calculate_user_useful_funny_cool(self): # self.df["user_useful_funny_cool"] = ( # self.df["user_useful"] + # self.df["user_funny"] + # self.df["user_cool"] # ) # self.df["user_useful_funny_cool"] = self.df["user_useful_funny_cool"].fillna(0).astype(int) # def compute_fake_score(self, row): # suspicion_points = 0 # # Linguistic Features # if row["pronoun_density"] < 0.01: # Low personal engagement # suspicion_points += 1 # if row["avg_sentence_length"] < 5 or row["avg_sentence_length"] > 30: # Extreme lengths # suspicion_points += 1 # if row["grammar_error_score"] > 5: # Many errors # suspicion_points += 1 # if row["repetitive_words_count"] > 5: # High repetition # suspicion_points += 1 # if row["code_switching_flag"] == 1: # Language mixing # suspicion_points += 1 # if row["excessive_punctuation_count"] > 3: # Overuse of punctuation # suspicion_points += 1 # if abs(row["sentiment_polarity"]) > 0.8: # Extreme sentiment # suspicion_points += 1 # # Review Patterns # if row["similarity_to_other_reviews"] > 0.8: # High duplication # suspicion_points += 1 # if row["user_review_burst_count"] > 5: # Spammy bursts # suspicion_points += 1 # if row["business_review_burst_count"] > 5: # Targeted bursts # suspicion_points += 1 # if abs(row["rating_deviation_from_business_average"]) > 2: # Large rating deviation # suspicion_points += 1 # if row["review_like_ratio"] > 0.9 or row["review_like_ratio"] < 0.1: # Extreme like ratio # suspicion_points += 1 # # User Behavior # if row["user_account_age"] < 30: # Very new account (days) # suspicion_points += 1 # if row["average_time_between_reviews"] < 24: # Rapid reviews (hours) # suspicion_points += 1 # if row["user_degree"] < 2: # Low business interaction # suspicion_points += 1 # if row["time_since_last_review_user"] < 24: # Recent burst (hours) # suspicion_points += 1 # # Threshold: 3 or more points = fake # return 1 if suspicion_points >= 3 else 0 # def dropping_unncessary_columns(self): # self.df.drop("review_text", axis=1, inplace=True) # self.df.drop("review_date", axis=1, inplace=True) # self.df.drop("business_name", axis=1, inplace=True) # self.df.drop("address", axis=1, inplace=True) # self.df.drop("city", axis=1, inplace=True) # self.df.drop("state", axis=1, inplace=True) # self.df.drop("postal_code", axis=1, inplace=True) # self.df.drop("categories", axis=1, inplace=True) # self.df.drop("user_name", axis=1, inplace=True) # self.df.drop("yelping_since", axis=1, inplace=True) # self.df.drop("checkin_date", axis=1, inplace=True) # self.df.drop("review_useful", axis=1, inplace=True) # self.df.drop("review_funny", axis=1, inplace=True) # self.df.drop("review_cool", axis=1, inplace=True) # self.df.drop("user_useful", axis=1, inplace=True) # self.df.drop("user_funny", axis=1, inplace=True) # self.df.drop("user_cool", axis=1, inplace=True) # self.df.drop("is_open", axis=1, inplace=True) # self.df.drop("compliment_hot", axis=1, inplace=True) # self.df.drop("compliment_more", axis=1, inplace=True) # self.df.drop("compliment_profile", axis=1, inplace=True) # self.df.drop("compliment_cute", axis=1, inplace=True) # self.df.drop("compliment_list", axis=1, inplace=True) # self.df.drop("compliment_note", axis=1, inplace=True) # self.df.drop("compliment_plain", axis=1, inplace=True) # self.df.drop("compliment_cool", axis=1, inplace=True) # self.df.drop("compliment_funny", axis=1, inplace=True) # self.df.drop("compliment_writer", axis=1, inplace=True) # self.df.drop("compliment_photos", axis=1, inplace=True) # def run_pipeline(self): # logger.info("FINALYZING HOURS COLUMN ...") # self.df["hours"] = self.df["hours"].apply(self.get_avg_duration) # self.df["hours"] = self.df["hours"].fillna(0) # print(self.df["hours"][:10]) # print(self.df["hours"].isnull().sum()) # logger.info("FINALYZING ATTRIBUTES COLUMN ...") # self.df.drop("attributes",axis=1,inplace=True) # logger.info("CREATING time_since_last_review_user COLUMN ...") # self.calculate_time_since_last_review() # print(np.unique(self.df["time_since_last_review_user"] )) # logger.info("CREATING time_since_last_review_business COLUMN ...") # self.calculate_time_since_last_review_business() # print(np.unique(self.df["time_since_last_review_business"] )) # logger.info("CREATING user_account_age COLUMN ...") # self.calculate_user_account_age() # print(np.unique(self.df["user_account_age"] )) # logger.info("CREATING average_time_between_reviews COLUMN ...") # self.calculate_avg_time_between_reviews() # print(np.unique(self.df["average_time_between_reviews"] )) # logger.info("CREATING user_degree COLUMN ...") # self.calculate_user_degree() # print(np.unique(self.df["user_degree"] )) # logger.info("CREATING business_degree COLUMN ...") # self.calculate_business_degree() # print(np.unique(self.df["business_degree"] )) # logger.info("CREATING rating_variance_user COLUMN ...") # self.calculate_rating_variance_user() # print(np.unique(self.df["rating_variance_user"] )) # logger.info("CREATING user_review_burst_count COLUMN ...") # self.calculate_user_review_burst_count() # print(np.unique(self.df["user_review_burst_count"] )) # logger.info("CREATING business_review_burst_count COLUMN ...") # self.calculate_business_review_burst_count() # print(np.unique(self.df["business_review_burst_count"] )) # logger.info("CREATING temporal_similarity COLUMN ...") # self.calculate_temporal_similarity() # print(np.unique(self.df["temporal_similarity"] )) # logger.info("CREATING rating_deviation_from_business_average COLUMN ...") # self.calculate_rating_deviation_from_business_average() # print(np.unique(self.df["rating_deviation_from_business_average"] )) # logger.info("CREATING review_like_ratio COLUMN ...") # self.calculate_review_like_ratio() # print(np.unique(self.df["review_like_ratio"] )) # logger.info("CREATING latest_checkin_hours COLUMN ...") # self.calculate_latest_checkin_hours() # print(np.unique(self.df["latest_checkin_hours"] )) # logger.info("CREATING pronoun_density COLUMN ...") # self.df["pronoun_density"] = self.df["review_text"].apply(self.compute_pronoun_density) # print(np.unique(self.df["pronoun_density"] )) # logger.info("CREATING avg_sentence_length COLUMN ...") # self.df["avg_sentence_length"] = self.df["review_text"].apply(self.compute_avg_sentence_length) # print(np.unique(self.df["avg_sentence_length"] )) # logger.info("CREATING excessive_punctuation_count COLUMN ...") # self.df["excessive_punctuation_count"] = self.df["review_text"].apply(self.compute_excessive_punctuation) # print(np.unique(self.df["excessive_punctuation_count"] )) # logger.info("CREATING sentiment_polarity COLUMN ...") # self.df["sentiment_polarity"] = self.df["review_text"].apply(self.compute_sentiment_polarity) # print(np.unique(self.df["sentiment_polarity"] )) # logger.info("CREATING good_severity and bad_severity COLUMNS ...") # severity_scores = self.df["review_text"].apply(self.calculate_sentiment_severity) # self.df[["good_severity", "bad_severity"]] = severity_scores # print(np.unique(self.df["good_severity"] )) # print(np.unique(self.df["bad_severity"] )) # logger.info("CREATING code_switching_flag COLUMN ...") # self.df["code_switching_flag"] = self.df["review_text"].apply(self.compute_code_switching_flag) # print(np.unique(self.df["code_switching_flag"] )) # all_texts = self.df["review_text"].tolist() # tokenized_ids = self.batch_tokenize(all_texts, batch_size=32, max_length=512) # logger.info("CREATING grammar_error_score COLUMN ...") # self.df["grammar_error_score"] = self.compute_grammar_error_score(all_texts, tokenized_ids) # print(np.unique(self.df["grammar_error_score"] )) # logger.info("CREATING repetitive_words_count COLUMN ...") # self.df["repetitive_words_count"] = self.compute_repetitive_words_count(all_texts, tokenized_ids) # print(np.unique(self.df["repetitive_words_count"] )) # logger.info("CREATING similarity_to_other_reviews COLUMN ...") # similarity_scores = self.compute_similarity_to_other_reviews(batch_size=32, max_length=512) # self.df["similarity_to_other_reviews"] = self.df["review_id"].map(similarity_scores) # print(np.unique(self.df["similarity_to_other_reviews"] )) # logger.info("CREATING friends COLUMN ...") # self.calculate_friend_count() # print(self.df["friends"].value_counts()) # logger.info("CREATING elite COLUMN ...") # self.transform_elite_status() # print(self.df["elite"].value_counts()) # logger.info("CREATING review_useful_funny_cool COLUMN ...") # self.calculate_review_useful_funny_cool() # print(self.df["review_useful_funny_cool"].value_counts()) # logger.info("CREATING user_useful_funny_cool COLUMN ...") # self.calculate_user_useful_funny_cool() # print(self.df["user_useful_funny_cool"].value_counts()) # # logger.info("CREATING LABEL COLUMN ...") # # self.df["fake"] = self.df.apply(self.compute_fake_score, axis=1) # # print(self.df["fake"].value_counts()) # logger.info("DELETING THE UNWANTED COLUMNS ...") # self.dropping_unncessary_columns() # print() # logger.info("SEEING NULL VALUES IN FINAL COLUMNS.....") # print(set(self.df.isnull().sum().values)) # return self.df from loguru import logger import pandas as pd import json from datetime import datetime import ast import numpy as np from pymongo import MongoClient from collections import defaultdict from tqdm import tqdm import time import requests import json import os import pandas as pd import nltk from nltk.tokenize import sent_tokenize, word_tokenize from nltk.corpus import stopwords from textblob import TextBlob import re from transformers import BertTokenizer, BertModel from transformers import RobertaTokenizer, RobertaModel import torch from sklearn.metrics.pairwise import cosine_similarity import numpy as np from sklearn.preprocessing import StandardScaler from sklearn.preprocessing import LabelEncoder from transformers import BertTokenizer, BertModel import torch from sklearn.metrics.pairwise import cosine_similarity import re from tqdm import tqdm from nltk.sentiment.vader import SentimentIntensityAnalyzer import nltk from scipy.stats import zscore from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.metrics.pairwise import cosine_similarity # Download NLTK resources nltk.download('punkt') nltk.download('averaged_perceptron_tagger') nltk.download('stopwords') nltk.download('punkt_tab') nltk.download('averaged_perceptron_tagger_eng') class Preprocessor: def __init__(self,df): self.df=df def create_columns(self): self.df["topic_creation_text"]=self.df["desc"]+self.df["title"]+self.df["share_title"] self.df['create_time_dt'] = pd.to_datetime(self.df['create_time'], unit='s') self.df['create_year'] = self.df['create_time_dt'].dt.year self.df['create_month'] = self.df['create_time_dt'].dt.month self.df['create_day'] = self.df['create_time_dt'].dt.day self.df['create_hour'] = self.df['create_time_dt'].dt.hour self.df['create_day_of_week'] = self.df['create_time_dt'].dt.dayofweek self.df['create_is_weekend'] = self.df['create_day_of_week'].isin([5, 6]).astype(int) self.df['create_days_since_creation'] = (pd.Timestamp('2025-04-27') - self.df['create_time_dt']).dt.days self.df.drop(["create_time"],axis=1,inplace=True) def average_posting_freq_per_day(self): sec_id = list(set(self.df["sec_id"])) posts_per_user = [] for i, user in enumerate(self.df["sec_id"]): try: days = (pd.to_datetime("2025-05-11") - pd.to_datetime(self.df["birthday"].iloc[i])).days sec_id = self.df["sec_id"].iloc[i] posts = len(self.df[self.df["sec_id"] == sec_id]) posts_per_day = posts / days if days > 0 else 0 posts_per_user.append(posts_per_day) except (ValueError, TypeError): print("error") posts_per_user.append(None) self.df["posts_per_user"]=posts_per_user def diversity_topics_about(self): self.df["diversity_of_topics"] = self.df.groupby('sec_id')['topic'].transform('nunique') / 150 def avg_post_length(self): try: avg_post_length=[] for user in self.df["sec_id"]: df_user=self.df[self.df["sec_id"]==user] post_lengths=0 for posts in df_user["desc"]: post_lengths+=len(posts) post_lengths=(post_lengths/len(df_user)) avg_post_length.append(post_lengths) except Exception as e: print(f"ERROR for : {posts} as {e}") self.df["avg_post_length"]=avg_post_length def sentiment_profile_of_posts(self): nltk.download('vader_lexicon') try: # Initialize VADER sentiment analyzer sia = SentimentIntensityAnalyzer() # Calculate average sentiment score per user self.df["avg_sentiment_score"] = self.df.groupby('sec_id')['desc'].transform( lambda x: x.apply(lambda text: sia.polarity_scores(text)['compound']).mean()) except Exception as e: logger.info(f"ERROR AS {e}") def get_words(self,text): if pd.isna(text): return [] # Convert to lowercase, remove punctuation, split into words text = re.sub(r'[^\w\s]', '', str(text).lower()) return text.split() def lexical_diversity(self): self.df['lexical_diversity'] = self.df.groupby('sec_id')['desc'].transform( lambda x: len(set(word for text in x for word in self.get_words(text))) / len([word for text in x for word in self.get_words(text)]) if len([word for text in x for word in self.get_words(text)]) > 0 else 0) def popularity(self): self.df['popularity'] = self.df.groupby('sec_id')['sec_id'].transform('count') def sentiment_distribution(self): nltk.download('vader_lexicon') sia = SentimentIntensityAnalyzer() self.df['avg_sentiment_score'] = self.df.groupby('sec_id')['desc'].transform( lambda x: x.apply(lambda text: sia.polarity_scores(str(text))['compound'] if not pd.isna(text) else 0).mean() ) self.df['sentiment_variance'] = self.df.groupby('sec_id')['desc'].transform( lambda x: x.apply(lambda text: sia.polarity_scores(str(text))['compound'] if not pd.isna(text) else 0).var(ddof=1) if len(x) > 1 else 0 ) def post_length(self): post_length=[] for posts in self.df["desc"]: post_length.append(len(posts)) self.df["post_length"]=post_length def sentiment_scores_post(self): sia = SentimentIntensityAnalyzer() self.df['sentiment_score'] = self.df['desc'].apply( lambda x: sia.polarity_scores(str(x))['compound'] if not pd.isna(x) else 0) def timing_post(self): self.df = self.df.sort_values(['sec_id', 'create_time_dt']) self.df['time_since_prev_post'] = self.df.groupby('sec_id')['create_time_dt'].diff().dt.total_seconds().fillna(0) def compute_lexical_similarity(self,group): group = group.sort_values('create_time_dt') texts = group['desc'].fillna('').astype(str).tolist() similarity_scores = [] if len(texts) <= 1: return [0] * len(texts) vectorizer = TfidfVectorizer() tfidf_matrix = vectorizer.fit_transform(texts) for i in range(len(texts)): if i == 0: similarity_scores.append(0) # No previous posts for the first post else: sim = cosine_similarity(tfidf_matrix[i:i+1], tfidf_matrix[:i]).flatten() similarity_scores.append(np.mean(sim) if len(sim) > 0 else 0) return similarity_scores def lexical_similarity(self): self.df['lexical_similarity'] = self.df.groupby('sec_id').apply(self.compute_lexical_similarity).explode().reset_index(drop=True).astype(float) def ratio(self): ratio=[] for ratios in self.df["ratio"]: ratio.append(int(ratios[:-1])) self.df["ratio"]=ratio def encoding(self): le=LabelEncoder() cat_columns=list(self.df.select_dtypes(exclude=np.number).columns) cat_columns.remove("sec_id") for col in cat_columns: self.df[col]=le.fit_transform(self.df[col]) def label_creation(self): user_stats = self.df.groupby('sec_id').agg({ 'aweme_count': 'mean', 'create_days_since_creation': 'mean', 'digg_count': 'mean', 'play_count': 'mean', 'comment_count': 'mean', 'follower_count': 'mean', 'lexical_similarity': 'mean', 'diversity_of_topics': 'mean', 'sentiment_score': 'mean', 'posts_per_user': 'mean' }).reset_index() user_stats['posting_rate'] = user_stats['posts_per_user'] / (user_stats['create_days_since_creation'] + 1) user_stats['engagement_ratio'] = (user_stats['digg_count'] + user_stats['comment_count'] + user_stats['play_count']) / (user_stats['follower_count'] + 1) features = ['posting_rate', 'engagement_ratio', 'lexical_similarity', 'diversity_of_topics', 'create_days_since_creation'] for feature in features: user_stats[f'{feature}_z'] = zscore(user_stats[feature].fillna(user_stats[feature].mean())) user_stats['fake_score'] = ( 1.0 * user_stats['posting_rate_z'] + # High posting rate -> fake -1.0 * user_stats['engagement_ratio_z'] + # Low engagement -> fake 1.0 * user_stats['lexical_similarity_z'] + # High similarity -> fake -1.0 * user_stats['diversity_of_topics_z'] + # Low diversity -> fake -1.0 * user_stats['create_days_since_creation_z'] # New account -> fake ) user_stats = user_stats.sort_values(by='fake_score', ascending=False) n_users = len(user_stats) n_fake = int(0.5 * n_users) user_stats['is_fake'] = 0 # Initialize all as genuine user_stats.iloc[:n_fake, user_stats.columns.get_loc('is_fake')] = 1 # Set top 75% to fake user_labels = user_stats[['sec_id', 'is_fake']] # fake_percentage = user_labels['is_fake'].mean() * 100 self.df = self.df.merge(user_labels[['sec_id', 'is_fake']], on='sec_id', how='left') def run_pipeline(self): logger.info("INSIDE TEST PREPROCESS") print(self.df) self.df["signature"].fillna("",inplace=True) self.df["desc"].fillna("",inplace=True) self.df["title"].fillna("",inplace=True) self.df.dropna(inplace=True) self.df["topic_creation_text"]=self.df["desc"]+self.df["title"]+self.df["share_title"] logger.info("CREATING COLUMN PREDICT...") self.create_columns() logger.info("AVERAGE POSTING PER FREQ PER DAY ...") self.average_posting_freq_per_day() logger.info("DIVERSITY TOPICS ABOUT ...") self.diversity_topics_about() logger.info("AVG POST LENGTH ...") self.avg_post_length() logger.info("SENTIMENT PROFILE OF POSTS...") self.sentiment_profile_of_posts() logger.info(f"{self.df.isnull().sum()} NAN values after creating dataset ") logger.info("LEXICAL DIVERSITY ...") self.lexical_diversity() logger.info("POPULARITY...") self.popularity() logger.info("SENTIMENT DISTRIBUTION ...") self.sentiment_distribution() logger.info("POST LENGTH ...") self.post_length() logger.info("SENTIMENT SCORES POST...") self.sentiment_scores_post() logger.info("TIMING POST...") self.timing_post() logger.info("LEXICAL SIMILARITY ...") self.lexical_similarity() self.df.drop(["desc","title","share_title","birthday","topic_creation_text","signature","create_time_dt"],axis=1,inplace=True) self.df.reset_index(drop=True,inplace=True) logger.info("ENCODING ...") self.encoding() # logger.info(" LABEL CREATION ...") # self.label_creation() self.df.dropna(inplace=True) print(self.df.info()) print(self.df.isnull().sum()) # self.df.to_csv(os.path.join(self.path,self.filename),index=False) return self.df