Table of Contents¶

  • Required Packages
  • Conducting Initial Data Investigation
    • Findings
  • Conducting Initial Topic Modelling
    • Findings
  • Performing Further Data Investigation
    • Findings
  • Conducting Emotion Analysis
    • Findings
  • Using a Large Language Model from Hugging Face (PHI)
    • Findings
  • Using Gensim for Topic Modelling
    • Findings

Required Packages¶

In [ ]:
!pip install bertopic gensim pyLDAvis ipykernel
In [2]:
import numpy as np
import pandas as pd
from nltk.corpus import stopwords
from nltk import word_tokenize
import nltk
from nltk.probability import FreqDist
import matplotlib.pyplot as plt
from wordcloud import WordCloud, STOPWORDS
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from bertopic import BERTopic
import torch
from gensim import corpora
from gensim.models.ldamodel import LdaModel
import pyLDAvis.gensim_models as gensimvis
import pyLDAvis
In [3]:
nltk.download("stopwords")
nltk.download("punkt_tab")

[nltk_data] Downloading package stopwords to /root/nltk_data...
[nltk_data]   Unzipping corpora/stopwords.zip.
[nltk_data] Downloading package punkt_tab to /root/nltk_data...
[nltk_data]   Unzipping tokenizers/punkt_tab.zip.
Out[3]:
True

Conducting initial data investigation¶

In [4]:
googleReview = pd.read_excel("Google_12_months.xlsx")
trustPilot = pd.read_excel("Trustpilot_12_months.xlsx")
In [5]:
googleReview.rename(columns= {"Club's Name": "Location Name", "Overall Score": "Review Score"}, inplace = True)
trustPilot.rename(columns= {"Review Content": "Comment", "Review Stars": "Review Score"}, inplace = True)
In [6]:
def data_preprocessing(data, col_name):

  data = data.copy()
  print(data.shape)
  stop_words = set(stopwords.words("english"))
  data = data.dropna(subset = [col_name])
  classifier = pipeline("text-classification", model="papluca/xlm-roberta-base-language-detection", truncation=True)

  data[col_name] = data[col_name].str.replace(r'\d+', '', regex = True)

  def process_text(text):

    lang_result = classifier(text)

    if lang_result[0]["label"] == "en":

      return " ".join([word for word in word_tokenize(text.lower()) if word not in stop_words and word.isalpha()])
    return ""

  data[col_name] = data[col_name].apply(process_text)
  return data
In [7]:
col_name = "Comment"
print("Google Reviews")
clean_google = data_preprocessing(googleReview, col_name)
print("Trustpilot Reviews")
clean_pilot = data_preprocessing(trustPilot, col_name)

Google Reviews
(23250, 7)

config.json:   0%|          | 0.00/1.42k [00:00<?, ?B/s]
model.safetensors:   0%|          | 0.00/1.11G [00:00<?, ?B/s]
tokenizer_config.json:   0%|          | 0.00/502 [00:00<?, ?B/s]
(…)6de7952f7c11ab059eca145a0a727afce0db2865:   0%|          | 0.00/5.07M [00:00<?, ?B/s]
tokenizer.json:   0%|          | 0.00/9.08M [00:00<?, ?B/s]
special_tokens_map.json:   0%|          | 0.00/239 [00:00<?, ?B/s]
Device set to use cuda:0
You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset

Trustpilot Reviews
(16673, 15)

Device set to use cuda:0
In [8]:
def wordcloud(cloud):

  custom_stopwords = {"gym", "puregym"}
  stopwords = STOPWORDS.union(custom_stopwords)

  for title, i in  cloud.items():

    wordcloud = WordCloud(stopwords= stopwords, width=800, height=400, background_color='white').generate(" ".join(i))

    plt.figure(figsize=(10, 5))
    plt.imshow(wordcloud, interpolation='bilinear')
    plt.axis('off')
    plt.title(f"The wordcloud for: {title}")
    plt.show()
In [9]:
def analysis(data):

  neg_reviews = {}

  for name, df in data.items():

    all_words = [word for comment in df['Comment'] for word in word_tokenize(comment)]

    print(f"{name} have {df['Location Name'].str.lower().nunique()} unique location")
    print(f"{name} frequency distribution is {FreqDist(all_words)}")
    print(f"{name} has this shape {df.shape}")

    common_words = FreqDist(all_words).most_common(10)
    neg_reviews[name.split()[0]] = df[df["Review Score"] < 3]
    neg_words = [word for comment in neg_reviews[name.split()[0]]['Comment'] for word in word_tokenize(comment)]

    print(f"{name} frequency distribution of negative words is {FreqDist(neg_words)}")
    print(f"{name} negative reviews has this shape {neg_reviews[name.split()[0]].shape}")

    words, counts = zip(*common_words)

    plt.figure(figsize=(10, 5))
    plt.bar(words, counts, color='skyblue')
    plt.xlabel("Words")
    plt.ylabel("Frequency")
    plt.title(f"Top 10 Most Common Words")
    plt.show()

    cloud = {f"All Reviews for {name.split()[0]}": all_words, f"Negative Reviews for {name.split()[0]}":neg_words}

    wordcloud(cloud)

  return neg_reviews
In [10]:
datasets = {"google review": clean_google, "trustpilot reviews": clean_pilot}
neg_dataset = analysis(datasets)

google review have 512 unique location
google review frequency distribution is <FreqDist with 13492 samples and 241042 outcomes>
google review has this shape (13898, 7)
google review frequency distribution of negative words is <FreqDist with 8414 samples and 81163 outcomes>
google review negative reviews has this shape (2785, 7)

trustpilot reviews have 374 unique location
trustpilot reviews frequency distribution is <FreqDist with 12859 samples and 277004 outcomes>
trustpilot reviews has this shape (16673, 15)
trustpilot reviews frequency distribution of negative words is <FreqDist with 8410 samples and 95444 outcomes>
trustpilot reviews negative reviews has this shape (3543, 15)
In [11]:
common_locations = len(set(clean_google["Location Name"].str.lower()).intersection(set(clean_pilot["Location Name"].str.lower())))
print(f'Number of common locations: {common_locations}')

Number of common locations: 312

Findings¶

Dataset Before Cleaning¶

  • Google dataset contains 23,250 data points with 7 features, while Trustpilot dataset contains 16,673 data points with 15 features.

Dataset Cleaning¶

  • Removed rows where the comments column is empty and where the language is not English.
  • Text processing steps:
    • Stopwords removed.
    • Non-alphabetic characters removed.
    • Text tokenized.

Initial Analysis After Cleaning¶

PureGym Reviews on Google¶

  • The cleaned Google dataset now has 13,898 data points.
  • Comments feature:
    • 13,492 unique words.
    • 241,042 total words.
  • Location Name feature:
    • 512 unique locations.
  • Top 10 words indicate a positive sentiment, featuring words like clean and great.
  • Word cloud analysis also suggests a positive sentiment with prominent words like nice and love.

Negative Reviews (Score < 3)¶

  • 2,785 data points classified as negative.
  • 8,414 unique words, 81,163 total words.
  • Word cloud reveals more negative words such as dirty and broken.

PureGym Reviews on Trustpilot¶

  • The cleaned Trustpilot dataset still has 16,673 data points.
  • Comments feature:
    • 12,859 unique words.
    • 277,004 total words.
  • Location Name feature:
    • 374 unique locations.
  • Top 10 words suggest a positive sentiment similar to Google reviews, with words like clean and great.
  • Word cloud is less conclusive, showing words like great and good, making it harder to determine an overall sentiment.

Negative Reviews (Score < 3)¶

  • 3,543 data points classified as negative.
  • 8,410 unique words, 95,444 total words.
  • Word cloud highlights specific negative topics, such as staff and showers, which may contribute to the dissatisfaction.

Common Locations¶

  • Number of common locations between Google and Trustpilot: 312

Conducting initial topic modelling¶

In [12]:
google_neg = None
trustpilot_neg = None


for name, df in neg_dataset.items():
  if name == "google":
    google_neg = df
  else:
    trustpilot_neg = df
In [13]:
common_google = google_neg[google_neg["Location Name"].isin(trustpilot_neg["Location Name"])][["Comment","Location Name"]]
common_trustpilot = trustpilot_neg[trustpilot_neg["Location Name"].isin(common_google["Location Name"])][["Comment", "Location Name"]]
merged_data = pd.concat([common_google, common_trustpilot], ignore_index=True)
In [14]:
merged_data.shape
Out[14]:
(3888, 2)
In [37]:
def bert_modelling(data):

  data = data.copy()

  merged_neg_clean = data["Comment"].tolist()
  model = BERTopic(verbose=True)
  model.fit(merged_neg_clean)
  topics, probabilities = model.transform(merged_neg_clean)

  data["topic"] = topics

  return model, data, topics, probabilities
In [38]:
model, topics_data, topics, probabilities = bert_modelling(merged_data)

2025-04-01 03:01:29,078 - BERTopic - Embedding - Transforming documents to embeddings.

Batches:   0%|          | 0/122 [00:00<?, ?it/s]
2025-04-01 03:01:31,289 - BERTopic - Embedding - Completed ✓
2025-04-01 03:01:31,290 - BERTopic - Dimensionality - Fitting the dimensionality reduction algorithm
2025-04-01 03:01:54,395 - BERTopic - Dimensionality - Completed ✓
2025-04-01 03:01:54,396 - BERTopic - Cluster - Start clustering the reduced embeddings
2025-04-01 03:01:54,540 - BERTopic - Cluster - Completed ✓
2025-04-01 03:01:54,544 - BERTopic - Representation - Fine-tuning topics using representation models.
2025-04-01 03:01:54,694 - BERTopic - Representation - Completed ✓

Batches:   0%|          | 0/122 [00:00<?, ?it/s]
2025-04-01 03:01:55,924 - BERTopic - Dimensionality - Reducing dimensionality of input embeddings.
2025-04-01 03:01:55,936 - BERTopic - Dimensionality - Completed ✓
2025-04-01 03:01:55,937 - BERTopic - Clustering - Approximating new points with `hdbscan_model`
2025-04-01 03:01:56,069 - BERTopic - Cluster - Completed ✓
In [40]:
def bert_visuals(bert_model):

  print(f"Total number of topics: {len(bert_model.get_topics())}")
  topic_freq = bert_model.get_topic_freq()
  print(f"Outliers:\n {topic_freq[topic_freq['Topic'] == -1]} \n")
  print(f"Top topics along with their document frequencies: \n {topic_freq[topic_freq['Topic'] != -1].head(10)} \n")
  top_2_topics = topic_freq[topic_freq['Topic'] != -1].head(2)["Topic"].tolist()
  for tp in top_2_topics:
    print(f"Topic {tp} has the following top words: \n {bert_model.get_topic(tp)} \n")

  fig1 = bert_model.visualize_topics()
  fig1.show()

  fig2 = bert_model.visualize_barchart()
  fig2.show()

  fig3 = bert_model.visualize_heatmap()
  fig3.show()
In [41]:
bert_visuals(model)

Total number of topics: 50
Outliers:
    Topic  Count
4     -1   1628 

Top topics along with their document frequencies: 
     Topic  Count
2       0    195
11      1    150
10      2    142
3       3    137
12      4    123
20      5    115
27      6     83
7       7     71
8       8     58
41      9     57 

Topic 0 has the following top words: 
 [('air', 0.06535853622220478), ('conditioning', 0.039109409267406005), ('hot', 0.0341153751577972), ('aircon', 0.030904603329895447), ('con', 0.02539572030351039), ('ac', 0.02335473164547456), ('summer', 0.022840948850730125), ('heat', 0.022368245534383064), ('working', 0.018771570123393193), ('temperature', 0.017792814503912378)] 

Topic 1 has the following top words: 
 [('rude', 0.021586991254173425), ('manager', 0.02069975537194293), ('staff', 0.01869429753354046), ('member', 0.01668603498125788), ('gym', 0.01588568912675759), ('women', 0.012400406597117608), ('personal', 0.011682768418990724), ('reviews', 0.011341051914656263), ('men', 0.010887589494607062), ('told', 0.01059952186619874)]
In [19]:
top_10_topic_Location = topics_data[(topics_data["topic"] > -1) & (topics_data["topic"] < 10)]
top_10_topic_Location[["Location Name", "topic"]].value_counts().head(10)
Out[19]:
count
Location Name topic
Leicester Walnut Street 0 28
Nottingham Colwick 5 16
Billericay 5 13
London Stratford 2 12
Aylesbury 1 10
Walsall Crown Wharf 5 10
Reading Calcot 0 9
Norwich Riverside 1 8
Burnham 2 8
Paisley 0 7

Findings¶

Pre-processing before topic modelling¶

  • The PureGym reviews datasets from Google and Trustpilot have been merged where both have a common location and the review score is negative (<3). The Comments feature was used for topic modelling after merging.
  • The merged results contain only 3,888 data points.

Topic Modelling Results (Bertopic)¶

  • Bertopic classified the datasets into 51 topics, including outliers.
  • 1,516 data points have been classed as outliers, which is a significant number given we only have 3,888 data points.
  • Looking at words in the top two topics (Topic 0: 185 and Topic 1: 169), we can see that air conditioning and facility issues (cold showers) may be the issue here.
  • The intertopic distance map shows 8 distinct topics, with the remaining topics clustered together as part of these 8 topics.
  • Looking at the words from the top 8 topics, the issues become clear. From these, we can deduce that the Topic 8 issues are:
    • Air conditioning at the gym.
    • Rude staff.
    • Problem with classes.
    • Access issues (pin code).
    • Dirty toilet.
    • Parking issues (fines).
    • Dirty gym.
    • Loud music.
  • The similarity matrix between topics indicates that Bertopic has done a good job in separating the topics, given how lighter the colors are.
  • Nottingham Colwick, Billericay, and Walsall Crown Wharf lead the way with 16, 13, and 10 occurrences of Topic 4, which appears to be related to parking issues.
  • London Stratford leads the way on Topic 1 with 10, which appears to be related to rude staff/staff issues.
  • Aylesbury also leads the way for Topic 0, which has to do with air conditioning at the gym.

Performing further data investigation¶

In [20]:
print(f"google top 20 locations with the most negative reviews {common_google['Location Name'].value_counts().head(20)}")
print("\n")
print(f"Trustpilot top 20 locations with the most negative reviews {common_trustpilot['Location Name'].value_counts().head(20)}")

google top 20 locations with the most negative reviews Location Name
London Stratford            59
London Canary Wharf         26
London Enfield              25
London Swiss Cottage        24
Birmingham City Centre      21
London Leytonstone          21
New Barnet                  20
Wakefield                   19
Bradford Thornbury          19
Walsall Crown Wharf         18
London Hoxton               18
London Seven Sisters        18
London Hayes                17
Manchester Exchange Quay    17
Sutton Times Square         16
London Bermondsey           16
Nottingham Colwick          16
London Piccadilly           15
Leeds City Centre North     15
London Muswell Hill         15
Name: count, dtype: int64


Trustpilot top 20 locations with the most negative reviews Location Name
Leicester Walnut Street      50
London Enfield               23
London Stratford             22
Burnham                      20
London Bermondsey            18
York                         16
Maidenhead                   16
London Seven Sisters         16
London Finchley              16
London Hayes                 16
London Swiss Cottage         15
London Hammersmith Palais    15
Northwich                    15
London Bromley               15
Dudley Tipton                14
Watford Waterfields          14
Basildon                     14
Bradford Thornbury           14
Telford                      14
Birmingham City Centre       14
Name: count, dtype: int64
In [21]:
google_top_20 = common_google['Location Name'].value_counts().head(20).reset_index()
google_top_20.columns = ['Location Name', 'Google Negative Reviews']

trustpilot_top_20 = common_trustpilot['Location Name'].value_counts().head(20).reset_index()
trustpilot_top_20.columns = ['Location Name', 'Trustpilot Negative Reviews']

merged_top_20 = pd.merge(google_top_20, trustpilot_top_20, on="Location Name", how="inner")

print("Common Locations with Negative Reviews from Both Google and Trustpilot:")
merged_top_20

Common Locations with Negative Reviews from Both Google and Trustpilot:
Out[21]:
Location Name Google Negative Reviews Trustpilot Negative Reviews
0 London Stratford 59 22
1 London Enfield 25 23
2 London Swiss Cottage 24 15
3 Birmingham City Centre 21 14
4 Bradford Thornbury 19 14
5 London Seven Sisters 18 16
6 London Hayes 17 16
7 London Bermondsey 16 18
In [22]:
google_neg_loc = google_neg["Location Name"].value_counts().reset_index()
pilot_neg_loc = trustpilot_neg["Location Name"].value_counts().reset_index()
In [23]:
merged_count = pd.merge(google_neg_loc, pilot_neg_loc, on='Location Name', how='outer')
merged_count["count_x"] = merged_count["count_x"].fillna(0)
merged_count["count_y"] = merged_count["count_y"].fillna(0)

merged_count["Total by loaction"] = merged_count["count_x"] + merged_count["count_y"]
merged_count = merged_count.sort_values(by='Total by loaction', ascending= False)
In [24]:
merged_count
Out[24]:
Location Name count_x count_y Total by loaction
362 London Stratford 59.0 22.0 81.0
289 Leicester Walnut Street 11.0 50.0 61.0
326 London Enfield 25.0 23.0 48.0
0 345 0.0 45.0 45.0
364 London Swiss Cottage 24.0 15.0 39.0
... ... ... ... ...
278 La Chaux de Fonds 1.0 0.0 1.0
59 247 - Esbjerg, Randersvej 1.0 0.0 1.0
61 249 - Ikast, Strøget 1.0 0.0 1.0
64 255 - Kalundborg, Sct. Jørgensbjerg 1.0 0.0 1.0
12 124 - Kbh. NV., Lygten 1.0 0.0 1.0

539 rows × 4 columns

In [81]:
google_top_30 = google_neg[google_neg["Location Name"].isin(merged_count["Location Name"].head(30))][["Comment", "Location Name"]]
pilot_top_30 = trustpilot_neg[trustpilot_neg["Location Name"].isin(merged_count["Location Name"].head(30))][["Comment", "Location Name"]]

top_30_wordcloud = {"Google": google_top_30["Comment"], "Trustpilot": pilot_top_30["Comment"]}
wordcloud(top_30_wordcloud)
In [86]:
top_30_comments =  pd.concat([google_top_30,  pilot_top_30], ignore_index = True)
df_duplicated = pd.concat([top_30_comments, top_30_comments], ignore_index=True)
model_30, top_30_data, topic_30, probabilities_30  = bert_modelling(df_duplicated)

2025-04-01 04:03:29,736 - BERTopic - Embedding - Transforming documents to embeddings.

Batches:   0%|          | 0/62 [00:00<?, ?it/s]
2025-04-01 04:03:30,963 - BERTopic - Embedding - Completed ✓
2025-04-01 04:03:30,964 - BERTopic - Dimensionality - Fitting the dimensionality reduction algorithm
2025-04-01 04:03:38,200 - BERTopic - Dimensionality - Completed ✓
2025-04-01 04:03:38,201 - BERTopic - Cluster - Start clustering the reduced embeddings
2025-04-01 04:03:38,270 - BERTopic - Cluster - Completed ✓
2025-04-01 04:03:38,274 - BERTopic - Representation - Fine-tuning topics using representation models.
2025-04-01 04:03:38,389 - BERTopic - Representation - Completed ✓

Batches:   0%|          | 0/62 [00:00<?, ?it/s]
2025-04-01 04:03:39,148 - BERTopic - Dimensionality - Reducing dimensionality of input embeddings.
2025-04-01 04:03:39,155 - BERTopic - Dimensionality - Completed ✓
2025-04-01 04:03:39,155 - BERTopic - Clustering - Approximating new points with `hdbscan_model`
2025-04-01 04:03:39,221 - BERTopic - Cluster - Completed ✓
In [87]:
bert_visuals(model_30)

Total number of topics: 70
Outliers:
    Topic  Count
1     -1    304 

Top topics along with their document frequencies: 
     Topic  Count
6       0     88
17      1     72
4       2     68
9       3     62
42      4     58
22      5     42
0       6     42
14      7     38
68      8     36
10      9     36 

Topic 0 has the following top words: 
 [('smell', 0.03209920574772303), ('smelly', 0.030228486724505444), ('air', 0.029442284051825408), ('disgusting', 0.024554686369513766), ('aircon', 0.02451718576468661), ('worst', 0.023792796209157436), ('stinks', 0.02200092847548957), ('ever', 0.021161352942561253), ('ventilation', 0.0196847984820827), ('urine', 0.018953204835633626)] 

Topic 1 has the following top words: 
 [('pin', 0.07810738736606905), ('pass', 0.06562447771361557), ('code', 0.05415072747804314), ('day', 0.05070777212915479), ('sent', 0.041654405752340876), ('app', 0.039985745448062736), ('access', 0.038156671096345636), ('bought', 0.03340465626022348), ('work', 0.030147532664350097), ('number', 0.029989309086047054)]

Findings¶

Top Locations for Bad Reviews¶

  • The top three locations for negative reviews in both datasets are:
    • London Stratford
    • Leicester Walnut Street
    • London Enfield
  • Each of these locations has at least 50+ negative reviews.

Topic Modeling¶

(Note: Due to a lack of data, the top 30 location data has been duplicated to ensure enough data for BERTopic and the intertopic distance map.)

  • The topics follow a similar theme to the previous BERTopic results.
  • However, there are two noticeable additions to the top topics:
    • Gym equipment issues.
    • Shower problems.

Conducting emotion analysis¶

In [46]:
model_name = "bhadresh-savani/bert-base-uncased-emotion"
classifier = pipeline("text-classification", model = model_name, truncation=True)

config.json:   0%|          | 0.00/935 [00:00<?, ?B/s]
model.safetensors:   0%|          | 0.00/438M [00:00<?, ?B/s]
tokenizer_config.json:   0%|          | 0.00/285 [00:00<?, ?B/s]
vocab.txt:   0%|          | 0.00/232k [00:00<?, ?B/s]
tokenizer.json:   0%|          | 0.00/466k [00:00<?, ?B/s]
special_tokens_map.json:   0%|          | 0.00/112 [00:00<?, ?B/s]
Device set to use cuda:0
In [47]:
google_neg_emotion = google_neg[["Comment", "Location Name"]].copy()
trustpilot_neg_emotion = trustpilot_neg[["Comment", "Location Name"]].copy()
In [48]:
def emotion_analysis(text):

  result = classifier(text)[0]["label"]

  return result
In [49]:
google_neg_emotion["emotion"] = google_neg_emotion["Comment"].apply(emotion_analysis)
trustpilot_neg_emotion["emotion"] = trustpilot_neg_emotion["Comment"].apply(emotion_analysis)
In [50]:
def emotion_chart(values):

  for name, i in values.items():

    plt.bar(i.index, i.values)
    plt.xlabel("Emotion")
    plt.ylabel("Distribution")
    plt.title(f"Distribution for all negative reviews in {name}")
    plt.show()
In [51]:
emotion_count_google = google_neg_emotion["emotion"].value_counts()
emotion_count_pilot = trustpilot_neg_emotion["emotion"].value_counts()
In [52]:
emotion_chart({"google": emotion_count_google, "Trustpilot" : emotion_count_pilot})
In [56]:
neg_only= pd.concat([google_neg_emotion[google_neg_emotion["emotion"]== "anger"], trustpilot_neg_emotion[trustpilot_neg_emotion["emotion"]== "anger"]], ignore_index = True)
In [57]:
model_anger, anger_data, topic_anger, probabilities_anger  = bert_modelling(neg_only)

2025-04-01 03:13:58,891 - BERTopic - Embedding - Transforming documents to embeddings.

Batches:   0%|          | 0/76 [00:00<?, ?it/s]
2025-04-01 03:14:00,256 - BERTopic - Embedding - Completed ✓
2025-04-01 03:14:00,257 - BERTopic - Dimensionality - Fitting the dimensionality reduction algorithm
2025-04-01 03:14:12,737 - BERTopic - Dimensionality - Completed ✓
2025-04-01 03:14:12,739 - BERTopic - Cluster - Start clustering the reduced embeddings
2025-04-01 03:14:12,824 - BERTopic - Cluster - Completed ✓
2025-04-01 03:14:12,827 - BERTopic - Representation - Fine-tuning topics using representation models.
2025-04-01 03:14:12,916 - BERTopic - Representation - Completed ✓

Batches:   0%|          | 0/76 [00:00<?, ?it/s]
2025-04-01 03:14:13,683 - BERTopic - Dimensionality - Reducing dimensionality of input embeddings.
2025-04-01 03:14:13,691 - BERTopic - Dimensionality - Completed ✓
2025-04-01 03:14:13,692 - BERTopic - Clustering - Approximating new points with `hdbscan_model`
2025-04-01 03:14:13,776 - BERTopic - Cluster - Completed ✓
In [58]:
bert_visuals(model_anger)

Total number of topics: 36
Outliers:
    Topic  Count
0     -1    826 

Top topics along with their document frequencies: 
     Topic  Count
1       0    308
2       1    107
3       2     83
5       3     75
21      4     69
16      5     66
4       6     57
25      7     55
9       8     51
22      9     48 

Topic 0 has the following top words: 
 [('nah', 3.619248420447182), ('know', 1.4617849136528802), ('', 1e-05), ('', 1e-05), ('', 1e-05), ('', 1e-05), ('', 1e-05), ('', 1e-05), ('', 1e-05), ('', 1e-05)] 

Topic 1 has the following top words: 
 [('rude', 0.03431141872944007), ('staff', 0.02840506990443558), ('member', 0.02520934244959189), ('manager', 0.023528607119674173), ('gym', 0.02139922651380406), ('personal', 0.018953258511792522), ('aggressive', 0.016668275528548066), ('reviews', 0.016001951029404715), ('puregym', 0.015331679815607645), ('one', 0.01292573313094035)]
In [62]:
neg_only[["Location Name", "emotion"]].value_counts().head(3)
Out[62]:
count
Location Name emotion
London Stratford anger 37
Leicester Walnut Street anger 22
London Enfield anger 19

Findings¶

Conducting Emotion Analysis Focused on Negative Reviews for PureGym (bert-base-uncased-emotion)¶

  • Both datasets contained the same emotions in the same order, with the only difference being that sadness and joy swapped places based on distribution.
  • Anger was the overwhelming emotion in both datasets, with over 1,000 occurrences in each.

Topic Modelling Where the Emotion is Anger¶

  • Looking at the intertopic distance map, we identified 6 distinct topics, with other topics clustered in or around these main topics.
  • The top topic words remain very similar to previous Bertopic results, with topics centered around:
    • Rude staff.
    • Issues with the toilets.
    • Access problems.
  • The similarity matrix between topics indicates that Bertopic has done a good job in separating the topics, given how light the colors are.
  • The top three locations where anger is the dominant emotion are:
    • London Stratford – 37 occurrences.
    • Leicester Walnut Street – 22 occurrences.
    • London Enfield – 19 occurrences.

Using a large language model from Hugging Face (PHI)¶

In [88]:
phi_neg = pd.concat([google_neg[["Comment", "Location Name"]],  trustpilot_neg[["Comment", "Location Name"]]], ignore_index = True)
In [89]:
torch.random.manual_seed(0)
Out[89]:
<torch._C.Generator at 0x7e26d6572930>
In [90]:
model_path = "microsoft/Phi-4-mini-instruct"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="cuda",
    torch_dtype="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_path)

config.json:   0%|          | 0.00/2.50k [00:00<?, ?B/s]
configuration_phi3.py:   0%|          | 0.00/10.9k [00:00<?, ?B/s]
A new version of the following files was downloaded from https://huggingface.co/microsoft/Phi-4-mini-instruct:
- configuration_phi3.py
. Make sure to double-check they do not contain any added malicious code. To avoid downloading new versions of the code file, you can pin a revision.

modeling_phi3.py:   0%|          | 0.00/54.3k [00:00<?, ?B/s]
A new version of the following files was downloaded from https://huggingface.co/microsoft/Phi-4-mini-instruct:
- modeling_phi3.py
. Make sure to double-check they do not contain any added malicious code. To avoid downloading new versions of the code file, you can pin a revision.

model.safetensors.index.json:   0%|          | 0.00/16.3k [00:00<?, ?B/s]
Fetching 2 files:   0%|          | 0/2 [00:00<?, ?it/s]
model-00001-of-00002.safetensors:   0%|          | 0.00/4.90G [00:00<?, ?B/s]
model-00002-of-00002.safetensors:   0%|          | 0.00/2.77G [00:00<?, ?B/s]
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
generation_config.json:   0%|          | 0.00/168 [00:00<?, ?B/s]
tokenizer_config.json:   0%|          | 0.00/2.93k [00:00<?, ?B/s]
vocab.json:   0%|          | 0.00/3.91M [00:00<?, ?B/s]
merges.txt:   0%|          | 0.00/2.42M [00:00<?, ?B/s]
tokenizer.json:   0%|          | 0.00/15.5M [00:00<?, ?B/s]
added_tokens.json:   0%|          | 0.00/249 [00:00<?, ?B/s]
special_tokens_map.json:   0%|          | 0.00/587 [00:00<?, ?B/s]
In [91]:
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)

Device set to use cuda
In [92]:
generation_args = {
    "max_new_tokens": 500,
    "return_full_text": False,
    "temperature": 0.0,
    "do_sample": False,
}
In [93]:
def phi_model(text):

    messages = [
        {"role": "system", "content": "You work as a data analyst insights guru for a top gym in the UK and you want to find topics for improvements from customer reviews. You must return exactly 3 key topics as strings, strictly formatted as 'topic 1', 'topic 2', 'topic 3'."},
        {"role": "user", "content": "In the following customer review interaction, pick out exactly 3 main topics and return them as a valid Python list: Broken equipment, overcrowded, dirty locker rooms, and unhelpful staff. Look elsewhere!."},
        {"role": "assistant", "content": "'Broken equipment', 'overcrowded', 'unhelpful staff'"},
        {"role": "user", "content": f"In the following customer review interaction, pick out exactly 3 main topics and return only those 3 as a string: {text}"},
        ]
    output = pipe(messages, **generation_args)[0]['generated_text'].replace("'", '"')

    return output
In [94]:
phi_neg["topics"] = phi_neg["Comment"].apply(phi_model)
In [95]:
phi_neg.to_csv("phi_model.csv", index=False)
In [104]:
model_phi = BERTopic(verbose=True)
topics_phi, probabilities_phi = model_phi.fit_transform(phi_neg['topics'].tolist())

2025-04-01 05:20:24,451 - BERTopic - Embedding - Transforming documents to embeddings.

Batches:   0%|          | 0/198 [00:00<?, ?it/s]
2025-04-01 05:20:26,769 - BERTopic - Embedding - Completed ✓
2025-04-01 05:20:26,770 - BERTopic - Dimensionality - Fitting the dimensionality reduction algorithm
2025-04-01 05:20:41,070 - BERTopic - Dimensionality - Completed ✓
2025-04-01 05:20:41,072 - BERTopic - Cluster - Start clustering the reduced embeddings
2025-04-01 05:20:41,312 - BERTopic - Cluster - Completed ✓
2025-04-01 05:20:41,316 - BERTopic - Representation - Fine-tuning topics using representation models.
2025-04-01 05:20:41,456 - BERTopic - Representation - Completed ✓
In [105]:
bert_visuals(model_phi)

Total number of topics: 101
Outliers:
    Topic  Count
1     -1   2247 

Top topics along with their document frequencies: 
     Topic  Count
2       0    373
7       1    199
11      2    182
23      3    144
3       4    141
36      5    139
26      6     97
49      7     93
38      8     88
6       9     80 

Topic 0 has the following top words: 
 [('unhelpful', 0.12300326579982984), ('overcrowded', 0.10022180631271357), ('broken', 0.09263096587012651), ('staff', 0.0736963666938909), ('equipment', 0.038364977494423844), ('away', 0.0023281513814752044), ('overcrowdedstaff', 0.002150786506725446), ('unavailability', 0.0019639597145591097), ('properly', 0.0017536863381075911), ('put', 0.0017204835562337562)] 

Topic 1 has the following top words: 
 [('gym', 0.023842928637476644), ('experience', 0.020038469076131472), ('poor', 0.01742995207675169), ('workout', 0.01630277849121239), ('worse', 0.016142740962067064), ('inadequate', 0.01405116986141448), ('faulty', 0.01307668117884842), ('terrible', 0.012086116101734209), ('worst', 0.011972406023336333), ('equipment', 0.011579355751910698)]
In [106]:
def phi_improvements(text):

    messages_1 = [
        {"role": "system", "content": "You work as a data analyst insights guru for a top gym in the UK and you want to compress a given list into collated topics. You should return these in a numbered list as business insights that can be used to improve the business"},
        {"role": "user", "content": "In the following list containing the main extracted topics from customer reviews, group or compress the topics and return them with actionable insights in a numbered list: [Poor service experience','Comparison of gym and service','Overall satisfaction','Service improvement needed','bad quality','Cleanliness of establishment','Neutral dining experience','Potential health and safety concerns','Overall satisfaction',]"},
        {"role": "assistant", "content": "1. Cleanliness: Increase cleaning frequency and enforce hygiene rules. \n 2. Equipment maintenance: Repair faulty machines promptly and schedule regular maintenance. \n ..."},
        {"role": "user", "content": f"In the following list containing the main extracted topics from customer reviews, group or compress the topics and return them with actionable insights in a numbered list: {text}"},
        ]
    output = pipe(messages_1, **generation_args)[0]['generated_text']

    return output
In [115]:
split_topic = np.array_split( phi_neg['topics'].tolist(), 2)
In [117]:
for i in split_topic:
  print(phi_improvements(i))

1. Overcrowding: Implement a reservation system to manage gym capacity and reduce overcrowding.
2. Customer service: Train staff to improve customer service and professionalism.
3. Equipment maintenance: Regularly check and repair gym equipment to prevent breakdowns.
4. Hygiene: Increase cleaning frequency and enforce hygiene rules to maintain clean facilities.
5. Staff behavior: Provide training to staff to improve their behavior and professionalism.
6. Gym atmosphere: Create a welcoming and motivating gym atmosphere to enhance the customer experience.
7. App functionality: Improve the app's functionality to reduce wasted time and address day pass and turnstile access issues.
1. Temperature control: Install adjustable thermostats in showers and ensure consistent water temperature.
2. Membership fees: Review pricing structure and consider offering tiered membership options.
3. Outdoor facilities: Regularly maintain and upgrade outdoor amenities to enhance user experience.
4. Long wait times: Implement a more efficient booking system and increase staff during peak hours.
5. Lack of customer service: Provide regular training to staff on customer service skills and communication.
6. Improve communication: Establish clear communication channels and provide regular updates to members.
7. Hygiene cleanliness standards: Enforce strict hygiene protocols and conduct regular inspections.
8. Changing room maintenance: Regularly clean and maintain changing rooms to ensure a pleasant experience.
9. Staff management: Invest in staff training and development to improve overall service quality.
10. Shower temperature: Install adjustable thermostats in showers and ensure consistent water temperature.
11. Cost saving: Review and optimize operational costs to offer more competitive membership fees.
12. Millhouses: Regularly maintain and upgrade facilities to enhance user experience.
13. Overcrowded: Implement a booking system to manage peak times and prevent overcrowding.
14. Dirty locker rooms: Regularly clean and maintain locker rooms to ensure a pleasant experience.
15. Unhelpful staff: Provide regular training to staff on customer service skills and communication.

Findings¶

BERTopic Modelling¶

(Three new topics have been created for each negative review using the Phi model. These new topics were then passed to BERTopic for further topic modeling.)

  • The total number of topics has dramatically increased to over 100.
  • No new insights have been gained, as the top topics still highlight the same key issues:
    • Staff interactions
    • Air conditioning problems
    • Cleanliness of the gym
    • Issues with gym classes

Recommendations¶

(The newly created topics were passed back to the Phi model to generate actionable insights.)

  • The dataset was split into two and passed into the Phi model to avoid errors.
  • Due to this split, some recommendations were repeated—these duplicates have been ignored.

Final Recommendations:¶

  1. Overcrowding – Implement a reservation system to manage gym capacity and reduce congestion.
  2. Customer service – Train staff to improve professionalism and service quality.
  3. Equipment maintenance – Regularly inspect and repair gym equipment to prevent breakdowns.
  4. Hygiene – Increase cleaning frequency and enforce hygiene rules for cleaner facilities.
  5. Staff behavior – Provide additional training to improve staff interactions with customers.
  6. Gym atmosphere – Foster a welcoming and motivating environment for a better user experience.
  7. App functionality – Improve the app to enhance usability and resolve issues with day passes and turnstile access.
  8. Temperature control – Install adjustable thermostats in showers to ensure consistent water temperature.
  9. Membership fees – Reevaluate pricing and consider offering tiered membership options.
  10. Outdoor facilities – Regularly maintain and upgrade outdoor amenities for better usability.
  11. Long wait times – Introduce a more efficient booking system and increase staff presence during peak hours.
  12. Improve communication – Establish clear communication channels and provide regular updates to members.
  13. Changing room maintenance – Regularly clean and maintain changing rooms to improve the user experience.
  14. Millhouses – Regularly maintain and upgrade facilities to enhance the overall experience.

Using Gensim for topic Modelling¶

In [109]:
comments_neg = phi_neg["Comment"]
In [110]:
tokenized_docs = [word_tokenize(comment) for comment in comments_neg]
dictionary = corpora.Dictionary(tokenized_docs)
dictionary.filter_extremes(no_below=2, no_above=0.5)
In [111]:
corpus = [dictionary.doc2bow(doc) for doc in tokenized_docs]

num_topics = 10
passes = 20
In [112]:
lda_model = LdaModel(corpus=corpus, num_topics=num_topics, id2word=dictionary, passes=passes)
In [113]:
for idx, topic in lda_model.print_topics(-1):
    print("Topic: {} \nWords: {}".format(idx, topic))

Topic: 0 
Words: 0.027*"staff" + 0.017*"manager" + 0.015*"rude" + 0.013*"member" + 0.011*"one" + 0.011*"members" + 0.009*"said" + 0.008*"personal" + 0.008*"left" + 0.008*"customer"
Topic: 1 
Words: 0.046*"water" + 0.030*"showers" + 0.025*"cold" + 0.024*"broken" + 0.020*"shower" + 0.017*"months" + 0.015*"hot" + 0.015*"machine" + 0.014*"machines" + 0.013*"order"
Topic: 2 
Words: 0.018*"membership" + 0.017*"email" + 0.016*"day" + 0.015*"pin" + 0.015*"get" + 0.014*"code" + 0.014*"fee" + 0.013*"joining" + 0.013*"pass" + 0.011*"account"
Topic: 3 
Words: 0.023*"parking" + 0.014*"pure" + 0.013*"back" + 0.011*"puregym" + 0.010*"park" + 0.009*"one" + 0.009*"free" + 0.009*"car" + 0.007*"date" + 0.007*"would"
Topic: 4 
Words: 0.015*"staff" + 0.013*"members" + 0.013*"people" + 0.011*"management" + 0.010*"like" + 0.010*"puregym" + 0.008*"member" + 0.008*"machine" + 0.007*"time" + 0.007*"one"
Topic: 5 
Words: 0.030*"membership" + 0.013*"cancel" + 0.012*"pay" + 0.011*"month" + 0.011*"would" + 0.011*"use" + 0.010*"gyms" + 0.009*"get" + 0.009*"go" + 0.008*"puregym"
Topic: 6 
Words: 0.049*"air" + 0.019*"conditioning" + 0.016*"music" + 0.016*"hot" + 0.015*"con" + 0.015*"temperature" + 0.009*"classes" + 0.009*"loud" + 0.009*"pt" + 0.009*"working"
Topic: 7 
Words: 0.039*"classes" + 0.039*"class" + 0.017*"one" + 0.014*"time" + 0.013*"get" + 0.010*"minutes" + 0.010*"booked" + 0.009*"instructor" + 0.009*"first" + 0.009*"work"
Topic: 8 
Words: 0.018*"equipment" + 0.016*"changing" + 0.015*"machines" + 0.010*"dirty" + 0.010*"toilets" + 0.010*"always" + 0.010*"place" + 0.009*"staff" + 0.009*"cleaning" + 0.009*"weights"
Topic: 9 
Words: 0.052*"equipment" + 0.025*"people" + 0.024*"use" + 0.019*"busy" + 0.018*"gyms" + 0.018*"machines" + 0.013*"many" + 0.012*"time" + 0.011*"go" + 0.011*"get"
In [114]:
pyLDAvis.enable_notebook()
vis = gensimvis.prepare(lda_model, corpus, dictionary)
pyLDAvis.display(vis)
Out[114]:

Findings¶

Using Gensim for Topic Modeling (Max Topics Set to 10)¶

  • Using Gensim does not provide any new insights.
  • The identified topics remain consistent with previous BERTopic results.
  • The key recurring issues are:
    • Staff interactions
    • Air conditioning problems
    • Cleanliness of the gym
    • Issues with gym classes