Skip to content

vaibhawkhemka/Product-Categorization

BranchesTags

Repository files navigation

Product-Categorization

Observations and analysis: A brand can have multiple products like laptops and mobiles - So brand can't be only feature to determine There are brand names like 0, 168, NA etc.. Title is summarized texts containing brandname as well. So this feature is good.

Data Cleaning and Processing

Data_Clean.csv - Pre-Processed text with title and thumbnail Stop words, punctuations, lower case etc.

Topic Modelling

LDA Sklearn

sklearn - CountVectorizer as features. LDA - for Dimensionality Reduction

features: title (clean) and countervectorizer . Taking description can also improve the performance due to addition of lot of contexts.

Results: topic distribution after clustering

Topic Index Count
0 7433
1 6360
2 19780
3 9971
4 28693
5 11064

PylDAvis: LDA sklearn

Above shows clear distinction of 6 clusters which is good topic model.

Output: Topic: 0
bag, laptop, type, warranty, windows
Topic: 1
flip, phone, card, wallet, mobile
Topic: 2
compatible, watch, galaxy, leather, samsung
Topic: 3
iphone, watch, smart, women, bluetooth
Topic: 4
case, cover, charger, pro, charging
Topic: 5
vivo, inch, display, laptop, intel \

By above it can be seen clearly that

### Labelling Results - For LDA
    map_to_output = {0:"Laptop",1:"Others",2:"Printers",3:"Smartwatch", 4:"Mobile",5:"Desktop"}

Improvements

Can use tfidf features as well to improve the performance

Sentence tranformers with Kmeans Clustering

Sentence tranformers for creating doc embeddings
Kmeans for creating clusters

features: title (clean) is taken as a feature. Taking description can also improve the performance due to addition of lot of contexts.

Result: topic distribution after clustering \

Topic Index Count
1 11046
2 15911
3 23072
4 16569
5 9279
6 7424

WordClouds:
cluster 1 cluster 1

cluster 2 cluster 2

cluster 3 cluster 3

cluster 4 cluster 4

cluster 5 cluster 5

cluster 6 cluster 6

output:

### Labelling Results - For Kmeans
    map_to_output = {0:"Mobile",1:"Printer",2:"Desktop",3:"Smartwatch", 4:"Others",5:"Laptop"}

Thus Both of the above is good a topic model having clear distinction.

Improvements

USE image as well for creating label using VIT tranformer
Can take description as well for labelling

Product Multiclassification using BERT model

Experiment 1 - (LDA Data, Trained with 50% of labelled data)

Evaluation Metric: F1-Score as data is imbalanced

Training:\ Epoch 1
Training loss: 0.20996832987158887
Validation loss: 0.31821184134476377
F1 Score (Weighted): 0.9020082201038174 \

Epoch 2
Training loss: 0.2120044876928212 \ Validation loss: 0.28195449785616633
F1 Score (Weighted): 0.9074309908181265
... ...

Results and evaluation on Validation Data:
Class: 0 Accuracy: 454/536

Class: 1 Accuracy: 384/459

Class: 2 Accuracy: 1305/1431

Class: 3 Accuracy: 677/714

Class: 4 Accuracy: 1967/2060

Class: 5 Accuracy: 719/800

F1-SCORE: 90%

Experiment 2 - (Kmeans Data, Trained with 50% of labelled data)

Data classes distribution: \

Label Data Type Train Val
0 train 4505 795
0 val
1 train 6548 1156
1 val
2 train 9439 1665
2 val
3 train 6739 1189
3 val
4 train 3761 664
4 val
5 train 3008 531
5 val

Evaluation Metric: Imbalanced data can handled by choosing correct evaluation metric which is Precision, Recall and F-Score

Training:
Epoch 1
Training loss: 0.18195189569042058
Validation loss: 0.07974208040073752
F1 Score (Weighted): 0.9773530337741395 \

Epoch 2
Training loss: 0.056020494600515267
Validation loss: 0.07013297464792675
F1 Score (Weighted): 0.9810772920329817 \

Epoch 3
Training loss: 0.03608494108226956
Validation loss: 0.07039217839755633
F1 Score (Weighted): 0.9826290609313495
... ...

Results and evaluation:

Class: 0 Accuracy: 785/795

Class: 1 Accuracy: 1140/1156

Class: 2 Accuracy: 1651/1665

Class: 3 Accuracy: 1162/1189

Class: 4 Accuracy: 631/664

Class: 5 Accuracy: 495/531

F1-SCORE: 97%

Improvements:
Might be overfitted but can be tackled by dropout and regularization - Didn't get time to fix that but can be improved in next iteration Increase the data for training

Usage

API Service: running the server

uvicorn main:app --host 0.0.0.0 --port 8000

Change the data present in test.py

data = '{"title": "Xiaomi 11i 5G Hypercharge (Stealth Black, 6GB RAM, 128GB Storage)", "thumbnail": "https://m.media-amazon.com/images/I/71BQ8Kjt29L._SL1500_.jpg"}'

for testing ,run test.py with passed data json

python test.py

for choosing model - BERT_LDA or BERT_Kmeans

model.load_state_dict(torch.load('checkpoints/finetuned_BERT_Kmeans_epoch.model', map_location=torch.device('cpu')))

for mapping

### Labelling Results - For Kmeans
    map_to_output = {0:"Mobile",1:"Printer",2:"Desktop",3:"Smartwatch", 4:"Others",5:"Laptop"}

Final Results

BERT_LDA

C:\Users\vaibh\OneDrive\Desktop\Product-Categorization>python test.py
Request successful!
{'title': 'Xiaomi 11i 5G Hypercharge (Stealth Black, 6GB RAM, 128GB Storage)', 'Thumbnail': 'https://m.media-amazon.com/images/I/71BQ8Kjt29L._SL1500_.jpg', 'product_type': 'Mobile', 'Product_Type_conf_score': 0.7956177592277527, 'Other_Possible_Product_Type': {'Laptop': 0.015701234340667725, 'Others': 0.001399507513269782, 'Printers': 0.0022549154236912727, 'Smartwatch': 0.001993010751903057, 'Mobile': 0.7956177592277527, 'Desktop': 0.1830335259437561}}

BERT_Kmeans

Request successful!
{'title': 'Xiaomi 11i 5G Hypercharge (Stealth Black, 6GB RAM, 128GB Storage)', 'Thumbnail': 'https://m.media-amazon.com/images/I/71BQ8Kjt29L._SL1500_.jpg', 'product_type': 'Mobile', 'Product_Type_conf_score': 0.9942101240158081, 'Other_Possible_Product_Type': {'Mobile': 0.9942101240158081, 'Printer': 0.0004031399730592966, 'Desktop': 0.0006456150440499187, 'Smartwatch': 0.0011665192432701588, 'Others': 0.0009345235885120928, 'Laptop': 0.0026401979848742485}}

NER - Attribute Tagger

MOBILE

Results: =========================== Initializing pipeline =========================== ✔ Initialized pipeline

============================= Training pipeline ============================= ℹ Pipeline: ['tok2vec', 'ner'] ℹ Initial learn rate: 0.001

Epoch Step LOSS TOK2VEC LOSS NER ENTS_F ENTS_P ENTS_R SCORE
0 0 0.00 106.69 0.00 0.00 0.00 0.00
0 200 414.20 4348.63 44.95 58.20 36.61 0.45
0 400 328.51 2794.96 41.46 47.00 37.09 0.41
0 600 58.27 3250.15 41.52 45.59 38.13 0.42
0 800 55.10 3031.42 52.04 49.57 54.78 0.52
1 1000 60.28 3333.52 39.04 36.83 41.53 0.39
1 1200 94.13 4071.16 54.20 50.66 58.28 0.54
2 1400 82.30 4359.56 52.53 49.46 56.01 0.53
2 1600 89.31 4853.72 60.34 55.03 66.79 0.60
3 1800 118.10 5541.28 52.31 50.13 54.68 0.52
3 2000 165.50 6147.18 54.29 51.17 57.81 0.54
4 2200 140.11 7054.81 59.93 58.78 61.12 0.60
5 2400 301.09 7902.63 61.24 57.01 66.13 0.61
... ... ... ... ... ... ... ...
13 3800 312.74 9729.62 60.84 56.06 66.51 0.61
15 4000 361.47 9526.00 57.64 58.95 56.39 0.58

testing: text = "EL D68 (Green, 32 GB) 3 GB RAM ['3 GB RAM | 32 GB ROM | Expandable Upto 128 GB', '15.46 cm (6.088 inch) Display', '13MP Rear Camera | 8MP Front Camera', '4000 mAh Battery', 'Quad-Core Processor']"

Green ->>>> COLOR
32 GB ->>>> STORAGE
3 GB RAM ->>>> RAM
3 GB RAM ->>>> RAM
32 GB ROM ->>>> STORAGE
Expandable Upto 128 GB ->>>> EXPANDABLE_STORAGE
15.46 cm (6.088 inch) ->>>> SCREEN_SIZE
13MP Rear Camera ->>>> BACK_CAMERA
8MP Front Camera ->>>> FRONT_CAMERA
4000 mAh Battery ->>>> BATTERY_CAPACITY
Quad-Core Processor ->>>> PROCESSOR_CORE \

Final Output: ![alt text](/Training Data/Mobile/final_output_mobile.png) And final result stored in raining Data/Mobile/Final_Result_Mobile.csv

### Testing
def test(text):
  # Load the trained spaCy NER model from the specified path
  nlp = spacy.load("Training Data/Mobile/output/model-best")

  #text = "EL D68 (Green, 32 GB) 3 GB RAM ['3 GB RAM | 32 GB ROM | Expandable Upto 128 GB', '15.46 cm (6.088 inch) Display', '13MP Rear Camera | 8MP Front Camera', '4000 mAh Battery', 'Quad-Core Processor']"
  docs = nlp(text)
  attributes = {}
  for ent in docs.ents:
    # Print the recognized text and its corresponding label
    attributes[ent.label_] = ent.text
  return attributes

LAPTOP

Results: =========================== Initializing pipeline =========================== ✔ Initialized pipeline

============================= Training pipeline ============================= ℹ Pipeline: ['tok2vec', 'ner'] ℹ Initial learn rate: 0.001

Epoch Step LOSS TOK2VEC LOSS NER ENTS_F ENTS_P ENTS_R SCORE
0 0 0.00 40.36 0.00 0.00 0.00 0.00
4 200 88.38 1605.06 48.15 43.33 54.17 0.48
8 400 23.58 439.22 59.26 53.33 66.67 0.59
14 600 104.49 289.68 61.90 72.22 54.17 0.62
20 800 39.03 196.97 65.22 68.18 62.50 0.65
27 1000 48.17 170.84 60.00 75.00 50.00 0.60
34 1200 31.62 165.31 69.77 78.95 62.50 0.70
42 1400 65.98 204.34 69.39 68.00 70.83 0.69
53 1600 36.93 173.93 72.73 80.00 66.67 0.73
64 1800 118.11 220.46 71.11 76.19 66.67 0.71
79 2000 75.25 231.58 76.60 78.26 75.00 0.77
97 2200 40.05 248.59 69.23 64.29 75.00 0.69
120 2400 29.12 277.95 66.67 66.67 66.67 0.67
148 2600 109.15 354.39 75.00 75.00 75.00 0.75
... ... ... ... ... ... ... ...
375 4200 193.09 403.85 73.47 72.00 75.00 0.73
404 4400 20.85 320.89 72.00 69.23 75.00 0.72

✔ Saved pipeline to output directory Training Data\Laptop\output\model-last ✔ Saved pipeline to output directory

Testing: text "ASUS FX506LHB-HN355W i5 10300H/ GTX1650- 4GB/ 8G/ 512G SSD/ 15.6 FHD-144hz/ Backlit KB- 1 Zone RGB/ 48Whr/ Win 11/ / / McAfee(1 Year)/ 1B-Black Plastic" \

ASUS FX506LHB-HN355W i5 ->>>> PROCESSOR
8G/ 512 ->>>> RAM \

Final Output: ![alt text](/Training Data/Laptop/final_output_laptop.png) And final result stored in raining Data/Laptop/Final_Result_Laptop.csv

Improvements

More data required - Span is giving NONE - this producing only around 100 data points And remaining 350 is not included which is still less
learining rate and epochs can be improved

About

No description, website, or topics provided.

Resources

Readme

License

MIT license
Activity

Stars

0 stars

Watchers

1 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages