In this project, we use transformers to predict product ratings and to determine the tone of the review based on preprocessed customer text reviews.
- ClothingReviewsClassification.ipynb : Jupyter Notebook with text preprocessing and model training
- /binary : Binary classification model's files
- /multiclass : Multiclass classification model's files
We use Women’s Clothing E-Commerce dataset revolving around the reviews written by customers
This dataset includes 23486 rows and 10 feature variables. Each row corresponds to a customer review, and includes the variables:
| Feature | Type | Description |
|---|---|---|
Clothing ID |
Integer | Variable that refers to the specific piece being reviewed. |
Age |
Integer | Variable of the reviewers age. |
Title |
String | Variable for the title of the review. |
Review Text |
String | Variable for the review body. |
Rating |
Integer | Variable for the product score granted by the customer from 1 Worst, to 5 Best. |
Recommended IND |
Integer | Variable stating where the customer recommends the product where 1 is recommended, 0 is not recommended. |
Positive Feedback Count |
Integer | Variable documenting the number of other customers who found this review positive. |
Division Name |
String | Categorical name of the product high level division. |
Department Name |
String | Categorical name of the product department name. |
Class Name |
String | Categorical name of the product class name. |
In the process of text preprocessing, we employ methods such as tokenization, POS-tagging, and lemmatization. All these operations are carried out using the nltk library. Additionally, we remove stop words and convert the text to lowercase.
We use bert-base-uncased to binary classification for predicting whether the review is positive or negative.
f1 score: 0.8507015684469733 number of epochs: 10
The classifier does its job well. For successful classification, it is necessary to write a review using a sufficient number of explicit evaluative words.
We use distilroberta-base to multiclass classification for predicting the rating of the product.
f1 score: 0.6033866825602262 number of epochs: 20
Overall accuracy for each class:
- Class 0: 0.24
- Class 1: 0.31
- Class 2: 0.34
- Class 3: 0.35
- Class 4: 0.81
The classifier confuses classes that are close in value, for example: class 2 with 1 and 3. The classifier distinguishes classes well at a distance, for example: class 4 is little confused with class 0. Most likely, this result was due to an unbalanced dataset.
Maxim Ivanov - GitHub, Telegram
This project was completed as part of the "Практический Deep Learning" course offered by AI Education.
This project is licensed under the MIT license. For more information, see the LICENSE file.