Enhancing Online Shopping Experience with Multi-Lingual Query-Based Product Ranking using BERT, BM25, and LGBMRanker.
The goal of this project is to develop a system that can rank products in an e-commerce environment based on their relevance to a given search query. This will enable e-commerce websites to deliver more relevant results to their customers, improving their search experience and potentially increasing sales.
The system takes as input a list of queries, each with a unique identifier. The queries can be considered as the search terms entered by users when looking for products on an e-commerce website. The system then processes the input data and generates a CSV file with the ranked products for each query.
To determine the relevance of a product to a given query, the system considers four degrees of relevance: Exacts, Substitutes, Compliments, and Irrelevants.
-
Exacts: These are the products that match the query exactly, meaning they have the same name, brand, and attributes as the query. These products should be ranked at the top of the list. -
Substitutes: These are products that have similar features to the query but may have different names or brands. These products should be ranked after the Exacts but still near the top of the list. -
Compliments: These are products that complement the query product and are often bought together with the query product. These products should be ranked after the Substitutes but still close to the top of the list. -
Irrelevants: These are products that are not related to the query and should be ranked at the bottom of the list.
By sorting the input data based on these four degrees of relevance, the system can generate a ranked list of products for each query, ensuring that the most relevant products are displayed first.
The project leverages multiple techniques to rank products based on multi-lingual queries, including three models: BERT, BM25, and LGBMRanker.
BERT : used as a cross encoder to evaluate the relevance of a query and product description. The query and product description are concatenated and fed into the BERT model, which outputs a score reflecting the relevance.
BM25 : used to evaluate the relevance of the query and product description. Like BERT, it takes the query and product description as inputs and then outputs a score.
LGBMRanker : takes the outputs from BERT and BM25, along with hand-crafted features, and uses them to rank the products based on the query. The LGBMRanker model combines the scores from BERT and BM25 with the hand-crafted features to create a final ranking of the products, which can be used to present the most relevant products to the user.
The input for this task is a CSV file with the following columns:
- query_id: a unique identifier for the query
- query: the text of the query
- query_locale: the locale of the query
- product_id: a unique identifier for the product
An example of the input file is shown below:
| query_id | query | query_locale | product_id | esci_label |
|---|---|---|---|---|
| Query_1 | "Query_1" | us | product_23 | exact |
| Query_2 | "Query_2" | us | product_24 | substitute |
This is a table of product catalog information. It is csv file that contains information about different products.
The columns are:
- product_id - a unique identifier for each product
- product_title - the title of the product
- product_description - a brief description of the product
- product_bullet_point - bullet points that highlight the key features of the product
- product_brand - the brand name of the product
- product_color_name - the color of the product
- product_locale - the locale where the product is available
Here is an example of a single row in the table:
| product_id | product_title | product_description | product_bullet_point | product_brand | product_color_name | product_locale |
|---|---|---|---|---|---|---|
| B075VXJ9VG | "BAZIC Pencil #2 HB Pencils, Latex Free Eraser" | Our goal is to provide each customer with long-lasting supplies at an affordable cost... | UN-SHARPENED #2 PREMIUM PENCILS. | BAZIC Products | Yellow | us |
The system will output a CSV file with the following columns:
- query_id: the unique identifier for the query
- product_id: the unique identifier for the product, ranked based on relevance
An example of the output file is shown below:
| query_id | product_id |
|---|---|
| Query_1 | product_50 |
| Query_1 | product_900 |
| Query_1 | product_80 |
| Query_2 | product_32 |
| Query_2 | product_987 |
| Query_2 | product_105 |
├── cli <- Entry code to the project
│ ├──learning.py <- Training code
│ └──prediction.py <- Prediction code
├── src <- Python package - scripts of the project
│ ├── configuration <- Setup configs
│ │ ├── resources <- configuration files
│ │ │ ├── data_structure.json <- data configuration file
│ │ │ ├── model.json <- model configuration file
│ │ │ └── path.json <- path configuration file
│ │ ├── app.py <- Environment configuration
│ │ └── data.py <- Data configuration
│ ├── models <- Modeling
│ │ ├── bert_model.py <- Bert model class (used in cross encoder)
│ │ ├── bm25_model.py <- BM25 model class
│ │ ├── cross_encoder_model.py <- Cross Encoder model class
│ │ ├── cross_encoder_dataset.py <- Cross Encoder Dataset class used to import data
│ │ └── cross_encoder_factory.py <- Cross Encoder Factory class used to get models
│ ├── pipeline <- Pipelines of the application
│ │ ├── catalogue_preprocessing_pipeline.py <- Pre-process the product catalogue
│ │ ├── create_features_pipeline.py <- Create features for final training
│ │ ├── data_preprocessing_pipeline.py <- Pre-porcess query product data
│ │ ├── get_data_pipeline.py <- Import data
│ │ ├── save_pipeline.py <- Save trained models
│ │ ├── train_bm25_pipeline.py <- Train BM25 model
│ │ ├── train_cross_encoder_pipeline.py <- Train cross encoder
│ │ └── train_ensemble_pipeline.py <- Train the ranking model (stacking of previous models)
│ └── utils <- Useful transversal functions
│ ├── constant.py <- Useful constant values
│ └── support.py <- System functions
│
│
├── .gitignore <- Files that should be ignored by git
├── README.md <- The top-level README of the repository
└── requirements.txt <- Python libraries used in the project
To get started, first make sure you have the necessary requirements installed. Run the following command in your terminal:
$ sh install.sh
Next, download the data for the project by running the following command:
$ sh download_data.sh
With these two simple steps, you're all set to start working on the project!
The cli/learning.py script trains a machine learning model on the provided data. The trained model will be saved to the specified location, ready for use in the query product ranking system.
To train a model, use the following command in your terminal:
$ python cli/learning.py \
--train_data [PATH_TO_TRAIN_DATA] \
--product_catalogue [PATH_TO_PRODUCT_CATALOGUE]
--model_save_dir [PATH_TO_SAVE_MODELS]
The following arguments must be passed to the script:
-
--train_data: The path to the training data file, in CSV format. -
--product_catalogue: The path to the product catalogue file, in CSV format. -
--model_save_dir: The directory where the trained models will be saved.
The script will output the training results to the console, including the final model's accuracy. The trained models will be saved to the specified model_save_dir for future use.
The cli/prediction.py script makes the prediction using a trained machine learning model and the provided test data. The predictions will be saved to the specified output location.
To make predictions, use the following command in your terminal:
$ python cli/prediction.py \
--test_data [PATH_TO_TEST_DATA] \
--product_catalogue [PATH_TO_PRODUCT_CATALOGUE]
--model_load_dir [PATH_TO_LOADED_MODEL]
--output [PATH_TO_OUTPUT]
The following arguments must be passed to the script:
-
--test_data: The path to the test data file, in CSV format. -
--product_catalogue: The path to the product catalogue file, in CSV format. This file should be the same as the one used for training. -
--model_load_dir: The directory where the trained model is located, in [Your preferred model format]. -
--output: The path to the output file where the predictions will be saved, in CSV format.
The script will output the predictions to the specified output file. The predictions will be in CSV format.