This repository contains code and the Commercial Dataset needed for reproducing the results presented in the following paper: Mend The Learning Approach, Not the Data: Insights for Ranking E-Commerce Products .

Note: Updates to the repository coming soon.

 pip3 install jupyter 

If you are not familiar with running notebook, click here.

Download the treac_eval tool and Mercateo Dataset.

• CRM_Training
  • crm_training_clicks.ipynb: Run this jupyter notebook for training CRM model from AtB click logs [Reproducing results of Table 2 of the paper].
  • crm_training_orders.ipynb: Run this jupyter notebook for training CRM model from order logs [Reproducing results of Table 3 of paper].
  • crm_model.py: Keras implementation of CNN for short text pairs with counterfactual risk minimization (CRM) loss function.
• Cross_Entropy_Training
  • cross_entropy_training_clicks.ipnyb: Run this jupyter notebook for training CNN model with cross entropy loss [Reproducing results of Table 2 of paper].
  • cross_entropy_training_orders.ipynb: Run this jupyter notebook for training CNN model with cross entropy loss [Reproducing results of Table 3 of paper].
  • model_cross_entropy.py: Keras implementation of CNN for short text pairs with cross-entropy loss.
• LambdaMART Training
  • Download the binary file of RankLib tool from here.
  • We used latest binary 'RankLib-2.1-patched.jar' for our experiments.
  • Train LambdaMART model, for Graded Order Labels [Table 3], by running this command:
    

     java -jar RankLib-2.1-patched.jar -train LambdaMART_files/New_Graded_Order_TrainFile.csv 
      -test LambdaMART_files/New_Graded_Order_TestFile.csv -validate LambdaMART_files/
      New_Graded_Order_DevFile.csv -ranker 6 -metric2t NDCG@10 -metric2T NDCG@10 
      -save Model_LMART_Graded_Orders.txt 

  • In order to evaluate the saved model on other metrics {NDCG@5,P@5,P@10,RR,MAP}, run this command:
    

     java -jar RankLib-2.1-patched.jar -load Model_LMART_Graded_Orders.txt -test
        LambdaMART_files/New_Graded_Order_DevFile.csv -metric2T NDCG@5 

• Affect of DNN architecture
  • For reproducing the results in Table 4 of the paper, refer to MatchZoo.
  • MatchZoo has comprehensive documentation on the dependencies and how to run the models.
  • For a fair comparison with S-CNN model, we modified the models in MatchZoo and added a fully connected layer before the last layer. This layer is added so that we can utilize the dense features.

• python 2.7 or higher
• numpy
• keras
• trec_eval

For more information about the dataset and model please refer to the paper. For any questions/bugs, you can report issue here.