shanmukh05/Amazon-ML-Challenge

Our team "Elementals" approach to Amazon ML Challenge

Amazon-ML-Challenge

Team : Elementals

Prateek | Manish | Adhesh | Shanmukh

Details of Files

• amazon_ml_preprocessing.ipynb : code to preprocess text
• amazon_ml_translation_csv.ipynb : code to translate non-english text
• amazon_ml_mode.ipynb : code to create submission file from multiple submission files using mode technique
• amazon_ml_training.ipynb : code to train embeddings and predict on test embeddings
• amazon_ml_embeddings.ipynb : code to generate embeddings from csv file
• submission_top-score.csv : submission file with top score [Accuracy : 66.85]

Details of Competition

• Competition : Multi Class Text Classification
• Host : Hacker-Earth
• Metric : Accuracy
• Time of Competition : 2days:23hrs:59min
• Checkout competition here

Details of Data

• Key column – PRODUCT_ID
• Input features – TITLE, DESCRIPTION, BULLET_POINTS, BRAND
• Target column – BROWSE_NODE_ID
• Train dataset size – 2,903,024
• Number of classes in Train – 9,919
• Overall Test dataset size – 110,775

Data Preprocessing

Libraries used:

• re
• langdetect
• deep-translator

Steps followed:

• Removed special characters and emojis using re.
• Translated non-english text to english using langdetect and deep-translator.
• Removed stop-words.
• Decontracted some of the words.

Our Approach

Libraries used:

• Sentence_transoformer
• RAPIDS

Steps followed:

• First the text is converted to embeddings using pre-trained models such as paraphrase-mpnet-base-v2 , paraphrase-MiniLM-L6-v2 paraphrase-MiniLM-L3-v2
• Dimension of Embeddings : 384
• Embeddings of training data are sent into KNNClassifier present in CuML library
• Then the trained KNNClassifier is used to predict on test embeddings.
• We also used mode technique i.e. using most frequently predicted label obtained from different experiments
• Cross Validation is also used to train KNNClassifier

Experiments

• Used NearestNeighbour, SVM, RandomForest Classifier techniques but results are not better compared to KNNClassifier.
• Different size embeddings (384,768)
• Combined TITLE,DESCRIPTION and TITLE,DESCRIPTION, BULLET_POINTS and TITLE, DESCRIPTION, BULLET_POINTS