Mobile App Classification

Model description

BigBird, is a sparse-attention based transformer which extends Transformer based models, such as BERT to much longer sequences. The model can handle input sequence of length up to 4,096 tokens.

The google/bigbird-roberta-base model is fine-tuned to classify an mobile app description into one of 6 play store categories. Trained on 9000 samples of English App Descriptions and associated categories of apps available in Google Play.

Fine-tuning

The model was fine-tuned for 5 epochs with a batch size of 16, a learning rate of 2e-05, and a maximum sequence length of 1024. Since this was a classification task, the model was trained with a cross-entropy loss function. The best evaluation f1 score achieved by the model was 0.8964259037209702, found after 4 epochs. The accuracy of the model on the test set was 0.8966.

How to use

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

tokenizer = AutoTokenizer.from_pretrained("nsi319/bigbird-roberta-base-finetuned-app")  
model = AutoModelForSequenceClassification.from_pretrained("nsi319/bigbird-roberta-base-finetuned-app")

classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)

classifier("From scores to signings, the ESPN App is here to keep you updated. Never miss another sporting moment with up-to-the-minute scores, latest news & a range of video content. Sign in and personalise the app to receive alerts for your teams and leagues. Wherever, whenever; the ESPN app keeps you connected.")

'''Output'''
[{'label': 'Sports', 'score': 0.9983325600624084}]

Limitations

Training data consists of apps from 6 play store categories namely Education, Entertainment, Productivity, Sports, News & Magazines and Photography.