leffff / ai-ijc

1st place solution to AI IJC Customer Service task

AI-IJC

Our top 1 solution to AI IJC competition

Team round included several tasks:

1. Prediction aggressive driving among taxi drivers using following data types: taxi route, customer comment about the ride, other table data about the order.
2. Extracting a reason for aggressive driving from comments
3. Detecting parts of taxi routs with aggressive driving
   
   

Final round task:

• Building a clustering system for taxi drivers, which in future could be used as a reward system
  
  

About the data:

• There were only two labels in the dataset: 0 which meant unaggressive driving and 1, which symbolized aggressive driving. The dataset was imbalanced with about 96% of 0 labels and only 4% of 1. These labels were only given for the first task.
  
• The dataset itself didn't have commnets for every ride, which made it hard to train big supervised models.
  However we had a whole dataset of unlabeled comments. Which we were to use.
  
  

Solution for the 1st task:

1. The best score was achieved with a text model sismetanin/xlm_roberta_base-ru-sentiment-rusentiment. However the stack also included XGBoost with Optuna and a GCN
2. We did EDA and cleaned short or senseless comments like "Да", "Нет", "." etc.
3. Implement UDA in order to utilize unlabeled samples. Original PyTorch UDA implementation
4. We used back translation to English as an augmentation of unlabeld data.
5. We chose Cross Entropy as supervised loss and KL Divergece as unsupervised loss

Solution for the 2nd task:

1. Delete senseless samples from text dataset
2. Use a QA model AlexKay/xlm-roberta-large-qa-multilingual-finedtuned-ru by asking it "как выражалось агрессивное вождение?"

Solution for 3rd task:

1. Training a GCN model for the task 1
2. Using the model to score driving agressiveness at every node individually

Solution to final round