KKday

Personalized city recommendation on KKday homepage. 

Based on language , customers purchase history and recently action, we generate cities interesting to customers. 

This model increased click rate over 1%.

Kaggle

This is a time model. For each time, it has only 5 features (shop_id/ item_id/ Category_name/ shop_name/ Category_id), So I need generate new features. I add  time delay data and embedding the shop name and item category as new features. I also train model to get new features. In the end, I use Linear model to reduce features and build 3 models to get the final answer.

Project 1 : Image Captioning

Given a picture, It will generate a short description for this picture.

This model use a pre-trained InceptionV3 model for CNN encoder and extract its last hidden layer as an embedding. 

Use about 82K training data and 40K validation date and each picture has 5 captions.

Project 2 : AI robot in Telegram

If you ask a programing problem, chat robot will return a closest StackOverflow link for you.

In this model, we will classify the training data to language type and use facebookresearch/StarSpace for embedding questions. For each input, we will classify it and embedding it to vector then found related link.

Project 3 : OpenAI CartPole-v0

Use Monte Coral tree search. That is, we choose road by root scores, if meet the tree leaf, do propagation (add score to root).

We can build a tree by playing games, then tree will tell us how to playing game. 

Project 4 : Playing Game

Use actor-critic training.

Input 4-frame image, train a DNN to catch the image information and predict the probability for 12 reactions. Then we according the DNN result to make a reaction for this game.