In this competition, to simulate the situation of an operating online recommender system, we will sequentially feed the test input in ascending order of the timestamp to test your model, and we also allow you to update your model during testing. This means your model will receive ground truth data from the testing set. To prevent disputes, we will ask you to upload your saved model to your Google Drive and submit the share link to the eeclass system, and we will then download and evaluate your model with an evaluation script. To make sure the evaluation process can be successfully executed, your model should follow the template below. Your model should inherit the tf.keras.Model class and at least implement the three methods: call(), eval_predict_onestep(), and eval_update_onestep(). Note that the @tf.function decorators on these methods are necessary for model.save() to record and save these methods.
For all evaluation rounds, we will ask you to upload your saved model to your Google Drive and submit the share link to the eeclass system, and we will then download and evaluate your model with an evaluation script. A download link to a sample evaluation script that uses part of the training set is provided here. Please make sure your saved model can be loaded and evaluated by the sample evaluation script. Note that the evaluation procedure will not provide any features of the users and the movies. If you want to use them during evaluation, you should store them in your model while training.
For each evaluation round, we will open an eeclass assignment for submission. You should upload your model to your Google Drive and submit the share link to the corresponding assignment. The first two evaluation rounds are optional. However, all groups should submit your model in the final evaluation round. Only the model performance in the final evaluation round will account for your score. 153554b96e