昵称登录后显示

Neural Network Morphism: The project's objective is to modify structure of previously trained neural network models to a new structure to avoid training new networks from scratch every time. I investigated how least squares can be applied to increase the layer size, decrease the layer size, add a layer, remove a layer, or change activation function of the pre-trained neural networks, thus transferring knowledge to a new neural networks structure and allowing to start training the new network from an already good network (in term of cost or training error).

Air ticket prediction:
For each query data, 5 features are computed, the flight number(encoded by dummy variables), the minimum price so far, the maximum price so far, the query-to-departure (number of days between the first query date (09.11.2015 in our case) and departure date), the days-to-departure (number of days between the query and departure date), and current price.

Classification:
For the output, I set the data of which the price is the the minimum price from the query date to the departure date to be 1(namely Class 1 - to buy), otherwise, I set it to be 0(namely Class 2 - to wait).
• The classification method is to predict to buy or to wait with the input features.
• As a result, if the prediction is to buy, we buy the ticket, and we only buy the earliest one.

Regression:
For regression, as the output is the minimum price for each departure date and each flight. Our regression method is to predict the expected minimum price for a given departure date and given flight with the input features. As a result, if the current price is less than the expected minimum price, we predict to buy; otherwise, we predict to wait.