joaopaulopmaues / Redes-Neurais

Trabalho de redes neurais

#First of all install and unzip ml-25m from ml-25m.zip from https://files.grouplens.org/datasets/movielens/ # Wide and Deep Learning For Recommender Systems This is my PyTorch implementation of the paper [Wide & Deep Learning for Recommender Systems](https://arxiv.org/pdf/1606.07792.pdf) by Google Inc. (2016). This model has been productionized and evaluated on [Google Play](https://play.google.com/store?hl=en_US), a commercial mobile app store with over one billion active users and over one million apps.

## Scripts * [data.py](https://github.com/khanhnamle1994/transfer-rec/blob/master/Multilayer-Perceptron-Experiments/Wide-and-Deep-PyTorch/data.py): This is the data processing script. * [layer.py](https://github.com/khanhnamle1994/transfer-rec/blob/master/Multilayer-Perceptron-Experiments/Wide-and-Deep-PyTorch/layer.py): This is the utility script that defines layer classes used in the Wide and Deep model. * [Wide_Deep.py](https://github.com/khanhnamle1994/transfer-rec/blob/master/Multilayer-Perceptron-Experiments/Wide-and-Deep-PyTorch/Wide_Deep.py): This is the model script that defines the Wide and Deep model. * [main.py](https://github.com/khanhnamle1994/transfer-rec/blob/master/Multilayer-Perceptron-Experiments/Wide-and-Deep-PyTorch/main.py): This is the main training script. You can simply run `python3 main.py` to execute it. ## Requirements ``` PyTorch 1.3 Python 3.6 Numpy Pandas ``` ## Dataset You can download the MovieLens-1M dataset from [this folder](https://github.com/khanhnamle1994/transfer-rec/tree/master/ml-1m).

## Results Here are the model hyper-parameters chosen: - Number of Dense Embedding Dimensions used in the Deep Component = 16 - Number of Hidden Layers used in the Deep Component = 16 - Activation Function = Sigmoid - Learning Rate = 0.001 - Batch Size = 512 - Weight Decay = 0.000001 - Optimizer Method = Adam - Dropout Rate = 0.5 After being trained for 100 epochs, the model achieves **validation AUC = 0.7995** and **test AUC = 0.7991** with **runtime = 1h 12m 15s** The results can be viewed at [this Weights & Biases link](https://app.wandb.ai/khanhnamle1994/multi_layer_perceptron_collaborative_filtering/runs/4jjdo87k). # Redes-Neurais