Softmax-function in Hands-on Tutorial

Owm112 · December 20, 2020, 9:05pm

Hello,
I was watching the WWW20-Hands-on-Tutorial and I saw the implementation for GCMC:

dglai/WWW20-Hands-on-Tutorial/blob/master/applications/recsys.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Recommender Systems with GNNs\n",
    "\n",
    "**NOTE: Though runnable, this tutorial is outdated.  We have an example of [GCMC](https://github.com/dmlc/dgl/tree/master/examples/pytorch/gcmc) with the latest APIs but we currently lack a tutorial.  Please stay tuned.**\n",
    "\n",
    "The academia and industries have witnessed various attempts and applications of applying GNNs to business applications including recommender systems, fraud detection, etc.\n",
    "\n",
    "In this hands-on tutorial, we will see how GNNs can be used for a recommender system.\n",
    "\n",
    "## Recommender System: Overview\n",
    "\n",
    "Recommender systems are typically used on E-commerce or similar platforms.  It is responsible for predicting the items that are most likely to be interacted by a user, given his/her previous interaction history.\n",
    "\n",
    "Based on whether the interactions have a *rating* that signifies how a user likes an item, we could categorize the interaction data as having either *explicit feedback* if the data has ratings, or *implicit feedback* otherwise.\n",
    "\n",

This file has been truncated. show original

In the GCMC paper of van Berg et al. they introduced a bilinear decoder with a softmax function. Why is it missing in the notebook. Is it becaus of simplicity?

Greetings!

zihao · December 21, 2020, 3:31am

Hi @Owm112, would you mind checking DGL’s example here:

where the bi-linear decoder was defined in model.py, the softmax function was not explicitly exposed because we use CrossEntropyLoss in train.py, which applies a softmax on the input: