Neural Collaborative Filtering

• ✓Build GMF, MLP and NeuMF in PyTorch and train them on real MovieLens interactions
• ✓Train on implicit feedback with sampled negatives, resampled every epoch the way the paper specifies
• ✓Learn user and item embeddings end-to-end instead of freezing factors from a matrix decomposition
• ✓Turn a pairwise scoring network into a ranked catalog, masking items the user has already seen
• ✓Measure the neural models against the item-KNN baseline on one split with one evaluation function
• ✓Show why a model that wins by under a percent may still not be the model to ship
• ✓Locate the two failures — linear scoring cost and ID-indexed cold start — that motivate two-tower retrieval

Build the architecture that took recommenders from hand-picked similarity functions to learned ones — and, just as importantly, the comparison discipline that tells you whether the neural version actually earned its complexity on your data.