Like2Vec with neural network embeddings
summary
Applies Word2Vec-style shallow neural embeddings to user–item graphs, building Like2Vec embeddings from an item co-occurrence matrix weighted by log‑likelihood ratio and trained with DL4J/Skymind-style tooling. Embeddings use cosine similarity to rank candidates and evaluate with a Rec@N protocol; results show Rec@N twice the hits at N=1 versus a log‑likelihood baseline and lower RMSE trends when increasing neighbor K with weighted averaging. Implements symmetric and proposed asymmetric (time‑ordered) graphs for session ordering and reports practical gains on top‑N recommendation tasks.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:40Matrix factorization
- 6:42nonlinear auto encoding
- 16:17Advanced versions
- 17:29Mutual information graph
- 18:54For the deep walk
- 19:26Continuous backward
- 19:51Lower dimensional representation
- 20:40User symmetric matrix
- 21:00Log likelihood ratio
- 21:53Like2Vec recipe
- 22:54Items recipe
- 23:18Results
- 23:45Evaluation metric
- 25:28Evolution metric
- 29:04Recall metric results
- 31:49Predicted ratings
- 33:43Future directions
- 35:45Conclusions
related talks
21mGraph Representation Learning with Deep Embedding Approach
28m"Embeddings: How Computers Learned to Read"
38mEmbeddings! Embeddings everywhere!
38mWord2Vec model to generate synonyms on the fly in Apache Lucene
43mpackage2vec: getting to know PyPI packages with ML
30mDeep Learning for Recommender Systems
28mEmbeddings
55mVector Similarity Search in Spring with Redis Stack