{"id":8223,"date":"2021-04-09T22:26:27","date_gmt":"2021-04-09T22:26:27","guid":{"rendered":"https:\/\/wealthrevelation.com\/data-science\/2021\/04\/09\/deep-learning-recommendation-models-dlrm-a-deep-dive\/"},"modified":"2021-04-09T22:26:27","modified_gmt":"2021-04-09T22:26:27","slug":"deep-learning-recommendation-models-dlrm-a-deep-dive","status":"publish","type":"post","link":"https:\/\/wealthrevelation.com\/data-science\/2021\/04\/09\/deep-learning-recommendation-models-dlrm-a-deep-dive\/","title":{"rendered":"Deep Learning Recommendation Models (DLRM): A Deep Dive"},"content":{"rendered":"
\n <\/p>\n

By Nishant Kumar<\/a>, Data Science Professional<\/b>.<\/p>\n

Recommendation systems are built to predict what users might like, especially when there are lots of choices available.<\/p>\n

The DLRM algorithm was open-sourced by Facebook on March 31, 2019, and is\u00a0part of the popular MLPerf Benchmark.<\/p>\n

\"\"<\/p>\n

Deep Learning Recommendation Model Architecture (DLRM)<\/a><\/em><\/p>\n

Why should you consider using DLRM?<\/strong><\/p>\n

This paper attempts to combine 2 important concepts that are driving the architectural changes in recommendation systems:<\/p>\n

    \n
  1. From the view of Recommendation Systems, initially, content filtering systems were employed that matched users to products based on their preferences. This subsequently evolved to use collaborative filtering where recommendations were based on past user behaviors.<\/li>\n
  2. From the view of Predictive Analytics, it relies on statistical models to classify or predict the probability of events based on the given data. These models shifted from simple models such as linear and logistic regression to models that incorporate deep networks.<\/li>\n<\/ol>\n

    In this paper, the authors claim to succeed in unifying these 2 perspectives in the DLRM Model.<\/p>\n

    A few notable features:<\/p>\n

      \n
    1. Extensive use of Embedding Tables<\/strong>: Embedding provides a rich and meaningful representation of the data of the users.<\/li>\n
    2. Exploits Multi-layer Perceptron (MLP):\u00a0<\/strong>MLP presents a flavor of Deep Learning. They can well address the limitations presented by the statistical methods.<\/li>\n
    3. Model Parallelism:\u00a0<\/strong>Poses less overhead on memory and speeds it up.<\/li>\n
    4. Interaction between Embeddings<\/strong>: Used to interpret latent factors (i.e., hidden factors) between feature interactions. An example would be how likely a user who likes comedy and horror movies would like a horror-comedy movie. Such interactions play a major role in the working of recommendation systems.<\/li>\n<\/ol>\n

      <\/p>\n

      LET’S START<\/strong><\/p>\n

      \u00a0<\/p>\n

      Model Workflow:<\/h3>\n

      \u00a0<\/p>\n

      <\/p>\n

      DLRM Workflow.<\/em><\/p>\n