AutoRec: An Automated Recommender System

Abstract

Recommender systems are highly specialized to handle specific data and tasks. For example, Neural Collaborative Filtering [1] takes the implicit interaction between user and item IDs as the input data for the rating prediction task. Wide & Deep learning [2] ingests user and application attributes to predict app downloads for Google Play. And DeepFM [3] leverages both numerical and categorical data to estimate the click-through rate (CTR) for ad campaigns. However, a high degree of specialization comes at the expense of model adaptability and model tuning complexity. As shown in Figure 1, the originally apt model often either becomes obsolete or requires hyper-parameter tuning as the recommendation task at hand changes and additional types and amounts of data are collected over time. The efforts required to re-tune or re-build a model is often high.

So far, several modular pipelines for building recommender systems, such as Open-Rec [4] and SMORe [5], have been proposed to address the adaptability issue. They contribute to the community by defining unified pipeline schema which divide recommendation models into a series of components (blocks) with specific functions and provide selectable modules for each. This design allows developers to quickly build and iterate recommendation models by assembling and swapping for the promising parts. Nevertheless, 1) determining which modules to use for each block and 2) hyper-parameter tuning for recommendation models remain challenging when models need to be adapted for continuously changing tasks and data.