Unleash the Power of Context: Enhancing Large-Scale Recommender Systems with Context-Based Prediction Models

In online advertising systems, the accuracy of these models is crucial for the success of advertising campaigns and the revenue generated by publishers. Advertisers rely on CTR prediction models to target their ads to the right audience and optimize their advertising budget, while publishers use these models to maximize their revenue by displaying ads that are most likely to be clicked. CTR prediction techniques continue to be an active area of research in both industry and academia [1, 5, 8]. In many commercial use cases, the CTR prediction model consists of billions of weights and must perform inference billions of times per second [3]. Therefore, any improvements applied to the model must be carefully balanced with the cost of serving. In this paper, we introduce the notion of Context-Based Prediction Models and demonstrate its effectiveness. A Context-Based Prediction Model determines the likelihood of an action (such as a click or a conversion) by solely considering user and contextual features, without taking into account any specific characteristics of the item itself.