FinalMLP: An Enhanced Two-Stream MLP Model for CTR Prediction

Click-through rate (CTR) prediction is one of the fundamental tasks in online advertising and recommendation. Multi-layer perceptron (MLP) serves as a core component many deep CTR models, but it has been widely shown that applying vanilla MLP network alone ineffective learning complex feature interactions. As such, two-stream models (e.g., Wide&Deep, DeepFM, DCN) have recently proposed, aiming to integrate two parallel sub-networks learn interactions from different views for enhanced prediction. In addition stream learns implicitly, most existing research focuses on designing another complement with explicitly Instead, this paper presents simple interaction model, namely FinalMLP, which employs only MLPs both streams yet achieves surprisingly strong performance. contrast sophisticated design each stream, our work enhances modeling through selection module, produces differentiated inputs streams, group-wise bilinear fusion effectively captures stream-level across streams. We show FinalMLP competitive or even better performance against four open benchmark datasets also brings significant improvements during an A/B test industrial news recommender system. envision effective model could serve new baseline future development models. Our source code will be available at MindSpore/models FuxiCTR/model_zoo.