Improved DeepFM-Based Model with LSTM on Click-Through-Rate
Yu Li, Chiawei Chu, Xingyu Liu · 2024
In recent years, the integration of deep learning into recommendation systems has significantly improved their ability to make accurate suggestions by capturing complex user interaction patterns. However, conventional deep learning models often overlook sequential patterns and prioritise high and low order features. To address this limitation, our work proposes a novel model called lstmDeepFM, which combines DeepFM with Long Short-Term Memory (LSTM) networks to optimise click-through rate (CTR) prediction tasks. Our approach involves transforming word features into dense, low-dimensional vectors using an embedding layer, and enriching article features with context using Doc2Vec. These features are then fused and fed into the FM, DNN and LSTM networks, with particular emphasis on integrating a self-attention mechanism into the DNN component to effectively capture essential information. Finally, a novel output function is applied to refine the predictions. Evaluation on real datasets, Criteo and Avazu, using mAUC as the metric, shows significant performance improvements, with our model achieving improvements of 1.31% and 3.28% on the respective datasets.