<p>In the field of user response prediction and intelligent recommendation, mitigating the negative impact of redundant features and learning feature interaction effectively are crucial for Click-through Rate prediction(CTR) tasks. Current research often focuses on selecting a fixed subset of feature fields. This setting is highly restrictive in today’s dynamic and complex recommendation scenarios because the contribution of different features varies significantly in user-item interactions. Moreover, existing feature selection methods rarely consider the comprehensive performance of features in diverse feature interactions, potentially leading to biased feature selection. This work proposes a model named FF-MAT which sets up a global feature filter before feature interaction modules. This filter, designed with a special structure, can receive feedback from subsequenLirui Dengt modules during training to capture more effective feature filtering weights dynamically for different upcoming inputs. Additionally, we designed two feature interaction modules specifically aligned with the filter to capture both explicit and implicit feature interactions. This network architecture enhances the filter’s learning process, allowing it to effectively evaluate feature performance across various interactions for more equitable filtering. We also incorporate a multi-level optimizer to ensure model stability and prevent overfitting. Experiments show that our model outperforms baseline models in terms of performance.</p>