The mainstream multimedia recommendation models investigate to fuse multimodal collaborative signals by graph learning for modeling users’ interests; however, the standard graph convolution technique suffers from the over-smoothing problem when deepening the convolutional layers. More convolutions would hinder the personality and the recommendation performance. On this point, we propose a path-based interesting mining (PIN) model to propagate collaborative signals along behavioral paths for multimedia recommendation. Different from the convolutional models, PIN adopts a non-convolutional manner for inter-node interest aggregation. PIN adaptively weights nodes on various paths through a path-aware attention mechanism and then fuses the modality signals adaptively for interest modeling. This robust long-path aggregation captures long-distance collaborative signals beyond the convolutional layer limitation. The extensive experiments on three real-world datasets demonstrate the superior performance of PIN, especially improving 41.23% over the state-of-the-art baseline on the Baby dataset, as well as confirming its effectiveness in mitigating the over-smoothing issue for multimedia recommendation.