Research on e-commerce special commodity recommendation system based on attention mechanism and Dense Net model

This paper constructs two cross-domain recommendation models based on the perspective of user sharing and non-sharing, both of which rely on intensive convolutional networks and attention mechanisms. This research introduces lightweight Dense Net and fine-grained pruning for model optimization. Lightweight Dense Net retains the core advantages by optimizing the repeat structure while reducing redundant parameters. Compared with the original network, the accuracy loss is not >2 %, the number of parameters is reduced to 204.96Mb, the compression ratio is 8.38, and the computational amount is reduced by 0.96Gflops, which facilitates the hardware deployment. Given the problem that lightweight Dense Net has no practical optimization in storage and computing after sparsing, this paper innovatively proposes a CSB compression storage method and supporting sparse convolution algorithm, which can effectively reduce the computing and storage requirements of inference network, realize the real computing acceleration and storage optimization, and overcome the hardware deployment problems. Compared with the original network, the accuracy loss is not >2 %, the number of parameters is reduced to 204.96Mb, the compression ratio is 8.38, and the computational amount is reduced by 0.96Gflops, which facilitates the hardware deployment. Given the problem that lightweight Dense Net has no practical optimization in storage and computing after sparsing, this paper innovatively proposes a CSB compression storage method and supporting sparse convolution algorithm, which can effectively reduce the computing and storage requirements of inference network, realize the real computing acceleration and storage optimization, and overcome the hardware deployment problems.