Blockchain sharding technology is a key approach to addressing scalability issues in blockchain systems. It enhances the network's processing capabilities and overall performance by dividing the network into multiple shards, which can process transactions and computations in parallel. However, traditional sharding techniques often lead to a high proportion of cross-shard transactions and a complex transaction verification process, which curtails the potential increase in blockchain throughput. To address this issue, we propose a transaction sharding algorithm based on community detection, CDShard. This algorithm partitions the blockchain network into multiple shards, each processing distinct sets of transactions in parallel, aiming to cluster accounts with closely related transactions within the same shard. This strategy significantly reduces the number of cross-shard transactions and the complexity of inter-shard communication, ultimately presenting the transaction-sharding results in the form of account-to-shard mappings. Experimental results demonstrate that the CDShard algorithm reduces the cross-shard transaction ratio from 80% to approximately 20%, thereby effectively enhancing the system's throughput.