Aiming at the problem that naive Bayes（ NB） algorithm consume a large amount of system and network resources in the early training stage

which seriously affects the classification efficiency

the idea of using tree structure is proposed. The array is used in the NB algorithm to maintain the feature words in the training samples.The number of occurrences is optimized for improvement. For the NB algorithm

when the number of mail sample attributes is large

the classification effect is poor

and the conditional probability of the feature words is taken as the rooting

which increases the sensitivity of the system to high-frequency vocabulary. The experimental results show that compared with the NB algorithm

the improved algorithm has better effects in training time

precision

reconciliation rate

etc. By adjusting the z value of the rooting times

the false positive of spam is reduced

the experiment found that when the z value is taken to 3

the classification performance indicators achieve a satisfactory effect.