The invention discloses an employee demission risk
estimation method based on vectorization semantic analysis. The method comprises the following steps: acquiring an employee communication
record through an enterprise communication platform API and performing sensitive information desensitization
processing; a pre-training
language model BERT or text-embeding-ad-002 is adopted to generate a 768-dimensional vector representation, and the 768-dimensional vector representation is stored in a vector
database;
dynamic clustering is carried out on the employees by using
cosine similarity calculation and an incremental K-means
algorithm, and clustering precision is optimized by using a
CLIQUE algorithm; when a demission application is received, calculating a demission risk density and dynamically setting a threshold according to a clustering scale to trigger early warning; a
state space including clustering risk density, employee characteristics and historical intervention effects is constructed, and personalized intervention strategies including communication intervention, salary adjustment, occupational development and the like are generated by using a PPO
reinforcement learning algorithm. According to the method, the hidden demission risk which is difficult to find by a traditional method can be identified, the
risk identification is advanced by 3-6 months, the high-risk employee identification accuracy is improved by 40%, the retention success rate is improved by 35%, the employee privacy is protected, and the data compliance requirement is met.