Aiming at the multi-objective problem in preschool education program formulation, this paper proposes a multi-objective optimization mining algorithm for typical data of preschool education platform based on deep learning. Extracting content characteristics and structural characteristics of typical data of a multi-source heterogeneous education platform by adopting a time window and a frequent item set, and inputting a convolutional neural network model; introducing a multi-layer heterogeneous attention mechanism into a convolution layer of the model, and mapping the extracted characteristic results; reconstructing the mapping result by utilizing a batch normalization layer, and segmenting and reconstructing by utilizing a pooling layer to obtain the characteristic results; the typical data features after segmentation are combined by the fully connected layer of the model, and the typical data are divided by the Softmax classifier to obtain the optimized typical data multi-objective optimization mining results. The test results show that the algorithm has a good feature extraction effect, the root mean square error is less than 0.12, the specificity results of data mining are more than 0.927, and the novelty results of typical data mining are more than 91.6%.