Algorithm 2 Sentiment prediction model training algorithm.

Require:  Dateset $Data=\left\{x_n\right\}$; The expected number of hidden layers N; The number of pre-training iterations is T Ensure:  Optimization results; Sentiment prediction; Weights  1:Collect early-stage user comments on food public sentiment, analyze them, calculate the number of negative and positive sentiments, and use them as the training dataset $Data$.  2:Collect user sentiment feedback on current food public sentiment within the initial time period t.  3:Set the weight of odd-numbered terms as $\xi$, Sparsity parameter p; Initialize the matrix and random bias vector.  4:Greedy layer-wise algorithm training for hidden layers.  5:Train the first layer using Algorithm 1. The input dataset is the training dataset $Data$.  6:Starting from the second layer, use the output of the hidden layer as the input for the next layer, obtaining the encoding-decoding matrix and bias vector for the next hidden layer.  7:Utilize backpropagation combined with gradient optimization techniques to perform top-down layered adjustments of parameters across the entire neural network.  8:Based on the data collected from step 2, make predictions to obtain overall positive and negative results.  9:Construct the weight quantifier: $\vartheta$ is the number of positive comments and $\omega$ is the number of negative comments If the prediction result is positive: Public sentiment index $\epsilon =1+{\displaystyle \frac{\vartheta }{\vartheta +\omega }}$, increase recommendation weight If the prediction result is negative: Public sentiment index $\epsilon =1-{\displaystyle \frac{\omega }{\vartheta +\omega }}$, decrease recommendation weight If the prediction result is neutral($\vartheta =\omega$): Public sentiment index $\epsilon =1$.