A high performance computing technology powered multimedia fusion model in university English translation

Long short term memory (LSTM) model construction.

The deep averaging network (DAN) model uses user and material content interaction behavior to capture similarities between users and items. The model focuses only on implicit feedback, i.e., whether they have read a given article and in what order they viewed it. The order in which a user reads an article encapsulates information about the user’s interests, and the DAN model uses LSTM to learn the order dependencies of the articles read to capture the user’s interests (Shen et al., 2022). Users’ interests change dynamically over time, so a bidirectional LSTM structure is used. We select each user’s specific browsing history data as the input to the LSTM. A neural attention mechanism is also introduced to capture the similarity between users and items. Then, in order to learn the parameters of the model, the DAN model uses an objective function based on ranking learning. Finally, material is recommended to the user based on the similarity of the computed inner product vector between the user and the item.

DAN is designed as a parallel convolutional neural network (PCNN) consisting of two convolutional neural networks (CNN). The PCNN fuses the feature information of the material. DAN introduces an artificial neural network and designs an ARNN component to capture the historical readings of the user with potential sequential features (Henderson & Corry, 2021). ARNN is an attention-based recurrent neural network can obtain richer sequential features at different click times. An attention mechanism can be used to capture the different effects of user clicks on the candidate material to model the current interest of the user. The model structure is described in detail below.

Suppose the user’s historical click material is {X_n−t, X_n−t+1, …, x_n}. where x_n is the nth clip clicked by the user. Given the candidate material as x_n, the recommendation system takes {X_n−t, X_n−t+1, K, …, x_n} as input and needs to calculate the probability of the user clicking on the candidate clip x_n. (1)${\displaystyle X_t=\frac{1}{n}\sum _{i=1}^m{V}_i{C}_t.}$

V_i represents the click speed, X_t represents the specific user, and m represents the number of user clicks. In the input gate, which indicates the information to be updated, the i_t values range from 0 to 1 and are used to select the proportion of new information to be remembered, and C_t, which indicates the alternative cell state (Rodriguez-Medina, Dominguez-Isidro & Ramirez-Martinell, 2022). (2)${\displaystyle i_t=\left(\sigma \left(W_i\cdot \left[h_{t-1},x_t\right]\right)\right)+b_i}$ (3)${\displaystyle C_t=tanh\left(W_c\cdot \left[h_{t-1},x_t\right]\right)+b_c.}$

The LSTM output results using sigmoid as the activation function, where the resulting user vector formula is shown in Eq. (4). (4)${\displaystyle y_{user}=sig\mathrm{mo}id\left(w\cdot \left(y_t+y_{feature}\right)\right)=\frac{1}{1+e^{\left(w\cdot \left(y_t+y_{feature}\right)\right)}}.}$

In Eq. (3), w represents the weights; y represents the output value; and dot product represents the convolution operation.

CNN model construction.

The input of the CNN is the word embedding vector of the material, denoted as {X₁, X₂, K…, x_n}, and U_I denotes the result of the 1st convolution (Zhao et al., 2021). (5)${\displaystyle U_I=\sum _{i=I}^{I+M}{X}_i\otimes W}$

Use global average pooling layer for dimensionality reduction: (6)${\displaystyle y_{poll}=\frac{1}{m}\sum _{i=1}^m{U}_i.}$

Using the Sigmoid activation function, obtain the material text vector: (7)${\displaystyle y_{material}=sig\mathrm{mo}id\left(y_{poll}\right)=\frac{1}{1+e^{\left(y_{poll}\right)}}.}$

Obtain the user prediction score with the formula: (8)${\displaystyle y_{pred}=u_i{v}_j^T=\sum _{k=1}^K{u}_{ik}{v}_{kj}.}$

where u represents the user vector and v represents the item vector.