3.1 CNN Entrepreneurship Project Recommendation Algorithm Design for Student Entrepreneurship Needs

The screening of industry-education-integrated entrepreneurial projects can be extremely time-consuming and difficult for students. As recommendation systems allow for more accurate and personalized recommendations because they actively search and analyze the existing and historical behavioral information of users. Therefore, the research makes an in-depth analysis of the recommendation system for student-oriented entrepreneurial projects. First, the historical behavior of the student users can be used to obtain their rating calculation for a particular entrepreneurial project, as in Eq. (1).

where $r_{ui}$ stands for the student user’s final rating of the item $i$ by the student user $u$; $a_{i}$ represents the weight of the $j$ interaction dimension; $q_{u,i,j}$ stands for the number of interactions of the item $u$ by the student user $j$ on the interaction dimension. There was no negative sample because the user behavior data is implicit feedback data. For this reason, the sampling method proposed by Xiang Liang was adopted to carry out negative sample sampling ^[16]. A collaborative filtering algorithm based on PMF was used to decompose the implicit characteristics of users and projects through the matrix ^[17]. First, it was assumed that the implicit feature vectors of students and entrepreneurial projects obey the Gaussian distribution. The calculation is shown in Eq. (2).

where $N\left(u_{i}\left| \mu ,{\sigma }_{V}^{2}I\right.\right),N\left(v_{j}\left| \mu ,{\sigma }_{V}^{2}I\right.\right)$ is the probability density function; $\mu $ represents mean value; ${\sigma }_{V}^{2}$ represents variance. $U,V$ represent the implied eigenmatrices of the student users and entrepreneurial projects, respectively; $u_{i},v_{i}$ are the implied eigenvectors of student users and entrepreneurial projects, respectively, and $I$ is the unit matrix. Assuming the conditional probabilities of the observed ratings followed a Gaussian prior distribution, as expressed in Eq. (3).

where ${I}_{i,j}^{R}$ is the indicator, which is 1 when the user $i$ rates the item $j$ and 0 when there is no rating. The implied characteristic posterior probabilities of the user and the item were obtained using the Bayesian function as in Eq. (4).

where the implied features of the venture $V$are not obtained by the probability MDA, but they are calculated using DNN. Assume $W$ is the series of parameters of the DNN, and it is Gaussian distributed. The feature $X_{j}$$V_{j}$ representing the $j$\textsuperscript{th} venture is input, and the $j$\textsuperscript{th} venture was calculated using Eq. (5).

where $dnm\left(W,X_{j}\right)$ stands for DNN, and $\varepsilon _{j}$is the random error. Assuming that the set of parameters $W$ also conforms to a Gaussian distribution. The probability distribution of $W$ and the conditional probability distribution of the item implied feature $V$ is expressed as Eq. (6).

Once the implied feature matrix $V$, the student user implied features $U$, and the set of DNN parameters $W$ are solved, the probability distribution can be calculated in Eq. (7).

Startup projects contain considerable characteristic information, both structured and unstructured data. More useful characteristics can be extracted from this content by combining the CNN with word embedding techniques and one-hot techniques to build a CNN, which is used to extract the implicit features of the projects. Fig. 1 presents the DNN structure ^[18].

The input layer of this network structure is the cleaned and processed startup project feature data (Fig. 1), which consists of four parts. The word vector method is mainly used to process the labels, and the word2vec algorithm is used to train all the label sequences and obtain the label vectors of the entrepreneurial projects. The trained label vectors cannot be used directly as input to neural networks because there are multiple labels for each entrepreneurial project. The study used the label vector summation and averaging method, assuming the set of labels for each entrepreneurial project is $T=\left\{t_{1},t_{2},\ldots ,t_{l}\right\}$, where $t_{i}\in p^{1\times k}$. Each label is then processed as expressed in Eq. (8).

where $l$ is the number of labels of the entrepreneurial project and $p$ denotes the processing result. This study used a CNN to process the text description of the venture, whose structure is shown in Fig. 2, to optimize the accuracy rate of the implied features of the entrepreneurial projects ^[19].

The network structure consisted of seven main components (Fig. 2). First, the embedding layer aims to convert the sequence of word separation results of a piece of text into a matrix that can be fed into this network, which is then used as input to the text CNN. The convolutional layer is designed to extract the local features. Its main purpose is to extract local features of text or images using local sensory undertakings, shared weights, and biases. The convolution operation of the CNN for this study is expressed as Eq. (9).]

where ${c}_{i}^{j}$is the activation value on a convolutional kernel; $f$ represents the activation function of a neuron; ${b}_{c}^{j}$ is the shared bias; $w_{i,j}$ represents the shared weights; $D$ stands for the input to the convolutional layer. The goal of the pooling layer is to simplify the output of the convolutional layer, in effect, a sampling operation. The fusion layer accepts the output of the CNN under study and directly splices it horizontally with the region. The domain and label serve as input to the lower layer. The individual features that have been fused need to be extracted from the abstract characteristics of the items by the fully connected layer, which is calculated using Eq. (10).

where $f$ stands for activation function; $W_{fc}$ is the set of parameters; $b_{fc}$ is the set of biases. Finally, the output layer is the one that converts the output of the previous layer into a dimension-specific vector of implicit features of the items.

The above algorithm is optimized using a maximum posteriori estimation approach with the optimization objective shown in Eq. (11).

where $\lambda _{U}=\frac{\sigma ^{2}}{{\sigma }_{U}^{2}},\lambda _{V}=\frac{\sigma ^{2}}{{\sigma }_{V}^{2}},\lambda _{W}=\frac{\sigma ^{2}}{{\sigma }_{W}^{2}}$ is the error function with a regularization parameter. $\lambda _{U},\lambda _{V},\lambda _{W}$ are the regularization parameters that need to be verified experimentally by a poorer difference. The optimal $U,V,W$ are obtained by optimizing Eq. (11) for the section class. The study uses the gradient descent algorithm to find the first-order partial derivative of $U,V$ in this equation and make it equal to 0 to find the direction of gradient descent as expressed in Eq. (12).

where $W$ is not directly available as a set of parameters for DNN. On the other hand, when $U,V$ are determined, it can be found by combining DNN with MDA. The implicit feature $U,V$ also predicts the missing rating information in the rating matrix $R$.

3.2 DNN Recommendation Algorithm based on the Construction of a Recommendation Model for The Integration of Learning, Production, and Education Entrepreneurship Projects

Suppose only neural networks are used to learn the user’s implicit characteristics from their content features. In that case, the recommendation algorithm does not consider the user’s behavioral information, which is equivalent to its degradation to a content-based recommendation algorithm. In general, the content features of student users tend to be static and unchanging, and modeling only the content feature information of student users will inevitably reduce the accuracy of the recommendation results. Hence, the historical behavior also needs to be modeled. Assuming that the implied features of the student user and the entrepreneurship project obtained from the decomposition of the rating matrix are $U,V$, respectively, the predicted rating of the project by this user can be calculated using Eq. (13).

where $i$ represents a student user and $j$ represents a business venture. The implicit feature of $u_{i}$ represents the implicit feature of the student user, and $i$ stands for the implicit feature of the entrepreneurial project. The optimized loss function of the collaborative algorithm based on the implied features is shown in Eq. (14) ^[20].

where $r_{i,j}$ is the true rating of the entrepreneurial project $j$ by the student user. ${\sum }_{i=1}^{n}{\sum }_{j=1}^{n}\left(r_{i,j}-\widehat{r_{i,j}}\right)^{2}$ is the mean square error between true and predicted ratings. $\lambda _{U}\left|\left|U\right|\right|^{2}+\lambda _{V}\left|\left|V\right|\right|^{2}$ is the regularization term and $\lambda _{U},\lambda _{V}$ are the regularization parameters. Eq. (14) can be transformed into (15), which represents the implicit feature $U$ as a function of the student user’s historical rating matrix $R$^[21].

where $f$ is a non-linear transformation; $R$ is the scoring matrix; $W$ is the set of parameters. The implicit interest features of the student user are also influenced by the user content features, i.e., the implicit user features $U$ are functionally related to the content features of the student user. Assuming the user content feature vector is $X$, this can be translated to Eq. (16).

For the student user $i$, the implicit feature vector is expressed as Eq. (17).

CNNs can be used for model building because of their ability to model arbitrary functions to mitigate the negative impact of data sparsity on the model. The student user ratings are reduced dimensionally, and the results are fed into the neural network. The RBM algorithm can reduce the dimensionality of the student user’s rating vector, where the visible layer represents the students’ rating data, and the hidden side stands for the reduced dimensional data. The reduced-dimensional rating data is used as an input to the CNN; Fig. 3 shows the structure of the student user.

In addition to using dimension reduction rating data as network input (Fig. 3), the other characteristics of the student user, such as the student’s region, profile, and area of interest, can also be used as network input. In summary, the research proposes a CNN recommendation algorithm based on a DNN recommendation algorithm as the recommendation model for the student production and education fusion entrepreneurship project, as shown in Fig. 4.

From Fig. 4, the left side of this model is the student user network model, and the right is the entrepreneurial project network model. The entrepreneurial project network structure uses a CNN to process the unstructured text information. Word embedding and one-hot are used to process the others. The student's model consists mainly of modeling the historical behavioral and content features of the student users. For the historical behavioral information of the student user, the study uses the RBM algorithm for processing. It then uses it as input to the network. The remainder of the processing is similar to the content features of the entrepreneurship project.

The network processing of the model was roughly divided into three stages. The first is the feature and stitching processing, where the student user and entrepreneurial project features are processed separately. The two are stitched together to generate the corresponding high-latitude spatial vector representation. It then approaches the fully connected layer-by-layer abstraction stage, where the high-dimensional spatial vectors of student users and entrepreneurial projects are again passed through fully connected layers, each extracting features from student users and entrepreneurial projects and abstracting them layer by layer. The final output is a vector of implicit features of the student user and the entrepreneurial project in the same vector space. The scoring stage is the final of the model, where an MDA can calculate the inner product of the implied features of the student user and the entrepreneurial project for the two network outputs, which is the student user score.

The pre-training of the model is training the RBM network, which aims to extract dense features from the sparse scoring data ^[22]. During the supervised training phase, the student rating features and other features are obtained from the pre-training combination and used as input to the student network. The text features and other features of the entrepreneurial project are used as input to the project network ^[23,24]. The calculation of each network layer is expressed in Eq. (18).

where $X_{h}$ represents the activation function of the $h$ layer; $W_{i}$is the set of parameters; $b_{i}$ represents the bias; $\sigma $represents the activation function. The study uses the function as the activation function, which is calculated using Eq. (19).

In the training of the supervised model, the objective function is expressed as Eq. (20).

where $W_{u},W_{v}$ are used to extract the set of neural network parameters for student user and entrepreneurial project features, respectively. The study uses the gradient descent method to obtain $W_{u},W_{v}$, which can calculate the ratings of all student users on entrepreneurial projects. After ranking from the largest to the smallest, the first $k$ entrepreneurial projects are selected as recommendation results.

4.1 Experimental Results Analysis of CNN Entrepreneurial Project Recommendation Algorithm for Students’ Entrepreneurial Needs

This study used a Python crawler to crawl data on the student users’ behaviors, characteristics, and projects from the entrepreneurship project information platform. The student users with a low number of behaviors with too much data and the startup project data were removed afterward. There were 7,780 boat business projects and 7,770 student users. After processing the data through implicit feedback, 248,610 student user ratings were obtained with a sparsity of 99.60% of the data. Table 1 lists the missing rate of each field of the student characteristics information and the rating data after implicit feedback processing.

From Table 1, a large proportion of student user information was missing, with 78.53% missing for user indirect, which is unusable because of its over-representation. The missing percentages of the student user region, area of interest, and their purpose on the website were 23.28%, 25.40%, and 20.10%, respectively, which could be retained because the fields were all processed using one-hot coding. Second, the rating data in Table 1 was calculated from the student user’s historical behavior data, with each row representing the student rating data, where the rating data includes the student number, the entrepreneurship project number, and the rating in three columns.

The effectiveness of the designed algorithm was assessed objectively. The PMF algorithm and the Conv MF algorithm, which use MDA to extract implicit characteristics of students from the scoring matrix, were selected for comparison with the research proposed DNN. Fig. 5 shows the RMSEs of the various algorithms with the recall rates at different amounts of content features.

From Fig. 5(a), the RMSE obtained by the DNN-MF algorithm studied was the lowest (0.5212). The RMSE value of the proposed algorithm was 0.1677 and 0.119, smaller than that of the PMF algorithm and ConMF algorithm, respectively. The recall rate of the three algorithms increased as the number of content features increased (Fig. 5(b)). Among them, under the condition of the same number of content features, the recall rate of the proposed algorithm was higher than the other two algorithms. Moreover, compared to the PMF algorithm, the improvement range of the proposed algorithm was 0.1420–0.2110, and compared with the ConMF algorithm, the improvement range was 0.0477–0.0519. In conclusion, the recommendation accuracy can be improved significantly by integrating the content characteristic information of entrepreneurial projects into the recommendation system.

The DNN-MF1 model is based on the Conv MF model with the label information of the project, and the DNN-MF2 model is based on the DNN-MF1 model with the information of the funding round and region. Fig. 6 presents the experimental results, where Fig. 6(a) shows the RMSE of various algorithms, and Fig. 6(b) shows the line graphs of different amounts of content features of various algorithms.

After gradually adding more project content information, the RMSE value of the DNN-MF2 algorithm studied was 0.5109 (Fig. 6(a)). The RMSE value of the proposed algorithm was 0.1293 and 0.1030 smaller than that of the ConMF algorithm and DNN-MF1 algorithm, respectively. From Fig. 6(b), under the condition of different amounts of content features, the recall rate improvement range of the DNN-MF2 algorithm compared to the DNN-MF1 algorithm was 0.0347–0.0743, and the recall rate improvement range of the DNN-MF2 algorithm compared to the ConMF algorithm was 0.0525–0.0849. Overall, the research on integrating more content features into the recommendation system can improve the recommendation accuracy of the model. The hyperparameters of the DNN-MF algorithm include the number of hidden features and the length of the word vector. A grid search was used to relate them to the recommendation accuracy of the DNN-MF algorithm. Fig. 7 shows the experimental results.

The RMSE value obtained by the research algorithm was relatively high when the number of word vectors was 100, as shown in Fig. 7(a). The mean RMSE value of the DNN-MF algorithm was 0.62786 in 50 to 250 implied feature numbers. From Fig. 7(b), the RMSE value of the research algorithm was the lowest in the range of 50–200 implied feature numbers when the number of word vectors was 200. The average RMSE value of the research algorithm in 50 to 250 implicit feature numbers was 0.6261. According to Fig. 7(c), the average RMSE value of the DNN-MF algorithm in 50 to 250 implied feature numbers was 0.62782 when the number of word vectors was 300. In conclusion, the recommendation effect of the research model is the best when the number of word vectors is 200, that is, the optimal word vector length is 200.

4.2 Experimental Results Analysis of the Recommendation Model of the Fusion of Learning, Production0.633 and Education Entrepreneurship Project based on CNN Recommendation Algorithm

This experiment chose PMF, Conv MF, and the recommendation algorithm DNN-MF in subsection 2.1, and the recommendation model DNN-DNN designed in section 2.2 based on the DNN recommendation algorithm for the fusion of learning, production, and teaching entrepreneurship projects for comparative analysis to demonstrate the effectiveness of the model designed by the Institute more objectively, and results are shown in Fig. 8.

Fig. 8 shows that the RMSE value of the DNN-DNN algorithm was 0.4931, which was 0.1959, 0.1489, and 0.0281 smaller than the PMF algorithm, Conv MF algorithm, and DNN-MF algorithm, respectively. In the recall rate index, the DNN-DNN model constructed in this study also showed great improvement compared to the other models. Under the condition that the number of content features was 50, 100, 200, and 300, the recall rate obtained using the study model was higher than that of the other models. Therefore, compared to PMF, Conv MF, and DNN-MF models, the recommendation model of learning-teaching-integrated entrepreneurial projects based on the CNN recommendation algorithm proposed in this study can significantly improve the recommendation accuracy of student-oriented projects with industry-education integrated entrepreneurial needs. The user model in the research model DNN-DNN can either use only the rating data of the student user as input, or it can include the content features of the student as input along with the rating data. The study conducted experiments on two models with or without adding student user features to determine the impact of adding student user features on the recommendation results. The DNN-DNN1 model indicates that no other features of the student user are added, and the DNN-DNN2 model indicates that the student user content features are added, including the student user’s region, area of interest, and the student user’s purpose. Fig. 9 shows the experimental results.

The RMSE of the DNN-DNN2 model with the added feature was 2.23% lower than that of the DNN-DNN1 model without adding student user features (Fig. 9(a)). The accuracy of the DNN-DNN2 model constructed in the study increased by approximately 2.35% compared to the DNN-DNN1 model (Fig. 9(b)). The results show that adding user characteristic information to the DNN-DNN model can help improve the accuracy of the recommendation algorithm.