Multimodal data-driven product crowdfunding success rate prediction method and system

Abstract

The application provides a kind of multimodal data-driven product crowdfunding success rate prediction method and system, it is related to deep learning technical field.The hierarchical hybrid model is constructed, and the structured data is filtered to obtain the structured features by structured branch;Text information and picture information are processed by text processing model and image processing model respectively to obtain semantic features and visual features;According to the activation probability of multiple expert networks obtained by semantic features and visual features, the expert network corresponding to the larger activation probability is activated, if the crowdfunding project focuses on aesthetic related products, the beauty expert or creativity expert will be activated, and the specialized implicit features are obtained;These specialized implicit features and structured features are input into the feature fusion model to obtain the prediction success rate.In processing unstructured data closely related to product appearance and aesthetic features, expert networks can highlight and strengthen these features, improving the prediction accuracy of aesthetic crowdfunding projects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a multi-modal data driven product crowdfunding success rate prediction method and system. BACKGROUND

[0002] As an innovative financing model, crowdfunding gathers small amounts of funds from individuals or organizations through an Internet platform to support creative projects, start-ups and public activities. After obtaining financial support, production begins, significantly reducing the pre-funding pressure of project initiators, enabling more creative and potential projects to start. At the same time, crowdfunding can also be an effective tool for testing market demand, helping product iteration by assessing market acceptance of new products and user feedback.

[0003] Traditional structured modeling evaluation methods (such as Mollick's random forest model) or traditional expert scoring-based evaluation methods are limited to structured variables such as target amount and fundraising period, and cannot handle unstructured data closely related to product appearance and aesthetic features, resulting in high prediction error for aesthetic crowdfunding projects and difficulty in accurately reflecting users' true evaluation of creative projects. SUMMARY

[0004] The problem to be solved by the present application is that existing evaluation and prediction methods are limited to structured data, resulting in high prediction error for aesthetic crowdfunding projects.

[0005] To solve the above problems, in a first aspect, the present application provides a multi-modal data driven product crowdfunding success rate prediction method, comprising:

[0006] inputting structured data of the crowdfunding project into a structured branch of a hierarchical mixed model to screen structured features, wherein the hierarchical mixed model comprises a structured branch, a hidden feature branch and a feature fusion model;

[0007] inputting text information of the crowdfunding project into a text processing model to obtain semantic features, wherein the hidden feature branch comprises a multi-modal deep learning model, the multi-modal deep learning model comprises a text processing model, an image processing model and an expert network;

[0008] inputting picture information of the crowdfunding project into an image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder is used for cross-modal parameter sharing and semantic alignment through a contrastive learning loss function;

[0009] According to the semantic features and the visual features, an activation probability of a plurality of expert networks is obtained, wherein the expert networks include a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert and a creative expert;

[0010] According to the activation probability, a corresponding expert network is activated sparsely to obtain specialized implicit features corresponding to the expert network.

[0011] The structured features and the specialized implicit features are input into a feature fusion model to obtain a predicted success rate.

[0012] In a second aspect, the present application further provides a multi-modal data driven product crowdfunding success rate prediction system, comprising:

[0013] A feature extraction layer is configured to input structured data of a crowdfunding project into a structured branch of a hierarchical hybrid model to obtain structured features through screening, wherein the hierarchical hybrid model comprises a structured branch, an implicit feature branch and a feature fusion model.

[0014] The implicit feature branch comprises a multi-modal deep learning model, and the multi-modal deep learning model comprises a text processing model, an image processing model and an expert network.

[0015] The image processing model is configured to input picture information of the crowdfunding project into the image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder is configured to share parameters and align semantics across modalities through a contrastive learning loss function.

[0016] The expert network comprises a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert and a creative expert.

[0017] The expert network comprises a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert and a creative expert.

[0018] The feature fusion layer is configured to input the structured features and the specialized implicit features into the feature fusion model to obtain a predicted success rate.

[0019] The present application provides a multi-modal data driven product crowdfunding success rate prediction method and system. Compared with the prior art, the present application has the following beneficial effects:

[0020] The structured data passes through a structured branch to extract structured features; the text information and the picture information pass through a text processing model and an image processing model in an implicit feature branch respectively to extract semantic features and visual features, and since the text processing model and the image processing model share a Transformer encoder, the text and the image can be aligned in a unified semantic space, so that the representation forms of the extracted semantic features and the visual features are consistent, laying a foundation for subsequent feature fusion and improving the accuracy of feature fusion; according to the semantic features and the visual features, activation probabilities of multiple expert networks can be obtained, and the expert network corresponding to a larger activation probability is activated, if the crowdfunding project focuses on products related to aesthetics, the beauty expert, the creativity expert or the functional expert is activated, and specialized implicit features are obtained from the extracted semantic features and the visual features, for example, the implicit features related to beauty, creativity or functionality are obtained by targeted strengthening; the specialized implicit features and the structured features are input into a feature fusion model to obtain a predicted success rate, the success rate predicted based on the specialized implicit features is more consistent with the actual success rate, and the accuracy of the prediction is ensured. Therefore, when processing unstructured data closely related to product appearance and aesthetic features, the expert network can highlight and strengthen such features, and the prediction accuracy of aesthetic crowdfunding projects is improved. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0022] Figure 1 A flowchart of a multi-modal data driven product crowdfunding success rate prediction method provided by an embodiment of the present application;

[0023] Figure 2 A structural diagram of a multi-modal data driven product crowdfunding success rate prediction system provided by an embodiment of the present application;

[0024] Figure 3 A structural diagram of implicit feature analysis provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0026] In order to better understand the above technical solutions, the above technical solutions will be described in detail in combination with the drawings of the specification and specific embodiments.

[0027] As shown in the figure, the multi-modal data driven product crowdfunding success rate prediction method provided by the embodiments of the present application comprises: Figure 1

[0028] S110: input the structured data of the crowdfunding project into the structured branch of the hierarchical mixed model, and screen to obtain structured features, wherein the hierarchical mixed model comprises a structured branch, a hidden feature branch and a feature fusion model.

[0029] S120: input the text information of the crowdfunding project into the text processing model to obtain semantic features, wherein the hidden feature branch comprises a multi-modal deep learning model, and the multi-modal deep learning model comprises a text processing model, an image processing model and an expert network.

[0030] S130: input the picture information of the crowdfunding project into the image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder is used for parameter sharing and semantic alignment across modalities through a contrastive learning loss function.

[0031] S140: according to the semantic features and the visual features, obtain the activation probability of a plurality of expert networks, and the expert networks comprise visual experts, language experts, cross-modal experts, functional experts, aesthetic experts and creative experts.

[0032] S150: according to the activation probability, activate the corresponding expert network with sparsity, and obtain the specialized hidden features corresponding to the expert network.

[0033] S160: input the structured features and the specialized hidden features into the feature fusion model to obtain the prediction success rate.

[0034] ​In the optional embodiment, a hierarchical hybrid model is constructed, structured data passes through a structured branch to extract structured features; text information and picture information pass through a text processing model and an image processing model in an implicit feature branch respectively to extract semantic features and visual features, and since the text processing model and the image processing model share a Transformer encoder, alignment of the text and the image in a unified semantic space can be ensured, the representation forms of the extracted semantic features and the visual features are consistent, a foundation is laid for subsequent feature fusion, and the accuracy of feature fusion is improved; according to the semantic features and the visual features, activation probabilities of multiple expert networks can be obtained, an expert network corresponding to a larger activation probability is activated, if the crowdfunding project focuses on products related to aesthetics, an aesthetic expert, a creativity expert or a functional expert is activated, since the expert network is a specialized expert network obtained through constraint training, processing through these expert networks can obtain specialized implicit features from the extracted semantic features and visual features, for example, aesthetic, creativity or functionality related implicit features are obtained through targeted reinforcement, and these specialized implicit features can highlight the characteristics of the crowdfunding product; the specialized implicit features and the structured features are input into a feature fusion model to obtain a predicted success rate, the success rate predicted based on the specialized implicit features is more consistent with the actual success rate, and the accuracy of the prediction is ensured. Therefore, when processing unstructured data closely related to product appearance and aesthetic features, the method can highlight and strengthen such features through expert networks, and improve the prediction accuracy of aesthetic crowdfunding projects.

[0035] As shown in Figure 2 A hierarchical hybrid model is built, the hierarchical hybrid model includes a structured branch and an implicit feature branch, the structured branch is used to process structured data, and the implicit feature branch is used to process unstructured data, mainly processing text information and picture information; in the structured branch, data type features or category type features in the structured data can be recognized or extracted in a statistical or machine learning manner. In the implicit feature branch, a multi-modal deep learning model is used to process text information and picture information, such as Figure 3As shown, the multi-modal deep learning model includes a text processing model, an image processing model, a Transformer encoder, and an expert network. The text information is processed by the text processing model, and the picture information is processed by the image processing model. By sharing the same Transformer encoder, unified semantic features and visual features can be obtained in the same semantic space. Then, the semantic features and the visual features are input into the expert network, and a few expert networks are selectively activated, so as to specialize the semantic features and the visual features, and further extract specialized implicit features. For example, only superficial features can be extracted from the text expression and the picture. After the constraint training of the expert network, the semantic features and the visual features can be combined to tend to dig the implicit features, and the corresponding specialized implicit features are obtained. According to the specialized implicit features, the functionality, the aesthetics, and the creativity of the crowdfunding product of the crowdfunding project and the scores of multiple dimensions of the characteristics can be obtained. The structured features and the specialized implicit features are input into the feature fusion model to obtain the prediction success rate, and then the prediction success rate is analyzed to obtain the output result that can guide the decision.

[0036] The following describes each step in detail.

[0037] S100: Data acquisition and preprocessing.

[0038] First, data acquisition and preliminary cleaning are carried out for the directly quantifiable core fields (such as target amount, fundraising period, project category, number of supporters, social media association flag, etc.) on the crowdfunding platform. After obtaining the relevant information through the Internet crawler or official API (Application Programming Interface), the information is uniformly loaded into a relational database or a distributed storage system, and managed with the project ID as the index key. In order to eliminate unreasonable extreme values, the box plot detection method is used to identify outliers, and the truncation method or quantile correction method is used to process the abnormal values; for missing fields, KNN (K-Nearest Neighbor) interpolation is used to infer more reliable values according to the distance measure of the most similar samples. Then, in order to alleviate the problem of unstable model training caused by the too skewed distribution or the difference in magnitude of the numerical columns, logarithmic transformation or interval scaling (Min-Max Scaling) can be performed on the amount and number of numerical columns, so as to pull different features into a comparable range, and to improve the efficiency and accuracy of subsequent feature fusion.

[0039] Among unstructured data, we prioritized textual information from crowdfunding projects, including common text data sources such as project descriptions, update announcements, and supporter comments. After obtaining the raw text, we first used regular expressions to remove HTML tags, redundant symbols, and high-noise characters. Records with mixed languages ​​or very short content were filtered as appropriate to ensure text quality for subsequent modeling. In addition to text, we also collected image data, such as product prototypes, promotional posters, or concept designs displayed on crowdfunding pages. To ensure consistent input scale, all images were scaled to 224×224 pixels and histogram equalization was performed to enhance contrast and legibility. If sample size was limited, augmentation strategies such as random cropping and horizontal flipping could be added as needed to increase the model's robustness to varying shooting environments and angles.

[0040] S110: Inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model to screen and obtain structured features, wherein the hierarchical hybrid model includes a structured branch, an implicit feature branch and a feature fusion model.

[0041] Specifically, the structured features include numerical features and categorical features. For structured data, statistical methods can be used to screen numerical features such as amount, period, and number of people. Machine learning methods can be used to identify specific structured features in structured data, such as categorical features such as project number, project name, and project type, using convolutional neural networks or recurrent neural networks.

[0042] After preliminary cleaning and standardization of the structured data of the crowdfunding platform, we further use single-factor analysis, multi-factor interaction analysis, and causal verification to accurately identify the core structural features that most affect the success rate of the project.

[0043] S111: Analyze the correlation between the numerical features and the success or failure of the project.

[0044] First, a single-factor screening was performed using statistical methods to evaluate the correlation between each candidate structured feature and the success or failure of the crowdfunding project. For numerical features (such as target amount, fundraising period, or number of supporters), the Pearson or Spearman correlation coefficient was used to measure the linear or monotonic correlation between them and the success or failure of the crowdfunding project (0 / 1 labels, corresponding to failure / success). If there is a linear correlation between the numerical feature and the success or failure of the project, the absolute value of the slope of the linear correlation is used to determine whether the numerical feature is a potential key feature. If the absolute value of the slope is greater than the preset slope threshold, the numerical feature is determined to be a potential key feature; otherwise, it is an auxiliary feature.

[0045] S112: If the numerical feature and the success of the project show a non-linear correlation, the numerical feature is segmented and processed, and the chi-square test is used to analyze whether each segment of the numerical feature is significantly associated with the success rate of crowdfunding.

[0046] S113: The chi-square test is used to analyze whether the category type feature is significantly associated with the success rate of crowdfunding.

[0047] S114: If the analysis result shows that a structured feature is significantly associated with the success rate of crowdfunding, the structured feature is considered as a potential key feature.

[0048] If some numerical features are found to have a non-linear relationship or strong skew distribution, equal frequency binning or segmentation is used, and then the chi-square test is used to determine whether there is a significant correlation between the success rate. At the level of category type features (such as project category or social media association mark), the chi-square independence test is also used to measure the dependence between the feature value distribution and the success or failure of the crowdfunding project. When a feature shows a high correlation at a preset significance level (such as p<0.05), it is considered a potential key factor and is retained in subsequent interactive analysis.

[0049] S115: Random forest and light gradient boosting machine are used for multi-factor interactive analysis of structured features, and when a structured feature shows double high phenomenon in two multi-factor interactive analysis, the structured feature is considered as a potential key feature.

[0050] Specifically, after single-factor screening is completed, random forest and light gradient boosting machine are used for multi-factor interactive analysis of structured features. Random forest measures the contribution of each feature to sample purity improvement when node splitting based on Gini importance (importance), and the formula is as follows:

[0051]

[0052] Where T represents the set of all trees, N t is the number of samples at node t, N is the total number of samples, and ΔGini(t,f) represents the reduction of Gini index of feature f at node t.

[0053] Light gradient boosting machine evaluates the information gain (importance) brought by feature splitting through Split Value Gain, and the formula is as follows:

[0054]

[0055] Where L and R represent the sample set of the left and right child nodes after splitting, respectively, I represents the sample set of the current node, and g iis the gradient of sample i, h i is the second-order derivative of sample i, and λ represents the regularization parameter, which is used to prevent overfitting and control the complexity of the model.

[0056] The "double-high" phenomenon refers to a high importance score (the calculated result is greater than or equal to the corresponding preset threshold) in both multi-factor interaction analyses. When a structured feature exhibits a high importance score in both multi-factor interaction analyses, it can be determined to have robust discriminative power for crowdfunding outcomes and can be considered a potential key feature. At the same time, features that are only highly important in one multi-factor interaction analysis should also be identified and comprehensively evaluated in conjunction with subsequent causal verification.

[0057] S116: Constructing a directed acyclic graph using the PC algorithm. When a structural feature is mediated by only one potential key feature in the directed acyclic graph, the structural feature is considered as an auxiliary feature.

[0058] Specifically, to further eliminate confounding effects and multicollinearity between features, the Peter-Clark (PC) algorithm was used to analyze correlations between variables to identify possible causal relationships. A partially directed acyclic graph (PDAG) was then constructed to identify possible causal associations. When a feature was determined to be mediated solely by another, more critical feature in the causal structure, it was classified as an auxiliary factor to reduce feature redundancy and strengthen the critical pathway. Mediation refers to a variable (called a mediating variable or mediator) that acts as a bridge between two other variables, where one variable (the independent variable) influences the other (the dependent variable) through the mediating variable. This means that a potential key feature influences the success or failure of a crowdfunding project through a structural feature. In this case, the mediating variable is considered an auxiliary feature.

[0059] At this point, through this progressive screening and verification, we can clearly define the structured features that play a major role in the success or failure of crowdfunding projects, and mainly use these high-value and high-explanatory structured features in the subsequent fusion modeling with unstructured features (text and images), thereby ensuring the robustness and interpretability of the hierarchical hybrid model.

[0060] S120: Inputting the text information of the crowdfunding project into a text processing model to obtain semantic features, wherein the implicit feature branch includes a multimodal deep learning model, and the multimodal deep learning model includes a text processing model, an image processing model, and an expert network.

[0061] Specifically, the processed text information is input into a pre-trained BERT model to obtain sentence or word vector representations; when more fine-grained expression is needed, a subword-based encoding method can also be selected to obtain more rich semantic features. Text vectorization not only significantly reduces the high-dimensional noise of data, but also provides a unified text representation for the cross-modal alignment process in subsequent multi-modal deep learning models, enabling more efficient feature fusion and semantic interaction.

[0062] In S130, picture information of the crowdfunding project is input into an image processing model to obtain visual features. The text processing model and the image processing model share the same Transformer encoder, which is used for cross-modal parameter sharing and semantic alignment through a contrastive learning loss function.

[0063] Specifically, the ViT model or convolutional neural network is used to obtain image vector representations, which are saved in the multi-modal database and associated with the project ID, so that the subsequent model training can conveniently load the text and image features corresponding to each project in batches.

[0064] After completing the structured data cleaning, text vectorization and image encoding, all results can be stored in the multi-modal database, and the association management of the same project data of different types is realized through the project ID. The database table can be divided into "structured information table", "text embedding table" and "image embedding table", and the different tables are connected with each other by project ID. This design not only enables batch query and one-key access to multi-modal input in the subsequent training stage, but also ensures the convenient access of incremental data. When new projects appear on the crowdfunding platform or existing projects are updated, only the corresponding records need to be added and the reference to the project ID is retained, so that they can be seamlessly integrated into the overall prediction process. Through such a data architecture, the unstructured noise of text and image is significantly reduced, and the structured features are also reasonably standardized in magnitude and dimension, thereby providing a clean and high-quality input data basis for subsequent multi-modal feature recognition, implicit feature quantization and hierarchical hybrid model construction.

[0065] In the successful prediction of crowdfunding projects, text and image and other unstructured data in the project description carry rich implicit information, such as the functionality, aesthetics and creativity of the project. These implicit features are difficult to quantify directly, but they have an important influence on the attractiveness of the project, user engagement and the final success rate. In order to accurately quantify these implicit features, a multi-modal deep learning model is used to realize deep analysis and feature extraction of text and image data through semantic embedding, cross-modal alignment and expert network division.

[0066] In the multi-modal deep learning model, the text part is encoded using the BERT pre-training model. BERT captures the context information in the text through a bidirectional Transformer architecture, generating high-quality semantic vectors. These vectors can effectively represent the semantic content of the text, especially the descriptions related to functionality, aesthetics, and creativity. The image part uses the ViT pre-training model to extract visual features. The ViT model divides the image into multiple patches and encodes them through a Transformer architecture, generating visual feature representations of the image. The advantage of the ViT model is that it can capture global information in the image, especially suitable for processing complex visual content such as design drawings, product prototypes, etc.

[0067] To ensure that the text and image are expressed consistently in a unified semantic space, the multi-modal deep learning model introduces a shared Transformer encoder. This encoder receives the embedding vectors of the text and image and performs cross-modal parameter sharing and semantic alignment through a contrastive loss function. The goal of contrastive learning is to maximize the similarity of positive sample pairs (i.e., the text and image of the same project) while minimizing the similarity of negative sample pairs (i.e., the text and image of different projects). The specific formula is as follows:

[0068]

[0069] where τ = 0.07 is the temperature parameter, used to adjust the smoothness of the similarity distribution; sim is the cosine similarity, used to measure the similarity between the text and image embedding vectors; z text is the embedding vector of the text, z image is the embedding vector of the image (positive sample), is the embedding vector of other images (negative sample), where j ranges from 1 to B, and B is the batch size (number of samples). In this way, the multi-modal deep learning model can ensure that the text and image are aligned in a unified semantic space, providing a consistent representation basis for subsequent feature quantization.

[0070] S140: According to the semantic features and visual features, obtain the activation probability of multiple expert networks, including visual experts, language experts, cross-modal experts, functionality experts, aesthetics experts, and creativity experts. The expert networks are functionally specialized through constrained training.

[0071] S150: According to the activation probability, activate the corresponding expert network sparsely, and obtain the specialized hidden features corresponding to the expert network.

[0072] Specifically, on the basis of semantic alignment, the multi-modal deep learning model introduces multiple expert networks that process different types of implicit features through a dynamic routing mechanism. Through the division of labor among multiple expert networks, each expert network focuses on processing a specific type of feature. The expert networks in the multi-modal deep learning model include visual experts, language experts, cross-modal experts, and functionally specialized experts, aesthetic experts, and creative experts. The activation probability of each expert network is calculated through a lightweight routing network. For example, when inputting picture information, the language expert will not be activated, while the visual expert and the cross-modal expert will be activated, and the functionally specialized experts, aesthetic experts, and creative experts will also be activated. However, the activation levels of these three networks will be different; when inputting picture information, the aesthetic expert will be more highly activated, while when inputting text information, the functional network will be more highly activated.

[0073] Specifically, the routing network generates routing weights based on the input semantic features and visual features, thereby screening the activated expert networks. To improve computational efficiency, the multi-modal deep learning model adopts a Top-K sparse activation strategy, that is, only the most relevant expert networks are activated for each sample. The routing decision function is as follows:

[0074] g(x) = softmax(W g x) p

[0075] where the softmax function is used in the output layer of a multi-classification problem, which can map the input to a probability distribution, so that the sum of all outputs is 1. W p ∈ R 1024×256 is a projection matrix used to map input features to a low-dimensional space, and W g ∈ R 256×3 is a routing weight used to calculate the activation probability of each expert network, and GELU (Gaussian Error Linear Unit) is an activation function that can better capture nonlinear features. The result of this formula is the probability distribution obtained after a series of transformations of the input x.

[0076] To further improve the feature extraction capability of the expert network, the multi-modal deep learning model functionally specializes the expert network through constraint training. Specifically, the functional expert focuses on extracting features related to project practicality, technical implementation, etc.; the aesthetic expert focuses on extracting features related to design aesthetics, visual appeal, etc.; and the creative expert focuses on extracting features related to innovation, uniqueness, etc. Through this division of labor, the multi-modal deep learning model can more accurately quantify different types of implicit features.

[0077] ​S151: Obtain a plurality of dimension scores according to the specialized implicit features, wherein the dimension scores include a functionality score, an aesthetic score, and a creativity score.

[0078] Specifically, based on the obtained specialized implicit features, the multi-modal deep learning model further maps the outputs of different types of expert networks to the score dimensions of functionality, aesthetics, and creativity. The functionality score is normalized using a Sigmoid function to map the output to the interval [0, 1]. The characteristics of the Sigmoid function make it suitable for processing binary classification or probability estimation problems, and it can effectively represent the functionality strength of the project. The aesthetic score is normalized using a Tanh function to map the output to the interval [-1, 1]. The symmetry of the Tanh function makes it suitable for capturing the bipolar distribution of aesthetic features, and it can effectively represent the degree of beauty of the design. The creativity score is normalized using a Softplus function, and the smoothness and non-negativity of the Softplus function make it suitable for capturing the differences in creativity features. The specific formulas are as follows:

[0079] f func =Sigmoid(W f z func +b f )

[0080] f aes =Tanh(W a z aes +b a )

[0081] f creat =Softplus(W c z creat +b c )

[0082] where W f , W a , W c are the weights corresponding to the functions, z func , z aes , z creat are input vectors, and b f , b a , b cis a bias term. According to the dimension score, the crowdfunding project is screened to obtain a target crowdfunding project, which can help the decision maker to quickly screen the project meeting the specific standard. According to the score, the decision maker can reasonably allocate resources, and can also recommend the project with high aesthetic or creative score according to the user's interest by analyzing the user's preference, so as to improve the user experience. Moreover, through the multi-dimensional score of the crowdfunding project, the success rate that the crowdfunding project can achieve can be intuitively felt. For example, if the scores of a crowdfunding project in functionality, aesthetics and creativity are all high, the prospects of the project are large, and the probability of being accepted by the public and being favored by the investors will be greatly improved, thus promoting the final crowdfunding success of the project.

[0083] S160: inputting the structured feature and the specialized implicit feature into a feature fusion model to obtain a predicted success rate. The feature fusion model includes a gated cross-attention mechanism, a bilinear pooling and a multi-layer perception.

[0084] Specifically, the feature fusion layer is responsible for deeply integrating the structured feature and the specialized implicit feature extracted by the multi-modal deep learning model, capturing the complex interaction relationship between the two, so as to provide high-quality input representation for the crowdfunding project success rate prediction. Through the gated cross-attention mechanism, the advantages of the two fusion modes are dynamically combined, in which the bilinear pooling captures the explicit interaction mode, and the multi-layer perception models the deep nonlinear relationship. Thus, the structured feature and the aesthetic, creative and functional specialized implicit features output by the multi-modal deep learning model are coupled, and then the success rate is predicted according to the fused features.

[0085] S161: inputting the structured feature and the specialized implicit feature into a gated cross-attention mechanism to obtain a weighted weight and an attention matrix.

[0086] Specifically, the gated cross-attention mechanism dynamically allocates weights to coordinate the two feature fusion modes of the bilinear pooling and the multi-layer perception. The gated cross-attention mechanism generates weights according to the input features, realizes the weighted combination of the two branches, and adaptively integrates the structured feature and the specialized implicit feature. This mechanism not only retains the nonlinear expression ability of the multi-layer perception, but also utilizes the explicit interaction modeling advantage of the bilinear pooling, realizing the complementary advantages of feature fusion.

[0087] S162: inputting the structured feature and the specialized implicit feature into a bilinear pooling to obtain a first fused feature.

[0088] Bilinear pooling is an explicit second-order feature interaction method that can capture the product effect between different feature dimensions. In crowdfunding project evaluation, certain feature combinations may have a significant impact on the prediction of project success rate. For example, the combination of "target amount x creativity" or "social communication x aesthetics" may reflect the potential success probability of the project more accurately than a single feature. Bilinear pooling can more accurately capture these high-order interaction relationships by explicitly calculating the product effect of these feature combinations.

[0089]

[0090] where vec() is to flatten the outer product matrix into a vector form, z struct is the structured feature, z limoe is the specialized implicit feature, W b is the dimension reduction matrix, used to control the computational complexity.

[0091] S163: input the structured feature and the specialized implicit feature into the multi-layer perception to obtain a second fused feature.

[0092] The multi-layer perception is used to perform nonlinear transformation on the features to capture complex feature relationships. The multi-layer perception includes an input layer, multiple hidden layers, and an output layer. The hidden layer is used to extract high-level features of the input data, perform weighted summation on the input data, and then introduce nonlinearity through an activation function. The multi-layer perception measures the accuracy of the prediction through a binary cross-entropy loss function and performs backpropagation optimization to ensure that the model can continuously adjust the parameters during the training process to improve the prediction accuracy. The following is the formula for the multi-layer perception to process the fusion of structured data and specialized implicit features.

[0093] h mlp = GELU(W2·GELU(W1·[z struct ; z limoe ]))

[0094] where W1, W2 are learnable parameters, and the GELU activation function enhances the non-linear representation capability, z struct is the structured feature, z limoe is the specialized implicit feature.

[0095] S164: according to the weighted weight, the first fused feature and the second fused feature, weighted to obtain a final fused feature.

[0096] The gating cross-attention mechanism dynamically allocates weights to coordinate the outputs of bilinear pooling and MLP. The gating unit generates weights according to the input features to realize the weighted combination of the two branches.

[0097] For example, the fusion method can be h = h mlp + h bilinσ(Wf+b), where h is the final fused feature, h mlp is the second fused feature output by the multi-layer perceptron, h bilin is the first fused feature output by the bilinear pooling, W and b are parameters (which can be regarded as dynamic weights) of the gated cross-attention mechanism, f is the input feature, and σ() is an activation function.

[0098] The fusion manner can also be h = a h mlp + (1-a) h bilin , where h is the final fused feature, h mlp is the second fused feature output by the multi-layer perceptron, h bilin is the first fused feature output by the bilinear pooling, and a is a dynamic weight allocated by the gated cross-attention mechanism.

[0099] S165: Obtain a predicted success rate according to the final fused feature.

[0100] Specifically, the output layer of the multi-layer perceptron can be used to input the final fused feature into the output layer, and the output result can be output according to requirements, for example, the class probability of the classification task (success or failure).

[0101] S170: Obtain a sparse attention matrix by using a sparsity constraint according to the attention matrix output by the gated cross-attention mechanism model.

[0102] Specifically, the multi-modal deep learning model has a sparse activation characteristic in the feature extraction layer. The sparsity constraint (such as L1 regularization or Top-K pruning strategy) is further introduced in the feature fusion layer. By introducing the sparsity constraint, the attention weight is forced to concentrate on the high-weight feature, which improves the feature selection accuracy and the model explainability. In the prediction explanation layer, the sparse attention matrix is used to intuitively display the high-weight feature, which is used for feature importance attribution, and explicit rules are extracted therefrom to automatically induce decision rules, thereby enhancing the explainability and operability of the model.

[0103] L1 regularization highlights the core elements in the multi-dimensional high-order interaction by forcing most of the weights to approach zero and only retaining a small number of feature combinations with the highest contribution, and the formula is as follows:

[0104]

[0105] L1 regularization is applied to the attention matrix A ∈ R d×d , where λ1 is a sparsity regularization coefficient, and the optimal value (such as λ1 = 0.3) is determined by grid search.

[0106] The Top-K pruning strategy further simplifies the model structure by retaining the K largest elements in the attention matrix. The pruned attention matrix can be used as an explicit representation of the "relationship network", making it easier for users to understand the core features that the model focuses on during decision-making. The formula is as follows:

[0107]

[0108] For the attention matrix, the top K elements with the highest weights are retained, and the rest are set to zero. The value of K is dynamically adjusted according to task requirements (e.g., K = 3).

[0109] S180: Use the Apriori algorithm to extract high-frequency feature combinations from the sparse attention matrix, and verify the significance of the high-frequency feature combinations through the chi-square test.

[0110] S190: When the extracted high-frequency feature combinations are significant, output the high-frequency feature combinations and the prediction success rate as decision rules.

[0111] Decision rule extraction: The Apriori algorithm is used to extract high-frequency feature combinations (e.g., support > 0.6, confidence > 0.8) from the sparse attention matrix, and the significance of the rules is verified through the chi-square test (p < 0.01). The output decision rule is as follows:

[0112] IF(A 价格,功能性 >0.7)AND(A 美观性,社交传播 >0.5)THEN success rate > 70

[0113] 0.7 and 0.5 represent the threshold or threshold of the features obtained through data analysis and algorithm, meaning that when the price and functionality score of the product exceed 0.7, and the aesthetic and social communication score exceed 0.5, the association rule considers that these conditions are met, and the prediction success rate is greater than 70%.

[0114] Readability expression and decision support: Through the above rule mining process, the complex model decision-making process can be converted into a series of easy-to-understand "IF-THEN" rules. For example, "IF (price < X AND functionality > Y) THEN (success probability > Z%) " such rules not only intuitive and easy to understand, but also provide clear decision-making basis for crowdfunding project initiators or evaluators. These rules can help project initiators make targeted adjustments during the project design phase, thereby improving the success rate of the project.

[0115] In addition to the above, during the model training process, in order to balance the complexity of different modal information and the learning goal of multi-stage feature interaction, a progressive training strategy is adopted, and each model is gradually unfrozen and jointly optimized in stages. Through this method, the training difficulty in the initial stage can be effectively reduced, the representation ability of each model can be stabilized, and higher prediction accuracy and interpretability can be achieved when integrating multi-modal features in the later stage.

[0116] The multi-modal data driven product crowdfunding success rate prediction method further includes model progressive training; the model progressive training includes:

[0117] The text processing model and the image processing model in the multi-modal deep learning model are pre-trained separately.

[0118] The multiple expert networks in the multi-modal deep learning model are constrainedly trained.

[0119] The parameters of the multi-modal deep learning model are frozen, and the structured branch and the feature fusion model are trained.

[0120] Specifically, in the initial stage, first, most of the parameters of the multi-modal deep learning model are fixed, and the structured branch and the feature fusion layer are mainly trained, so that the model can first master the numerical and categorical data of the crowdfunding project and establish a basic success rate prediction ability. In this stage, the text and image features are usually obtained in a pre-training manner and remain in a frozen state, avoiding simultaneous multi-modal alignment and structured feature learning, which can cause unstable training or gradient conflict. By optimizing only the structured branch and the feature fusion layer, the model can quickly converge in the early stage and learn the main linear or nonlinear interaction relationship, laying a good foundation for subsequent unfreezing of the implicit feature branch and parallel training.

[0121] The top layer parameters in the multi-modal deep learning model are unfrozen and jointly optimized with the feature fusion model, wherein the top layer parameters are the parameters in the Transformer encoder and the parameters before the Transformer encoder in the multi-modal deep learning model.

[0122] Specifically, when the structured branch has been relatively stable and the overall prediction accuracy reaches a certain level, the top-level parameters of the multi-modal deep learning model are gradually unfrozen, i.e., the parameters located at the uppermost layer of the network architecture in the model. These parameters are responsible for processing the features processed by the previous layers (such as feature extraction layers) and finally generating the output of the model. After unfreezing the top-level parameters, they start to be jointly optimized with the feature fusion layer. At this time, the model can more deeply align the text and image data across modalities, and strengthen the semantic consistency of text and image through contrastive learning, so as to establish a closer relationship between the implicit features such as functionality, aesthetics, and creativity and the structured features. In this stage, the weights of modules such as gated cross-attention mechanism and bilinear pooling are also constantly updated, so that the multi-modal interaction is more in line with the real market feedback mechanism of crowdfunding projects. With the gradual unfreezing and fine-tuning of the implicit feature branch, the model will further explore the potential influence of product appearance and text description on the success rate of crowdfunding, while maintaining the advantages of structured features, and integrating the information of unstructured features into the decision-making process.

[0123] All frozen parameters are unfrozen, and the layered hybrid model is fine-tuned using perturbed data, wherein the perturbed data is obtained using an adversarial data augmentation method.

[0124] Specifically, in the last stage, all frozen network parameters are unlocked for end-to-end unified training. To enhance the robustness of the model under data distribution mutation or noise input, an adversarial data augmentation method can be used to add a small perturbation to the text or image input, or to randomly jitter the structured features. In this way, the model can learn to maintain relatively stable prediction performance and explainability in extreme or uncertain environments. Finally, through progressive training, a stable learning process from structured dominance to multi-modal full fusion is gradually realized, which not only can alleviate the risk of gradient explosion or disappearance of the multi-modal deep learning model in the early training, but also can polish a high-precision and explainable crowdfunding success rate prediction model in the later stage, providing comprehensive and reliable decision support for crowdfunding platforms and project initiators.

[0125] To solve the high information entropy loss and insufficient prediction accuracy caused by the "modality split effect", the present application realizes cross-modal semantic alignment at the text and image level based on a multi-modal deep learning model, and combines mechanisms such as gated cross-attention and bilinear pooling to deeply couple structured features (such as target amount and fundraising period) with abstract functional, aesthetic, and creative features. Compared with models such as MMF-CNN that simply concatenate images and text, the present application effectively reduces the information entropy loss rate and reduces the prediction error of aesthetic-related features by nearly 10%-15% in experiments, showing higher precision and generalization ability in creative or aesthetic projects.

[0126] To address the problem of disconnection between explainability and business needs, the application introduces sparsity constraints and association rule mining in the prediction and explanation layer of the hierarchical hybrid model, breaking through the single limitation of methods such as LIME and SHAP that can only give numerical feature importance. Unlike ProtoPNet, which requires manual annotation of a large number of samples, the application automatically highlights core weights through Top-K pruning or L1 regularization during training, which can help project initiators or platform managers quickly identify weaknesses in appearance design, copywriting, or pricing mechanisms in crowdfunding practice, improving product iteration efficiency and reducing market investment risk.

[0127] To address the risk of model failure in dynamic market environments, the application uses a progressive training and incremental learning strategy to adapt to changes in the distribution of crowdfunding project types in real time by unfreezing the multi-modal deep learning model and feature fusion layer at different times. Traditional static modeling methods such as LightGBM+ genetic algorithm can have a precision decay of up to 23% when the proportion of project categories changes significantly. However, the application can quickly optimize the sparse attention matrix and maintain high prediction accuracy after incremental data injection, significantly reducing the cost of secondary optimization due to model mismatch. In addition, the feature stability index (FSI) of the model is significantly improved, allowing it to more flexibly capture market features of emerging categories such as AI hardware, adapting to the needs of the rapidly iterating crowdfunding ecosystem.

[0128] In summary, the application has made innovative breakthroughs in "multi-modal feature fusion", "explainability enhancement", and "dynamic market adaptation", not only significantly improving the prediction accuracy of crowdfunding project success rates, but also clearly presenting the contributions of each modality feature as executable rules, providing valuable technical support for innovation verification and market decision-making for crowdfunding platforms and project initiators. Compared with existing technologies, the application has higher levels of practicality and commercial value, fully meeting the requirements of the crowdfunding ecosystem for short cycle, high accuracy, and operability.

[0129] In summary, compared with existing technologies, the application has the following benefits:

[0130] 1. A multi-modal expert network is constructed based on a multi-modal deep learning model, which realizes semantic alignment of text and images through dynamic routing and sparse activation mechanism, effectively improving the quantification ability of product functionality, aesthetics and creativity and other implicit features.

[0131] 2. Structured features and unstructured abstract features are integrated in the hierarchical hybrid model, which uses a gating cross-attention mechanism, bilinear pooling, and multi-layer perceptron to mine multi-dimensional interaction relationships, significantly enhancing the prediction accuracy and stability of crowdfunding project success rates.

[0132] 3. Introducing sparsity constraints and rule extraction methods can extract high-confidence "IF-THEN" decision rules from attention matrices, improving the interpretability and operability of crowdfunding prediction results.

[0133] 4. Using progressive training strategies and feature evolution mechanisms, the model can maintain dynamic adaptability when market structure changes or project types fluctuate, ensuring continuous and stable prediction performance.

[0134] The above method is used for practical application test, and the test process and results are as follows.

[0135] To specifically demonstrate the practical application effect of the multi-modal crowdfunding success rate prediction method proposed in the present application in real crowdfunding projects, this embodiment takes five typical "design-toy" projects on a certain platform as the research object, and completes the whole process analysis of data collection and preprocessing, structured feature analysis, unstructured implicit feature quantization, success rate prediction and interpretable output according to the technical route of the present application.

[0136] Firstly, the structured fields of each project page are scraped from a certain platform, including project target amount, fundraising period, number of supporters, whether associated with social media and other information, and stored in a multi-modal database with project ID as the primary key. After boxplot outlier identification on continuous variables such as target amount and number of supporters, truncated or quantile method is used for outlier correction, and KNN is used to impute missing data. To avoid the influence of numerical feature scale difference on model stability, log transformation and interval scaling techniques are also used for normalization.

[0137] Secondly, in the unstructured data preprocessing stage, BERT and ViT are used to extract feature vectors from project description text and product design images provided on the crowdfunding page. Text data is removed of HTML tags, special symbols and noise characters through regular expressions, and is encoded by BERT model to form a unified semantic vector representation; image data is uniformly sized and enhanced by histogram equalization to enhance details, and then generates image vector feature representation through ViT, and then completes cross-modal semantic alignment through the shared Transformer encoder of the multi-modal deep learning model. The three types of implicit features of the project are precisely quantified by functional experts, aesthetic experts and creative experts.

[0138] Taking the "Modular Magnetic Building Blocks" project as an example, its functionality, aesthetics and creativity scores are of high level. This project clearly emphasizes its unique magnetic module design, good aesthetic display and clear functional advantages, which are accurately captured by the model and mapped to high-quality implicit feature scores.

[0139] Subsequently, the structured factor identification phase is entered. First, single-factor analysis is performed to test the correlation (Pearson and Chi-square tests) of structured variables such as target amount, fundraising period, and number of supporters with the success rate of crowdfunding, to confirm their significant relationship with the success rate. Multi-factor analysis uses random forest and LightGBM models to further confirm high-correlation variables based on Gini Importance and Gain, such as the number of supporters and social media association indicators, which have a high contribution to the success rate and show robust importance. In the causal verification stage, PC algorithm is used to further exclude redundant or mediated variables, and the final structured feature set is determined.

[0140] Next, in the hierarchical mixed model construction and prediction phase, structured features and implicit features quantified by multi-modal deep learning models are deeply integrated through a gated cross-attention mechanism and a bilinear pooling. The model automatically adjusts the attention allocation weight through a dynamic routing mechanism, enabling high-order interaction between structured features (such as the number of supporters and target amount) and unstructured features (such as creativity score), and outputting the predicted success probability of each project.

[0141] Taking the "magnetic levitation educational toy" project as an example, its text description quality is 0.89, image clarity is 0.92, functional score is 0.94, aesthetic score is 0.81, and creativity score is 0.88. The model predicts a success probability of 0.95 after considering all factors, and the actual crowdfunding amount reaches 56800 US dollars, far exceeding the initial target of 25000 US dollars, verifying the high accuracy of the model.

[0142] At the same time, to further enhance the explainability and practicality of the model, a sparse attention matrix is introduced in the prediction process, and algorithm (such as Apriori) is used to extract causal rule examples from the model output for practical operation:

[0143] IF (functional score > 0.85) AND (aesthetic score > 0.75) AND (target amount ≤ 20000) THEN (predicted success rate > 0.90)

[0144] This rule is verified to be effective in the "modular magnetic building blocks" and "creative glow-in-the-dark puzzle" projects. As shown in Table 1, taking the "creative glow-in-the-dark puzzle" project as an example, its functionality is 0.88, aesthetics is 0.76, and target amount is 18000, which meets the above rule conditions, and the actual fundraising amount reaches 40200, which is more than twice the target, highlighting the practicality and accuracy of the model's prediction rules. The "multi-dimensional geometric puzzle" project has a relatively low implicit feature score (creativity 0.59, aesthetics 0.51), with a predicted success rate of only 0.52, and the actual fundraising failed to reach the target amount, demonstrating the reliability of the model in predicting failure risks.

[0145] Table 1 toy crowdfunding project implicit feature score and predicted success rate

[0146] Project Name Functionality Rating Aesthetics Rating Creativity Rating Predicted Success Actual Success Modular magnetic building blocks 0.93 0.80 0.90 0.94 Yes Creative glow-in-the-dark puzzle 0.88 0.76 0.84 0.89 Yes Smart educational train set 0.75 0.60 0.70 0.66 Yes Custom-programmable gyroscopes 0.91 0.79 0.86 0.90 Yes Multi-dimensional geometric puzzle 0.68 0.51 0.59 0.52 No

[0147] From the above implementation process and case verification results, it can be seen that the method provided by the application can significantly improve the success rate prediction accuracy of the crowdfunding project and the effectiveness of the landing application, and by generating clear and operable decision rules, it provides a comprehensive, intelligent and efficient auxiliary decision tool for crowdfunding platforms and project initiators, and has wide application and promotion value.

[0148] As shown in Figure 2 The multi-modal data-driven product crowdfunding success rate prediction system provided by the embodiment of the application comprises:

[0149] The feature extraction layer is configured to input the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model, and filter to obtain structured features, wherein the hierarchical hybrid model comprises a structured branch, an implicit feature branch and a feature fusion model; is also configured to input the text information of the crowdfunding project into a text processing model to obtain semantic features, wherein the implicit feature branch comprises a multi-modal deep learning model, and the multi-modal deep learning model comprises a text processing model, an image processing model and an expert network; is also configured to input the picture information of the crowdfunding project into the image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder is configured to share parameters and align semantics across modalities through a contrastive learning loss function; is also configured to obtain the activation probability of the plurality of expert networks according to the semantic features and the visual features, wherein the expert networks comprise a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert and a creative expert; is also configured to activate the corresponding expert network according to the activation probability, and obtain the specialized implicit features corresponding to the expert network.

[0150] The feature fusion layer is configured to input the structured features and the specialized implicit features into the feature fusion model to obtain the predicted success rate.

[0151] The prediction explanation layer is configured to obtain a sparse attention matrix by using a sparsity constraint according to the attention matrix output by the gated cross-attention mechanism model; is also configured to extract a high-frequency feature combination from the sparse attention matrix by using an Apriori algorithm, and verify the significance of the high-frequency feature combination by a chi-square test; is also configured to output the high-frequency feature combination and the predicted success rate as a decision rule when the extracted high-frequency feature combination is significant.

[0152] In the embodiment, the beneficial effects of the multi-modal data driven product crowdfunding success rate prediction system are similar to those of the multi-modal data driven product crowdfunding success rate prediction method described above, and will not be repeated here.

[0153] The electronic device provided by the embodiment of the present application comprises a memory and a processor; the memory is used to store a computer program; and the processor is used to implement the multi-modal data driven product crowdfunding success rate prediction method as described above when executing the computer program.

[0154] The computer readable storage medium provided by the embodiment of the present application has the computer program stored thereon, and when the computer program is executed by the processor, the multi-modal data driven product crowdfunding success rate prediction method as described above is implemented.

[0155] In the embodiment, the beneficial effects of the electronic device and the computer readable storage medium are similar to those of the multi-modal data driven product crowdfunding success rate prediction method described above, and will not be repeated here.

[0156] It should be noted that, in this paper, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0157] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multimodal data-driven product crowdfunding success rate prediction method, characterized by: include: Inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model to screen and obtain structured features, wherein the hierarchical hybrid model includes a structured branch, an implicit feature branch, and a feature fusion model; Inputting text information of the crowdfunding project into a text processing model to obtain semantic features, wherein the implicit feature branch includes a multimodal deep learning model, and the multimodal deep learning model includes a text processing model, an image processing model, and an expert network; Inputting image information of crowdfunding projects into an image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder uses a contrastive learning loss function for cross-modal parameter sharing and semantic alignment; Obtaining activation probabilities of multiple expert networks based on the semantic features and visual features, wherein the expert networks include visual experts, language experts, cross-modal experts, functional experts, aesthetic experts, and creative experts; According to the activation probability, sparsely activate the corresponding expert network to obtain specialized implicit features corresponding to the expert network; Inputting the structured features and specialized latent features into a feature fusion model to obtain a prediction success rate; The feature fusion model includes a gated cross attention mechanism, bilinear pooling, and a multi-layer perceptron; Inputting the structured features and specialized latent features into the feature fusion model to obtain the prediction success rate includes: Inputting the structured features and specialized latent features into the gated cross attention mechanism to obtain weighted weights and attention matrix; Inputting the structured features and the specialized latent features into bilinear pooling to obtain a first fusion feature; Inputting the structured features and the specialized latent features into a multi-layer perceptron to obtain a second fusion feature; Obtaining a final fusion feature by weighting according to the weighted weight, the first fusion feature, and the second fusion feature; According to the final fusion features, the prediction success rate is obtained.

2. The multimodal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that: The structured features include numerical features and categorical features; The structured data of the crowdfunding project is input into the structured branch of the hierarchical hybrid model to screen out structured features, including: Analyze the correlation between the numerical characteristics and project success or failure; If there is a nonlinear correlation between the numerical feature and the success of the project, the numerical feature is segmented and the chi-square test is used to analyze whether there is a significant correlation between each segment of the numerical feature and the crowdfunding success rate; The chi-square test was used to analyze whether there was a significant correlation between categorical characteristics and crowdfunding success rate; If the analysis results show that a structured feature is significantly associated with the crowdfunding success rate, then the structured feature is considered a potential key feature.

3. The multimodal data-driven product crowdfunding success rate prediction method according to claim 2, characterized in that: The method of inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model and screening to obtain structured features further includes: Random forest and lightweight gradient boosting machine are used to perform multi-factor interaction analysis on structured features. When a structured feature shows a high double phenomenon in two multi-factor interaction analyses, it is considered as a potential key feature. The PC algorithm is used to construct a directed acyclic graph. When a structural feature is mediated by only one potential key feature in the directed acyclic graph, the structural feature is regarded as an auxiliary feature.

4. The multimodal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that: The expert network is functionally specialized through constrained training.

5. The multimodal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that: After sparsely activating the corresponding expert network according to the activation probability to obtain specialized implicit features corresponding to the expert network, the method further includes: Based on the specialized implicit features, multiple dimensional scores are obtained, wherein the dimensional scores include functionality scores, aesthetic scores, and creativity scores.

6. The multimodal data-driven product crowdfunding success rate prediction method according to claim 5, characterized in that: The multiple dimension scores obtained based on the specialized implicit features include: Sigmoid function was used to normalize the functional scores; The Tanh function was used to normalize the aesthetic scores; The Softplus function was used to normalize the creativity scores.

7. The multimodal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that: After inputting the structured features and the specialized latent features into the feature fusion model to obtain the prediction success rate, the method further includes: According to the attention matrix output by the gated cross attention mechanism model, a sparse attention matrix is ​​obtained by using sparsity constraints; The Apriori algorithm is used to extract high-frequency feature combinations from the sparse attention matrix, and the significance of the high-frequency feature combinations is verified by the chi-square test; When the extracted high-frequency feature combination is significant, the high-frequency feature combination and the prediction success rate are output as decision rules.

8. The multimodal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that: It also includes model progressive training; The model progressive training includes: Separately pre-train the text processing model and image processing model in the multimodal deep learning model; Constrained training of multiple expert networks in multimodal deep learning models; Freeze the parameters of the multimodal deep learning model and train the structured branch and feature fusion model; Unfreeze the top-level parameters in the multimodal deep learning model and perform joint optimization training with the feature fusion model, where the top-level parameters are the parameters in the Transformer encoder and the parameters before the Transformer encoder in the multimodal deep learning model; Unfreeze all frozen parameters and fine-tune the hierarchical hybrid model using perturbation data obtained using adversarial data augmentation methods.

9. A multimodal data-driven product crowdfunding success rate prediction system, characterized by: include: A feature extraction layer, configured to input the structured data of the crowdfunding project into a structured branch of a hierarchical hybrid model to filter out structured features, wherein the hierarchical hybrid model includes a structured branch, an implicit feature branch, and a feature fusion model; It is also used to input the text information of the crowdfunding project into a text processing model to obtain semantic features, wherein the implicit feature branch includes a multimodal deep learning model, and the multimodal deep learning model includes a text processing model, an image processing model and an expert network; It is also used to input the image information of the crowdfunding project into the image processing model to obtain visual features, wherein the text processing model and the image processing model share the same Transformer encoder, and the Transformer encoder uses a contrastive learning loss function for cross-modal parameter sharing and semantic alignment; The method is further configured to obtain activation probabilities of multiple expert networks based on the semantic features and visual features, wherein the expert networks include visual experts, language experts, cross-modal experts, functional experts, aesthetic experts, and creative experts; It is also used to sparsely activate the corresponding expert network according to the activation probability to obtain specialized implicit features corresponding to the expert network; A feature fusion layer is used to input the structured features and specialized latent features into a feature fusion model to obtain a prediction success rate; The feature fusion model includes a gated cross attention mechanism, bilinear pooling, and a multi-layer perceptron; Inputting the structured features and specialized latent features into the feature fusion model to obtain the prediction success rate includes: Inputting the structured features and specialized latent features into the gated cross attention mechanism to obtain weighted weights and attention matrix; Inputting the structured features and the specialized latent features into bilinear pooling to obtain a first fusion feature; Inputting the structured features and the specialized latent features into a multi-layer perceptron to obtain a second fusion feature; Obtaining a final fusion feature by weighting according to the weighted weight, the first fusion feature, and the second fusion feature; According to the final fusion features, the prediction success rate is obtained.

Images