Product crowd funding success rate prediction method and system driven by multi-modal data

Abstract

The invention provides a multi-modal data-driven product crowd funding success rate prediction method and system, and relates to the technical field of deep learning. The method comprises the following steps: constructing a hierarchical hybrid model, and performing structured branching and screening on structured data to obtain structured features; respectively processing the text information and the picture information through a text processing model and an image processing model to obtain semantic features and visual features; according to the semantic features and the visual features, activation probabilities of a plurality of expert networks can be obtained, the expert network corresponding to the larger activation probability is selected for activation, and if the crowd funding project focuses on aesthetics-related products, aesthetic experts or creative experts and the like are activated to obtain specific implicit features; and inputting the specific implicit features and the structured features into a feature fusion model to obtain a prediction success rate. When unstructured data closely related to product appearance and aesthetic features are processed, the features can be highlighted and strengthened through an expert network, and the prediction accuracy of aesthetic crowdfunding projects is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and in particular, to a method and system for predicting the success rate of product crowdfunding driven by multimodal data. Background Art

[0002] As an innovative financing model, crowdfunding pools small amounts of funds from individuals or organizations through an Internet platform to support creative projects, startup enterprises, and public welfare activities. Production only starts after obtaining financial support, significantly reducing the upfront financial pressure on project initiators and enabling more creative and potential projects to be launched. At the same time, crowdfunding can also be an effective tool for testing market demand, helping with product iteration and update by evaluating the market's acceptance of new products and user feedback.

[0003] Traditional structured modeling evaluation methods (such as the random forest model used by Mollick) or traditional expert-scoring-based evaluation methods are limited to structured variables in limited dimensions such as the target amount and fundraising cycle, and cannot handle unstructured data closely related to product appearance and aesthetic features, resulting in a relatively high prediction error for aesthetic crowdfunding projects and making it difficult to accurately reflect users' true evaluations of creative projects. Summary of the Invention

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

[0005] To solve the above problem, in a first aspect, the present invention provides a method for predicting the success rate of product crowdfunding driven by multimodal data, including:

[0006] Inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model, and screening to obtain structured features, where the hierarchical hybrid model includes a structured branch, an implicit feature branch, and a feature fusion model;

[0007] Inputting the text information of the crowdfunding project into the text processing model to obtain semantic features, where 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;

[0008] Inputting the image information of the crowdfunding project into the image processing model to obtain visual features, where 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] Based on the semantic features and visual features, obtain the activation probabilities of multiple expert networks, where the expert networks include visual experts, language experts, cross-modal experts, functional experts, aesthetic experts, and creative experts;

[0010] Based on the activation probabilities, sparsely activate the corresponding expert networks to obtain specialized implicit features corresponding to the expert networks;

[0011] Input the structured features and specialized implicit features into a feature fusion model to obtain the prediction success rate.

[0012] In a second aspect, the present invention also provides a multi-modal data-driven product crowdfunding success rate prediction system, including:

[0013] A feature extraction layer for inputting the structured data of a crowdfunding project into the structured branch of a hierarchical hybrid model, and screening to obtain structured features, where the hierarchical hybrid model includes a structured branch, an implicit feature branch, and a feature fusion model;

[0014] It is also used to input the text information of the crowdfunding project into a text processing model to obtain semantic features, where the implicit feature branch includes a multi-modal deep learning model, and the multi-modal deep learning model includes a text processing model, an image processing model, and expert networks;

[0015] It is also used to input the picture information of the crowdfunding project into an image processing model to obtain visual features, where 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;

[0016] It is also used to obtain the activation probabilities of multiple expert networks based on the semantic features and visual features, and the expert networks include visual experts, language experts, cross-modal experts, functional experts, aesthetic experts, and creative experts;

[0017] It is also used to sparsely activate the corresponding expert networks according to the activation probabilities to obtain specialized implicit features corresponding to the expert networks;

[0018] A feature fusion layer for inputting the structured features and specialized implicit features into a feature fusion model to obtain the prediction success rate.

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

[0020] By constructing a hierarchical hybrid model, structured data passes through the structured branch to extract structured features; text information and image information respectively pass through the text processing model and the image processing model in the implicit feature branch to extract semantic features and visual features. Since the text processing model and the image processing model share a Transformer encoder, it can ensure that text and images are aligned in a unified semantic space, making the representation forms of the extracted semantic features and visual features consistent, laying a foundation for subsequent feature fusion and improving the accuracy of feature fusion. According to the semantic features and visual features, the activation probabilities of multiple expert networks can be obtained, and the expert network corresponding to the larger activation probability is selected for activation. If the crowdfunding project focuses on aesthetic-related products, then the aesthetics expert, creativity expert, or functionality expert, etc., will be activated to obtain specialized implicit features from the extracted semantic features and visual features, such as specifically strengthening implicit features related to aesthetics, creativity, or functionality. These specialized implicit features and structured features are input into the feature fusion model to obtain the predicted success rate. The success rate predicted based on the specialized implicit features is more consistent with the actual success rate, ensuring the accuracy of the prediction. Therefore, when dealing with unstructured data closely related to product appearance and aesthetic features, this method can highlight and strengthen such features through expert networks, improving the prediction accuracy of aesthetic crowdfunding projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0022] Figure 1 It is a flowchart of a method for predicting the success rate of product crowdfunding driven by multi-modal data provided by an embodiment of the present invention;

[0023] Figure 2 It is a structural diagram of a system for predicting the success rate of product crowdfunding driven by multi-modal data provided by an embodiment of the present invention;

[0024] Figure 3 It is a structural diagram of implicit feature analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described clearly and completely. Obviously, the described embodiments are part of the embodiments of this application, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts belong to the scope of protection of this application.

[0026] To better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific implementation manners.

[0027] As Figure 1 shown, a method for predicting the success rate of product crowdfunding driven by multimodal data provided by an embodiment of this application includes:

[0028] S110: Input the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model, and screen to obtain structured features, where the hierarchical hybrid model includes a structured branch, a latent 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, where the latent 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.

[0030] S130: Input the image information of the crowdfunding project into the image processing model to obtain visual features, where 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.

[0031] S140: According to the semantic features and visual features, obtain the activation probabilities of multiple expert networks, where the expert networks include a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert, and a creative expert.

[0032] S150: According to the activation probabilities, sparsely activate the corresponding expert networks to obtain specialized latent features corresponding to the expert networks.

[0033] S160: Input the structured features and the specialized latent features into the feature fusion model to obtain the predicted success rate.

[0034] In this alternative embodiment, a hierarchical hybrid model is constructed. The structured data passes through the structured branch to extract structured features. The text information and image information respectively pass through the text processing model and the image processing model in the implicit feature branch to extract semantic features and visual features. Since the text processing model and the image processing model share a Transformer encoder, it can ensure that the text and the image are aligned in a unified semantic space, making the representation forms of the extracted semantic features and visual features consistent, laying a foundation for subsequent feature fusion and improving the accuracy of feature fusion. According to the semantic features and visual features, the activation probabilities of multiple expert networks can be obtained, and the expert network corresponding to the larger activation probability is selected for activation. If the crowdfunding project focuses on aesthetic-related products, the aesthetics expert, creativity expert, or functionality expert, etc. will be activated. Since the expert network is a specialized expert network obtained through constrained training, through the processing of these expert networks, specialized implicit features can be obtained from the extracted semantic features and visual features, such as specifically strengthening the implicit features related to aesthetics, creativity, or functionality. These specialized implicit features can better highlight the characteristics of the crowdfunding product. The specialized implicit features and the structured features are input into the feature fusion model to obtain the predicted success rate. The success rate predicted based on the specialized implicit features is more consistent with the actual success rate, ensuring the accuracy of the prediction. Therefore, when processing unstructured data closely related to product appearance and aesthetic features, this method can highlight and strengthen such features through the expert network, improving the prediction accuracy of aesthetic crowdfunding projects.

[0035] As Figure 2 shown, 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 image information. In the structured branch, statistical or machine learning methods can be used to identify or extract data-type features or category-type features in the structured data. In the implicit feature branch, a multimodal deep learning model is used to process text information and image information, as Figure 3As shown in the figure, the multimodal 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 image information is processed by the image processing model. By sharing the same Transformer encoder, it is possible to obtain semantic features and visual features with unified expressions in the same semantic space. Then, the semantic features and visual features are input into the expert network, and several expert networks are selectively and sparsely activated, so as to specialize the semantic features and visual features, and further extract specialized implicit features. For example, only superficial features can be extracted from the text expression and the picture. After being subjected to constrained training, the expert network can combine the semantic features and visual features, and dig deep into the implicit features tendentiously to obtain the corresponding specialized implicit features. It can also know the characteristics and scores of multiple dimensions such as the functionality, aesthetics, and creativity of the crowdfunding product of the crowdfunding project according to the specialized implicit features. The structured features and specialized implicit features are input into the feature fusion model to obtain the predicted success rate, and then the predicted success rate is analyzed to obtain the output result that can guide the decision-making.

[0036] The following describes each step in detail.

[0037] S100: Data collection and preprocessing.

[0038] First, data collection and preliminary cleaning are carried out for the core fields that can be directly quantified on the crowdfunding platform (such as the target amount, fundraising period, project category, number of supporters, social media association flag, etc.). After obtaining the relevant information through the method of Internet crawler or official API (Application Programming Interface), it 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 boxplot detection method is used to identify outliers, and the truncation method or the quantile correction method is used to process the outliers; for the missing fields, KNN (K-Nearest Neighbor) interpolation is used to infer more credible values based on the distance measure from the most similar samples. Then, in order to alleviate the problem of unstable model training caused by overly skewed numerical distributions or magnitude differences, logarithmic transformation or interval scaling (Min-Max Scaling) can also be performed on numerical columns such as amount and number of people to pull different features into a comparable range to improve the efficiency and accuracy in subsequent feature fusion.

[0039] In unstructured data, the text information of crowdfunding projects is mainly considered, including common text data sources such as project entity descriptions, update announcements, and supporter comments. After obtaining the original text, first use regular expressions to remove HTML tags, redundant symbols, and high-noise characters; for records with mixed languages or extremely short content, filter them as appropriate to ensure the text quality for subsequent modeling. In addition to text, image data such as product prototype diagrams, promotional posters, or concept designs shown on the crowdfunding page are further collected. To ensure the consistency of the input scale, all images are uniformly scaled to 224×224 pixels, and histogram equalization is performed simultaneously to enhance contrast and distinguishability. In the case of limited sample size, random cropping, horizontal flipping and other enhancement strategies can be added as needed to increase the robustness of the model to different shooting environments and angles.

[0040] S110: Input the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model to screen out structured features, where the hierarchical hybrid model includes a structured branch, a latent 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, and machine learning methods can be used to identify specific structured features in structured data using convolutional neural networks or recurrent neural networks, etc., such as categorical features like project number, project name, project type, etc.

[0042] After preliminary cleaning and standardization processing of the structured data of the crowdfunding platform, through single-factor analysis, multi-factor interaction analysis, and causal verification and other links, the core structured features that can most affect the project success rate are accurately mined.

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

[0044] First, use statistical methods for single-factor screening to evaluate the relevance between each candidate structured feature and the success or failure of the crowdfunding project. In terms of numerical features (such as target amount, fundraising period, or number of supporters, etc.), the Pearson or Spearman correlation coefficient is used to measure the linear or monotonic correlation degree between it and the success or failure of the crowdfunding project (0 / 1 label, corresponding to failure / success). If there is a linear correlation between the numerical feature and the success or failure of the project, judge whether the numerical feature is a potential key feature according to the absolute value of the slope of the linear correlation. 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 there is a non-linear correlation between the numerical feature and the project success or failure, then perform segmented processing on the numerical feature, and use the chi-square test to analyze whether there is a significant association between each segment of the numerical feature and the crowdfunding success rate.

[0046] S113: Use the chi-square test to analyze whether there is a significant association between the categorical feature and the crowdfunding success rate.

[0047] S114: If the analysis result shows that there is a significant association between a structured feature and the crowdfunding success rate, then regard this structured feature as a potential key feature.

[0048] If it is found that some numerical features show non-linear relationships or strong skewed distributions, then after equal-frequency binning or segmented processing, use the chi-square test to determine whether there is a significant association with the success rate. At the categorical feature level (such as project category or social media association flag, etc.), also use the chi-square independence test to measure the dependence relationship between the feature value distribution and the success or failure of crowdfunding projects. When a feature shows a high degree of correlation at the preset significance level (such as p < 0.05), it is regarded as a potential key factor and retained in the subsequent interaction analysis.

[0049] S115: Use random forest and light gradient boosting machine to perform multi-factor interaction analysis on the structured features. When a structured feature shows a double-high phenomenon in the two multi-factor interaction analyses, then regard this structured feature as a potential key feature.

[0050] Specifically, after completing the single-factor screening, continue to use random forest and light gradient boosting machine to perform multi-factor interaction analysis on the structured features. Random forest measures the contribution degree (importance) of each feature to the improvement of sample purity during node splitting based on the Gini coefficient (Gini 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 in the Gini index of feature f at node t.

[0053] And the 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 respectively represent the sample sets of the left and right child nodes after splitting, I represents the sample set of the current node, g iis the gradient of sample i, h i is the second derivative of sample i. 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 the situation where high importance is presented under the analysis of two multi-factor interactions (the calculation result is greater than or equal to the corresponding preset threshold). When a structured feature shows the double-high phenomenon in the multi-factor interaction analysis, it can be determined that it has a robust discrimination ability for the crowdfunding result and can be regarded as a potential key feature. At the same time, for features with relatively high importance in only one type of multi-factor interaction analysis, they also need to be screened and comprehensively evaluated in combination with subsequent causal verification links.

[0057] S116: Use the PC algorithm to construct a directed acyclic graph. When a structured feature is mediated by only one potential key feature in the directed acyclic graph, then regard this structured feature as an auxiliary feature.

[0058] Specifically, to further eliminate the confounding effects and multicollinearity among features, the PC (Peter-Clark) algorithm is used to analyze the correlation between variables to determine the possible causal relationships between variables, and a partial directed acyclic graph (PDAG) is constructed to identify possible causal associations. When it is determined that a certain feature is only mediated by another more critical feature in the causal structure, then this feature is classified as an auxiliary factor to reduce feature redundancy and strengthen the key path. Among them, mediation means that a variable (called a mediating variable or mediator) plays a bridging role between two other variables. One variable (independent variable) affects another variable (dependent variable) through the mediating variable, that is, a potential key feature affects the success or failure of a crowdfunding project through a structured feature. At this time, the mediating variable is regarded as an auxiliary feature.

[0059] So far, through this progressive screening and verification, the structured features that play a major role in the success or failure of crowdfunding projects can be clearly defined, and these structured features with high value and high interpretability will be mainly used in the subsequent fusion modeling with unstructured features (text and images), so as to ensure the robustness and interpretability of the hierarchical hybrid model.

[0060] S120: Input the text information of the crowdfunding project into the text processing model to obtain semantic features. Among them, the implicit feature branch includes a multi-modal deep learning model, and the multi-modal 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 vectors or word vector representations. When more fine-grained expressions are needed, a method based on sub-word encoding after word segmentation can also be selected to obtain richer semantic features. Text vectorization can not only significantly reduce the high-dimensional noise of data, but also provide 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] S130: Input the picture information of the crowdfunding project into an image processing model to obtain visual features. Among them, 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.

[0063] Specifically, use the ViT model or convolutional neural network to obtain the image vector representation, save it to the multi-modal database, and associate it with the project ID, so that the text and image features corresponding to each project can be conveniently loaded in batches during subsequent model training.

[0064] After completing the structured data cleaning, text vectorization, and image encoding, all results can be uniformly stored in the multi-modal database, and the associated management of various types of data for the same project can be achieved through the project ID. The database tables can be divided into "structured information table", "text embedding table", and "image embedding table", and different tables are connected to each other by the project ID. This design can not only obtain multi-modal inputs with one key through batch queries in the subsequent training stage, but also ensure the convenient access of incremental data. When a new project appears on the crowdfunding platform or an existing project is updated, just add the corresponding record and retain the reference to the project ID, and it can be seamlessly incorporated into the overall prediction process. Through such a data architecture, the unstructured noise of text and images is significantly reduced, and the structured features are also reasonably normalized in terms of magnitude and dimension, thus 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, unstructured data such as text and images in the project description carry rich implicit information, such as the functionality, aesthetics, and creativity of the project. Although these implicit features are difficult to directly quantify, they have an important impact on the attractiveness, user participation, and ultimate success rate of the project. In order to accurately quantify these implicit features, we adopted a multi-modal deep learning model to achieve in-depth analysis and feature extraction of text and image data through semantic embedding, cross-modal alignment, and expert network division of labor.

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

[0067] To ensure that the text and the 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 the image and performs cross-modal parameter sharing and semantic alignment through a contrastive learning loss function. The goal of contrastive learning is to maximize the similarity of positive sample pairs (i.e., the text and the image of the same item) while minimizing the similarity of negative sample pairs (i.e., the text and the image of different items). The specific formula is as follows:

[0068]

[0069] where τ = 0.07 is the temperature parameter, which is used to adjust the smoothness of the similarity distribution; sim is the cosine similarity, which is used to measure the similarity between the text and the 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 samples), 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 the alignment of the text and the image in a unified semantic space, thus providing a consistent representation basis for subsequent feature quantization.

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

[0071] S150: Sparsely activate the corresponding expert networks according to the activation probabilities to obtain specialized latent features corresponding to the expert networks.

[0072] Specifically, based on semantic alignment, the multi-modal deep learning model introduces multiple expert networks, and uses a dynamic routing mechanism to refine different types of implicit features. Through the division of labor and cooperation of multiple expert networks, each expert network focuses on processing specific types of features. The expert networks in the multi-modal deep learning model include visual experts, language experts, cross-modal experts, as well as specially designed functional experts, aesthetic experts, and creative experts. The activation probability of each expert network is calculated by a lightweight routing network. For example, when inputting picture information, the language expert will not be activated, while the visual expert and cross-modal experts will be activated. In addition, the functionally specialized functional experts, aesthetic experts, and creative experts will all be activated. However, the activation degrees of these three networks will be different; when inputting picture information, the aesthetic expert has a higher activation degree; while when inputting text information, the functional network has a higher activation degree.

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

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

[0075] Among them, the softmax function is used in the output layer of multi-classification problems. It can map the input to a probability distribution, making the sum of all outputs equal to 1. W p ∈R 1024×256 is the projection matrix, which is used to map the input features to a low-dimensional space. W g ∈R 256×3 is the routing weight, which is used to calculate the activation probability of each expert network. GELU (Gaussian Error LinearUnit) is an activation function that can better capture non-linear features. The result obtained by this formula is the probability distribution obtained after the input x undergoes a series of transformations.

[0076] To further improve the feature extraction ability of the expert networks, the multi-modal deep learning model specializes the functions of the expert networks through constrained training. Specifically, the functional experts focus on extracting features related to project practicality, technical implementation, etc.; the aesthetic experts focus on extracting features related to design aesthetics, visual attraction, etc.; and the creative experts focus 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 multiple dimension scores according to the specialized implicit features, where the dimension scores include functional scores, aesthetic scores, and creativity scores.

[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 scoring dimensions of functionality, aesthetics, and creativity. The Sigmoid function is used to normalize the functional scores and map the outputs to the interval [0, 1]. The characteristics of the Sigmoid function make it suitable for handling binary classification or probability estimation problems and can effectively represent the functional intensity of the item. The Tanh function is used to normalize the aesthetic scores and map the outputs to the interval [-1, 1]. The symmetry of the Tanh function makes it suitable for capturing the bipolar distribution of aesthetic features and can effectively represent the aesthetic degree of the design. The Softplus function is used to normalize the creativity scores. 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<00000!3>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 of the corresponding functions, z func 、z aes 、z creat are the input vectors, and b f 、b a 、b c It should be noted that there seems to be a typo in the original text where "W <00000!3> " should probably be "W f ". This has been corrected in the translation as much as possible while maintaining the original format.is the bias term. Screening crowdfunding projects according to the dimensionality scores to obtain target crowdfunding projects can help decision-makers quickly screen out projects that meet specific criteria. According to the scores, decision-makers can reasonably allocate resources. Additionally, by analyzing users' preferences, projects with high aesthetics or creativity scores can be recommended based on users' interests to improve the user experience. Moreover, through the multi-dimensional scoring of crowdfunding projects, the possible success rate of the crowdfunding project can be intuitively felt. For example, if a crowdfunding project has relatively high scores in functionality, aesthetics, and creativity, then the prospects of this project are relatively large, and the probability of being accepted by the public and favored by investors will be greatly increased, thus promoting the ultimate success of the crowdfunding for this project.

[0083] S160: Input 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, bilinear pooling, and a multi-layer perceptron.

[0084] Specifically, the feature fusion layer is responsible for deeply integrating the structured feature with the specialized implicit feature extracted by the multi-modal deep learning model, capturing the complex interaction relationship between the two, so as to provide a high-quality input representation for the prediction of the success rate of crowdfunding projects. The advantages of two fusion methods are dynamically combined through the gated cross-attention mechanism, where bilinear pooling captures the explicit interaction pattern and the multi-layer perceptron models the deep non-linear relationship. Thus, the structured feature is coupled with the specialized implicit features such as aesthetics, creativity, and functionality output by the multi-modal deep learning model, and then the success rate is predicted based on the fused features.

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

[0086] Specifically, the gated cross-attention mechanism coordinates two feature fusion methods, bilinear pooling and the multi-layer perceptron, through dynamic weight allocation. The gated cross-attention mechanism generates weights based on the input features to achieve the weighted combination of the two branches, thereby adaptively integrating the structured feature and the specialized implicit feature. This mechanism not only retains the non-linear expression ability of the multi-layer perceptron but also utilizes the explicit interaction modeling advantage of bilinear pooling, realizing the complementary advantages of feature fusion.

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

[0088] Bilinear pooling is an explicit second-order feature interaction method that can capture the multiplicative effects between different feature dimensions. In the evaluation of crowdfunding projects, certain feature combinations may have a significant impact on the prediction of project success rates. For example, combinations such as "target amount × creativity" or "social dissemination × aesthetics" may better reflect the potential success probability of a project than individual features. Bilinear pooling can more accurately capture these high-order interaction relationships by explicitly calculating the multiplicative effects of these feature combinations.

[0089]

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

[0091] S163: Input the structured feature and the specialized latent feature into a multi-layer perceptron to obtain a second fused feature.

[0092] The multi-layer perceptron is used to perform non-linear transformations on features and capture complex feature relationships. The multi-layer perceptron includes an input layer, multiple hidden layers, and an output layer. The hidden layers are used to extract high-level features of the input data, perform weighted summation on the input data, and then introduce non-linearity through an activation function. The multi-layer perceptron 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 parameters during training to improve the prediction accuracy. The following is the formula for the multi-layer perceptron to process the fusion of structured data and specialized latent features.

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

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

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

[0096] The gated cross-attention mechanism dynamically assigns weights to coordinate the outputs of bilinear pooling and the MLP. The gating unit generates weights based on the input features to achieve the weighted combination of the two branches.

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

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

[0099] S165: Obtain the prediction 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 output the result according to the requirements. For example, in a classification task (success or failure), the class probability is output.

[0101] S170: According to the attention matrix output by the gated cross-attention mechanism model, use the sparsity constraint to obtain the sparse attention matrix.

[0102] Specifically, in the feature extraction layer, the multi-modal deep learning model has the sparse activation characteristic. In the feature fusion layer, the sparsity constraint (such as L1 regularization or Top-K pruning strategy) is further introduced. By introducing the sparsity constraint, the attention weights are forced to concentrate on the high-weight features, improving the feature selection accuracy and the model interpretability. In the prediction interpretation layer, the sparse attention matrix is used to intuitively display the high-weight features for feature importance attribution, and explicit rules are extracted from it to automatically induce decision rules, enhancing the model's interpretability and operability.

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

[0104]

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

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

[0107]

[0108] For the attention matrix, retain the top K elements with the highest weights and set the rest to zero. The value of K is dynamically adjusted according to the 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 chi-square tests.

[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: Use the Apriori algorithm to extract high-frequency feature combinations (e.g., support > 0.6, confidence > 0.8) from the sparse attention matrix and verify the significance of the rules through chi-square tests (p < 0.01). The output decision rule form 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 thresholds or cutoffs of the features obtained through data analysis and algorithms, meaning that when the price and functional score of the product exceed 0.7, and the aesthetic and social dissemination scores exceed 0.5, the association rule considers these conditions to be met, then the prediction success rate is greater than 70%.

[0114] Readable expression and decision support: Through the above rule mining process, the complex model decision-making process can be transformed into a series of easy-to-understand "IF-THEN" rules. For example, rules like "IF (price < X AND functionality > Y) THEN (success probability > Z%)" are not only intuitive and easy to understand but also can provide clear decision-making basis for the initiators or evaluators of crowdfunding projects. These rules can help the project initiators make targeted adjustments during the project design stage, thereby increasing 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 objective of multi-stage feature interaction, a progressive training strategy is adopted to gradually unfreeze each model in stages and perform joint optimization. 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] Perform separate pre-training on the text processing model and the image processing model in the multi-modal deep learning model.

[0118] Perform constrained training on multiple expert networks in the multi-modal deep learning model.

[0119] Freeze the parameters of the multi-modal deep learning model and train the structured branch and the feature fusion model.

[0120] Specifically, in the initial stage, first fix most of the parameters of the multi-modal deep learning model, and mainly train the structured branch and the feature fusion layer, so that the model can first master the numerical and categorical data of the crowdfunding project and establish a basic success rate prediction ability. At this stage, the text and image features are usually obtained in a pre-trained manner and kept in an updated and frozen state to avoid simultaneous multi-modal alignment and structured feature learning, which may cause training instability or gradient conflicts. By only optimizing the structured branch and the feature fusion layer, the model can converge quickly in the early stage and learn the main linear or non-linear interaction relationships, laying a good foundation for subsequent unfreezing of the hidden feature branch for parallel training.

[0121] Unfreeze the top-level parameters in the multi-modal 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 in the multi-modal deep learning model and the parameters before the Transformer encoder.

[0122] Specifically, when the structured branch has become relatively stable and the overall prediction accuracy has reached a certain level, gradually unfreeze the top-layer parameters of the multi-modal deep learning model, that is, the parameters located at the top layer of the network architecture. These parameters are responsible for processing the features after being processed by the previous layers (such as the feature extraction layer) and finally generating the output of the model. After unfreezing the top-layer parameters, let them start joint optimization with the feature fusion layer. At this time, the model can more deeply perform cross-modal alignment on text and image data, strengthen the semantic consistency between text and image through contrastive learning, so as to establish a closer association between implicit features such as functionality, aesthetics, and creativity and structured features. At this stage, the weights of modules such as the gated cross-attention mechanism and bilinear pooling will also be continuously updated, making the multi-modal interaction 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, and integrate the information of unstructured features into the decision-making process while maintaining the advantages of structured features.

[0123] Unfreeze all the frozen parameters and perform fine-tuning training on the hierarchical hybrid model using perturbed data, where the perturbed data is data obtained by using the adversarial data augmentation method.

[0124] Specifically, in the final stage, unlock all the frozen network parameters and conduct end-to-end unified training. To enhance the robustness of the model under sudden changes in data distribution or noisy input, the adversarial data augmentation method can also be used to add small perturbations to text or image inputs, or perform random jitter on structured features. This can enable the model to learn to maintain relatively stable prediction performance and interpretability in extreme or uncertain environments. Finally, through progressive training, gradually achieve a robust learning process from structured dominance to full multi-modal fusion, which can not only alleviate the risk of gradient explosion or disappearance during the early training of multi-modal deep learning models, but also polish a high-precision and interpretable crowdfunding success rate prediction model in the later stage, providing comprehensive and reliable decision-making support for crowdfunding platforms and project initiators.

[0125] To address the high information entropy loss and insufficient prediction accuracy caused by the "modal fragmentation effect", this application achieves cross-modal semantic alignment at the text and image levels 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, fundraising cycle) with abstract functional, aesthetic, and creative features. Compared with models such as MMF-CNN that simply stitch text and images together, this application effectively reduces the information entropy loss rate in experiments and reduces the prediction error of aesthetics-related features by nearly 10% - 15%, showing higher accuracy and generalization ability in creative or aesthetic projects.

[0126] To address the disconnect between interpretability and business needs, this application introduces sparsity constraints and association rule mining in the prediction and interpretation layers of the hierarchical hybrid model, breaking through the limitations of methods like LIME and SHAP, which can only provide numerical feature importance. Unlike ProtoPNet, which requires manual annotation of a large number of samples, this application automatically highlights core weights during training through Top-K pruning or L1 regularization. In crowdfunding practice, this can help project initiators or platform managers quickly identify deficiencies in appearance design, creative copywriting, or pricing mechanisms, improve product iteration efficiency, and reduce market investment risks.

[0127] In order to cope with the risk of model failure in a dynamic market environment, this application adopts a progressive training and incremental learning strategy to achieve real-time adaptation to changes in the distribution of crowdfunding project types by unfreezing the multimodal deep learning model and feature fusion layer at different times. When the proportion of project categories changes significantly, the accuracy of traditional static modeling methods (such as LightGBM+genetic algorithm) can be reduced by up to 23%. After the injection of incremental data, this application can quickly tune the sparse attention matrix and maintain a high prediction accuracy, greatly reducing the secondary tuning cost caused by model mismatch. In addition, the feature stability index (FSI) of the model has also been significantly improved, which can more flexibly capture the market characteristics of emerging categories (such as AI hardware) and adapt to the rapidly iterating needs of the crowdfunding ecosystem.

[0128] In summary, this system demonstrates innovative breakthroughs in the areas of "multimodal feature fusion," "enhanced interpretability," and "dynamic market adaptation." It not only significantly improves the accuracy of crowdfunding project success predictions but also clearly presents the contributions of each modal feature as executable rules, providing valuable technical support for innovation verification and market decision-making for crowdfunding platforms and project initiators. Compared to existing technologies, it demonstrates superior practicality and commercial value, fully meeting the crowdfunding ecosystem's requirements for short-term, highly accurate, and actionable outputs.

[0129] In summary, compared with the existing technology, the present invention has the following beneficial effects:

[0130] 1. Build a multimodal expert network based on a multimodal deep learning model, achieve semantic alignment of text and images through dynamic routing and sparse activation mechanisms, and effectively improve the ability to quantify implicit features such as product functionality, aesthetics, and creativity.

[0131] 2. By integrating structured features with unstructured abstract features in a hierarchical hybrid model, and through gated cross-attention mechanism, bilinear pooling and multi-layer perceptron, we can explore multi-dimensional interactive relationships and significantly enhance the prediction accuracy and stability of the success rate of crowdfunding projects.

[0132] 3. By introducing sparsity constraints and rule extraction methods, high-confidence "IF-THEN" decision rules can be extracted from the attention matrix, improving the interpretability and operability of crowdfunding prediction results.

[0133] 4. By adopting a progressive training strategy and a feature evolution mechanism, the model can maintain dynamic adaptability when the market structure changes or the project type fluctuates, ensuring the continuous and stable output of prediction performance.

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

[0135] To specifically demonstrate the actual application effect of the multi-modal crowdfunding success rate prediction method proposed by the present invention in real crowdfunding projects, in this embodiment, five typical projects of the "Design - Toy" category on a certain platform are taken as the research object. According to the technical route of the present invention, the whole process analysis of data collection and preprocessing, structured feature analysis, unstructured implicit feature quantification, success rate prediction, and interpretable output is completed in sequence.

[0136] First, the structured fields of each project page are crawled from a certain platform, including information such as the project target amount, fundraising period, number of supporters, and whether it is associated with social media, and are stored in a multi-modal database with the project ID as the primary key. After identifying outliers in continuous variables such as the target amount and the number of supporters using box plots, the truncation or quantile method is used to correct outliers, and the KNN method is applied to impute missing data. To avoid the influence of numerical feature scale differences on model stability, logarithmic transformation and interval scaling techniques are also used for normalization processing.

[0137] Secondly, in the unstructured data preprocessing stage, the BERT and ViT are respectively used to extract feature vectors for the project description text and product design images provided on the crowdfunding page. The text data removes HTML tags, special symbols, and noise characters through regular expressions and is encoded by the BERT model to form a unified semantic vector representation; the image data is unified in size, enhanced in detail through histogram equalization, and then generates image vector feature representations via ViT. Subsequently, cross-modal semantic alignment is completed through the shared Transformer encoder of the multi-modal deep learning model. The three types of implicit features of the project are accurately quantified through functional experts, aesthetic experts, and creative expert networks.

[0138] Taking the "Modular Magnetic Building Blocks" project as an example, its functional, aesthetic, and creative scores are at a relatively 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 of the present invention and mapped into high-quality implicit feature scores.

[0139] Subsequently, it enters the structured factor identification stage. First, through single-factor analysis, correlation tests (Pearson and Chi-square tests) are conducted on structured variables such as the target amount, fundraising cycle, and number of supporters to confirm their significant relationship with the crowdfunding success rate. Multifactor analysis then uses the random forest and LightGBM models to further identify highly correlated variables based on Gini Importance and Gain. For example, the number of supporters and the social media association flag have a relatively high contribution to the success rate, showing robust importance. In the causal verification link, the PC algorithm is used to further eliminate redundant or mediated variables, and the final set of structured features is determined.

[0140] Next, in the hierarchical hybrid model construction and prediction stage, the structured features are deeply fused with the implicit features quantified by the multimodal deep learning model through the gated cross-attention mechanism and bilinear pooling. The model automatically adjusts the attention allocation weights through the dynamic routing mechanism, enabling high-order interactions between structured features (such as the number of supporters and the target amount) and unstructured features (such as the creativity score), and outputs the predicted success probability of each project.

[0141] Taking the "Maglev Educational Toy" project as an example, its text description quality is 0.89, image clarity is 0.92, functionality score is 0.94, aesthetics score is 0.81, and creativity score is 0.88. After the model synthesizes various factors, the predicted success probability is as high as 0.95, and the actual crowdfunding amount reaches $56,800, far exceeding its initial target of $25,000, verifying the high accuracy of the model.

[0142] At the same time, to further enhance the interpretability and practicality of the model, a sparse attention matrix is introduced during the prediction process, and algorithms (such as Apriori) are used to extract actionable causal rule examples from the model output:

[0143] IF (functionality score > 0.85) AND (aesthetics score > 0.75) AND (target amount ≤ 20,000) THEN (predicted success rate > 0.90)

[0144] This rule has been verified to be effective in both 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 18,000, meeting the above rule conditions. The actual fundraising amount reaches 40,200, more than twice the target, highlighting the practical guidance and accuracy of the model's prediction rules. For the "Multidimensional Geometric Insertion" project, due to the relatively low scores of the implicit features (creativity 0.59, aesthetics 0.51), the predicted success rate is only 0.52, and the actual fundraising fails to reach the target amount, demonstrating the reliability of the model in predicting failure risks.

[0145] Table 1 Implicit Feature Scores and Predicted Success Rates of Toy Crowdfunding Projects

[0146] Project Name Functional Evaluation Aesthetic Evaluation Creativity Evaluation Predicted Success Probability Actual Success or Not 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 Intelligent Teaching Aid Train Set 0.75 0.60 0.70 0.66 Yes Customizable Programming Top 0.91 0.79 0.86 0.90 Yes Multi-Dimensional Geometric Interlocking 0.68 0.51 0.59 0.52 No

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

[0148] As Figure 2 shown, a multi-modal data-driven product crowdfunding success rate prediction system provided by an embodiment of this application includes:

[0149] A feature extraction layer, configured to input the structured data of a crowdfunding project into the structured branch of a hierarchical hybrid model to screen out structured features, where the hierarchical hybrid model includes a structured branch, an implicit feature branch, and a feature fusion model; and is further configured to input the text information of the crowdfunding project into a text processing model to obtain semantic features, where the implicit feature branch includes a multi-modal deep learning model, and the multi-modal deep learning model includes a text processing model, an image processing model, and an expert network; and is further configured to input the image information of the crowdfunding project into the image processing model to obtain visual features, where 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; and is further configured to obtain the activation probabilities of multiple expert networks according to the semantic features and visual features, where the expert networks include a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert, and a creative expert; and is further configured to sparsely activate the corresponding expert networks according to the activation probabilities to obtain specialized implicit features corresponding to the expert networks.

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

[0151] A prediction explanation layer, configured to obtain a sparse attention matrix according to the attention matrix output by the gated cross-attention mechanism model by using sparsity constraints; and is further configured to extract high-frequency feature combinations from the sparse attention matrix by using the Apriori algorithm and verify the significance of the high-frequency feature combinations through chi-square tests; and is further configured to, when the extracted high-frequency feature combinations are significant, output the high-frequency feature combinations and the predicted success rate as decision rules.

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

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

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

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

[0156] It should be noted that in this article, relational 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 actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements 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, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the 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 method for predicting the success rate of product crowdfunding driven by multi-modal data, characterized in that, Including: Input the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model, and filter out structured features. Among them, the hierarchical hybrid model includes a structured branch, a latent feature branch, and a feature fusion model; Input the text information of the crowdfunding project into the text processing model to obtain semantic features. Among them, the latent 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; Input the picture information of the crowdfunding project into the image processing model to obtain visual features. Among them, 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; According to the semantic features and visual features, obtain the activation probabilities of multiple expert networks. Among them, the expert network includes a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert, and a creative expert; Sparsely activate the corresponding expert network according to the activation probability to obtain specialized latent features corresponding to the expert network; Input the structured features and specialized latent features into the feature fusion model to obtain the predicted success rate.

2. The multi-modal data-driven product crowdfunding success rate prediction method according to claim 1, wherein, The structured features include numerical features and categorical features; The inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model and filtering out structured features includes: Analyze the correlation between the numerical features and the success or failure of the project; If there is a non-linear correlation between the numerical features and the success or failure of the project, segment the numerical features, and use the chi-square test to analyze whether there is a significant association between each segment of numerical features and the crowdfunding success rate; Use the chi-square test to analyze whether there is a significant association between the categorical features and the crowdfunding success rate; If the analysis result shows that a structured feature has a significant association with the crowdfunding success rate, regard this structured feature as a potential key feature.

3. The multi-modal data-driven product crowdfunding success rate prediction method according to claim 2, wherein The inputting the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model and filtering out structured features further includes: Use random forest and light gradient boosting machine to perform multi-factor interaction analysis on the structured features. When a structured feature shows a double-high phenomenon in the two multi-factor interaction analyses, regard this structured feature as a potential key feature; Use the PC algorithm to construct a directed acyclic graph. When a structured feature is only mediated by one potential key feature in the directed acyclic graph, regard this structured feature as an auxiliary feature.

4. The method for predicting the success rate of product crowdfunding driven by multimodal data according to claim 1, wherein, The expert network is functionally specialized through constrained training.

5. The multi-modal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that After the corresponding expert network is sparsely activated according to the activation probability to obtain specialized latent features corresponding to the expert network, it further includes: According to the specialized latent features, obtain multiple dimension scores. Among them, the dimension scores include functional scores, aesthetic scores, and creative scores.

6. The multi-modal data-driven product crowdfunding success rate prediction method according to claim 5, wherein, The obtaining multiple dimension scores according to the specialized latent features includes: Use the Sigmoid function to normalize the functional score; Use the Tanh function to normalize the aesthetic score; The Softplus function is used to normalize the creativity score.

7. The method for predicting the success rate of product crowdfunding driven by multi-modal data according to claim 1, wherein, The feature fusion model includes a gated cross-attention mechanism, bilinear pooling, and a multi-layer perceptron; The steps of inputting the structured feature and the specialized implicit feature into the feature fusion model to obtain the prediction success rate include: Inputting the structured feature and the specialized implicit feature into the gated cross-attention mechanism to obtain a weighted weight and an attention matrix; Inputting the structured feature and the specialized implicit feature into bilinear pooling to obtain a first fusion feature; Inputting the structured feature and the specialized implicit feature into the multi-layer perceptron to obtain a second fusion feature; Weighting the weighted weight, the first fusion feature, and the second fusion feature to obtain a final fusion feature; Obtaining the prediction success rate according to the final fusion feature.

8. The method for predicting the success rate of product crowdfunding driven by multimodal data according to claim 7, wherein After inputting the structured feature and the specialized implicit feature into the feature fusion model to obtain the prediction success rate, the following steps are further included: According to the attention matrix output by the gated cross-attention mechanism model, using sparsity constraints to obtain a sparse attention matrix; Using the Apriori algorithm to extract high-frequency feature combinations from the sparse attention matrix, and verifying the significance of the high-frequency feature combinations through chi-square tests; When the extracted high-frequency feature combinations are significant, the high-frequency feature combinations and the prediction success rate are output as decision rules.

9. The multi-modal data-driven product crowdfunding success rate prediction method according to claim 1, characterized in that, It also includes progressive model training; The progressive model training includes: Separately pre-training the text processing model and the image processing model in the multi-modal deep learning model; Conducting constrained training on multiple expert networks in the multi-modal deep learning model; Freezing the parameters of the multi-modal deep learning model and training the structured branch and the feature fusion model; Thawing the top-layer parameters in the multi-modal deep learning model and jointly optimizing and training them with the feature fusion model, where the top-layer parameters are the parameters in the Transformer encoder in the multi-modal deep learning model and the parameters before the Transformer encoder; Thawing all the frozen parameters and fine-tuning the hierarchical hybrid model with perturbed data, where the perturbed data is data obtained by using the adversarial data augmentation method.

10. A multi-modal data-driven product crowdfunding success rate prediction system, characterized in that, It includes: A feature extraction layer, which is used to input the structured data of the crowdfunding project into the structured branch of the hierarchical hybrid model and screen out the structured features. 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 the text processing model to obtain semantic features. The implicit feature branch includes a multi-modal deep learning model, and the multi-modal 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. 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; It is also used to obtain the activation probabilities of multiple expert networks according to the semantic features and visual features, where the expert networks include a visual expert, a language expert, a cross-modal expert, a functional expert, an aesthetic expert, and a creative expert; It is also used to sparsely activate the corresponding expert networks according to the activation probabilities to obtain specialized implicit features corresponding to the expert networks; A feature fusion layer is used to input the structured features and the specialized implicit features into a feature fusion model to obtain a prediction success rate.

Images