A method for predicting the popularity of user-generated content based on retrieval enhancement prediction

Abstract

This invention relates to the field of information dissemination prediction technology and provides a method for predicting the popularity of user-generated content (UGC) based on retrieval-enhanced prediction (RAP). By constructing a UGC knowledge vector base and designing a retrieval-enhanced prediction (RAP) method, this invention employs a dual-path retrieval, high-speed recall, and a decoupled noise reduction architecture for the retrieved content, effectively solving the performance bottleneck problem of existing UGC popularity prediction methods. Specifically, the dual-path retrieval and re-ranking recall architecture addresses the long text representation bias problem; the decoupled noise reduction method in the downstream deep learning network module, which dynamically adjusts the feature fusion weights based on confidence, solves the retrieval noise interference problem; and the overall architecture design overcomes the strong coupling defect of the model, enabling the model to adapt to dynamic updates of the retrieval database, thus improving its versatility and generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of information dissemination prediction technology, and in particular relates to a method for predicting the popularity of user-generated content based on retrieval enhancement prediction. Background Technology

[0002] In the rapidly developing digital age, the continuous advancements in internet and mobile communication technologies have propelled the creation and dissemination of video at an unprecedented speed. User-generated content (UGC), as a crucial means for people to access information and entertainment on the internet, has seen its popularity and trend prediction become key research areas. This research not only helps content producers enrich and innovate video content but also assists internet platforms in optimizing resource allocation and improving user experience, thereby driving the overall development of industry technology. UGC is characterized by its vast quantity and diverse formats, encompassing multimodal data such as text, images, video, and audio. This explosive growth and complexity of data makes accurately predicting its popularity a critical technical challenge for business scenarios such as content recommendation, advertising, and public opinion monitoring.

[0003] Currently, existing methods for predicting UGC popularity mainly fall into the following categories: Traditional machine learning methods: As an important foundational tool, researchers personalize and improve classic models to adapt to prediction tasks. Commonly used methods include linear regression, support vector machines (SVM), decision trees, and random forests (RF). Multimodal methods: Integrating multi-source information such as visual, audio, text descriptions, user interaction behavior, and social network relationships from videos, combining heterogeneous data sources through early fusion and late fusion to improve the accuracy and robustness of predictions. Variational autoencoders (VAEs): As generative models, they learn the probability distribution of data to capture the latent representations of videos and user behavior, helping to understand the complex dynamics of video dissemination. Knowledge graphs (KGs): Integrating information such as video content and user behavior using structured semantic networks, they mine deep relationships in the data through relational reasoning and path deduction. Graph neural networks (GNNs): Leveraging their ability to process graph-structured data, they can effectively express and capture entity relationships; entities such as videos and users are easily embedded in graph structures. Heterogeneous graphs: Using different types of nodes and edges to make entity connections clearer and relationships more complex, improving the effectiveness of video popularity prediction. Information cascading prediction: This method analyzes the information cascading process to obtain the scale and popularity of dissemination. The cascading process represents the propagation network formed by video through user behavior in the network, characterizing the dynamics of information dissemination. Existing UGC popularity prediction methods have many limitations: At the data level, limitations such as the quality of UGC data and the explosive growth of long semantic information make it difficult to fully model the complex semantic relationships and external knowledge dependencies of UGC, leading to bottlenecks in prediction performance. At the deep learning network training level, it is difficult to fully capture the implicit cross-modal semantic relationships in UGC content, including the deep interaction relationships between multimodal features such as visual, auditory, and textual features. Even in the few methods that use retrieval libraries to enhance semantic understanding and UGC prediction tasks, the following technical limitations exist:

[0004] Search noise interference: If the vector retrieval library is of low quality and the Top-K information is not related to the target UGC content, it will introduce strong noise into the prediction model;

[0005] Inaccurate embedding of long texts: Traditional single-path retrieval (dense vector similarity retrieval) often fails to accurately express semantic information when embedding long texts, reducing prediction reliability;

[0006] Coupling defects: The strong coupling between the retrieval module and the prediction model makes it difficult for the model to adapt to the dynamic updates of the vector retrieval library, and it is prone to over-reliance on the retrieval library. When the retrieval fails, it cannot maintain the basic prediction ability and loses its versatility and generalization ability.

[0007] These limitations collectively lead to performance bottlenecks in prediction models, making traditional prediction methods based on a single deep learning network architecture inadequate. To address this, this invention proposes a user-generated content popularity prediction method based on Retrieval Augmented Prediction (RAP). Summary of the Invention

[0008] The purpose of this invention is to provide a method for predicting the popularity of user-generated content based on retrieval enhancement prediction, aiming to solve the problems mentioned in the background art.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A method for predicting the popularity of user-generated content based on retrieval-enhanced prediction includes the following steps:

[0011] Step 1: Data Collection;

[0012] Collect multimodal data from UGC websites, including video streams, titles, comment content, cover images, and popularity metrics such as likes, views, and comments.

[0013] Step 2: Data preprocessing;

[0014] The collected raw data is cleaned, duplicates are removed, modal missing items are processed, and text and images are standardized. The samples are then divided into a UGC knowledge vector base set, a training set, and a test set.

[0015] Step 3: Feature construction;

[0016] Multimodal embedding representations are generated by extracting features from three modalities: video, text, and image, and then performing modality augmentation and embedding transformation.

[0017] Step 4: Constructing a UGC knowledge vector library;

[0018] For text modalities, all text features are concatenated and long text embeddings are generated to form a UCG knowledge vector library;

[0019] Step 5: Dual-path retrieval and rearrangement recall;

[0020] For text embeddings of UGC samples, dense vector similarity retrieval and BM25 retrieval based on word frequency are performed sequentially in the UGC knowledge vector base. Then, the Top-K results are retrieved through a dual-path fusion using the RFF algorithm to obtain the final recall set.

[0021] Step 6: Retrieve enhanced prediction results to assist in predicting the popularity of UGC on downstream networks;

[0022] By combining deep learning models, model training and popularity prediction are performed based on the input multimodal features and recall set, and the downstream network has a mechanism to dynamically adjust the feature fusion weights according to the confidence of the retrieval results.

[0023] Furthermore, in step 3, modality amplification includes:

[0024] The Illava model is used to perform image-to-text processing on the video cover, generating text modal features:

[0025] ;

[0026] in Indicates the first The text modal features generated by modal amplification of the cover features of each sample using the llava model; The total number of words in the current sample. It is the first The first sample One token, It refers to the position of the word in the current sample. indivual;

[0027] For the video modality, take 10 frames at equal intervals and convert them into the image modality;

[0028] The text and image modal information are transformed into embeddings using the jina-clip-v2 model:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] in , , , , They are respectively , , , , Embedded, Indicates the first Video modal information of each sample, Indicates the first Title information of each sample, Indicates the first Comment information for each sample Indicates the first Cover information for each sample; This is a text feature embedding function based on the Jina-Clip-V2 model; This is an image feature embedding function based on the jina-clip-v2 model.

[0035] Furthermore, in step 4, the construction of the UGC knowledge vector base includes:

[0036] First, the text features of the title, cover text, and comments are concatenated. Then, the Jina-clip-v2 model is used to generate long text embeddings. The relevant formulas are as follows:

[0037] ;

[0038] ;

[0039] in Text embedding for UGC samples; It represents splicing; For text embedding in the concatenated UCG knowledge vector library; , , These correspond to the first in the UCG knowledge vector base. The title text features, cover text features, and comment text features of each sample are modally amplified.

[0040] Furthermore, in step 5, dense vector similarity retrieval is achieved by calculating cosine similarity, as shown in the following formula:

[0041] ;

[0042] in Text embedding representing UGC samples With the UCG knowledge vector base j vectors Cosine similarity; The total number of documents in the UGC knowledge vector library;

[0043] After sorting by cosine similarity in descending order, the Top-J results are retrieved and denoted as a set. .

[0044] Furthermore, in step 5, the relevance score calculation formula for BM25 retrieval based on word frequency is as follows:

[0045] ;

[0046] ;

[0047] in Score for BM25; For the UCG knowledge vector base The text content of the document; For query Terms in; Inverse document frequency; For terms exist The frequency in; and To adjust the parameters; For document Length; The average length of all documents in the UCG knowledge vector library; For included terms The number of documents;

[0048] After sorting the results in descending order of BM25 score, the Top-J results are retrieved and denoted as the set. .

[0049] Furthermore, in step 5, the fusion score of the RFF algorithm is:

[0050] ;

[0051] in For the fusion score; For the UCG knowledge vector base k One vector; For retrieval methods, For dense vector similarity retrieval, This is a BM25 search based on term frequency. For smoothing hyperparameters; for Ranking in the corresponding search results: for Dense vector similarity search results The ranking in for Corresponding document BM25 results Ranking in;

[0052] according to After sorting in descending order, select the Top-K highest-scoring items. The final recall set is obtained:

[0053] ;

[0054] in This refers to the recall set, i.e., the RAP results; These are the top K vectors sorted from highest to lowest score.

[0055] Furthermore, in step 6, the model training phase adopts a supervised learning paradigm, the optimizer is AdamW, and a dynamic learning rate decay strategy is set; the input of the model is the feature representation after multimodal fusion, the output is a continuous popularity score, and the training objective is to minimize the error between the predicted value and the true value.

[0056] Furthermore, the predictive performance of step 6 is evaluated using normalized mean square error and Spearman's rank correlation coefficient.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] This invention constructs a UGC knowledge vector base and designs a retrieval augmentation prediction (RAP) method. It employs a dual-path retrieval, high-speed recall, and content-decoupled noise reduction architecture to effectively address the performance bottleneck of existing UGC popularity prediction methods. Specifically, the dual-path retrieval and re-ranking recall architecture solves the problem of long text representation bias; the decoupled noise reduction method in the downstream deep learning network module, which dynamically adjusts feature fusion weights based on confidence, solves the problem of retrieval noise interference; and the overall architecture design overcomes the strong coupling defect of the model, enabling the model to adapt to dynamic updates of the retrieval database, thus improving its versatility and generalization ability. Attached Figure Description

[0059] Figure 1 This is a flowchart of the method of the present invention.

[0060] Figure 2 A schematic diagram is constructed for the features.

[0061] Figure 3 This is a schematic diagram of dual-path retrieval and rearrangement recall.

[0062] Figure 4 A flowchart for retrieving enhanced prediction results (RAP results) to assist in predicting the popularity of UGC on downstream networks. Detailed Implementation

[0063] In order to provide a clearer understanding of the technical features, objectives and beneficial effects of the present invention, the technical solution of the present invention will now be described in detail below, but it should not be construed as limiting the scope of implementation of the present invention.

[0064] This invention provides a method for predicting the popularity of user-generated content based on retrieval-enhanced prediction, the flowchart of which is shown below. Figure 1 As shown, the method includes the following steps:

[0065] Step 1: Data Collection;

[0066] We conduct systematic data collection on UGC websites, regularly crawling relevant information from UGC content, covering video content and its multimodal characteristics. Specifically, this includes the following dimensions:

[0067] UGC content video stream: used to extract visual features;

[0068] UGC content video titles typically contain keywords, trending topics, etc., and have high semantic value;

[0069] UGC content comments reflect user interaction and are rich in audience emotions and attitudes.

[0070] UGC content video covers: rich in visual information, which greatly affects click-through rates;

[0071] Likes, views, and comments on UGC content: As core behavioral indicators of the popularity of UGC content, they serve as significant monitoring signals.

[0072] Step 2: Data preprocessing;

[0073] To improve the effectiveness and generalization ability of model training, the original data needs to be preprocessed first. The specific process is as follows:

[0074] Data cleaning: Remove records with incorrect formats, standardize data encoding formats (such as UTF-8), field naming, and units;

[0075] Duplicate removal: Eliminate duplicate content to avoid bias in the model;

[0076] Modal missing processing: Remove samples with missing modalities (such as no title, no cover, no comments, etc.) to reduce noise interference;

[0077] Text and image standardization: performing word segmentation and stop word removal on text fields.

[0078] After data cleaning and standardization, the samples were divided into three parts according to the three-stage requirements of UCG knowledge vector base construction-training-testing:

[0079] UGC knowledge vector library (accounting for 50%): serving as the basic feature library for matching and retrieval, supporting RAP retrieval calculation;

[0080] Training set (40%): used for model learning and parameter optimization;

[0081] Test set (10%): Used for model performance evaluation.

[0082] The overall data partitioning ensures randomness and distribution consistency, avoiding training-test bias (data leakage).

[0083] Step 3: Feature construction;

[0084] like Figure 2 As shown, the multimodal feature information extracted from UGC includes three categories: video, text, and image. The specific process of feature construction is as follows:

[0085] Modality augmentation: For video cover images, the Illava model is used for image-to-text processing to augment the modality and increase the diversity of modal representations. For video modalities, 10 frames are taken at equal intervals and converted into image modalities.

[0086] Embedding transformation: Text and image modal information are transformed into embeddings using the jina-clip-v2 model.

[0087] The variables involved are defined as shown in Equations 1-10:

[0088] Formula 1: ;

[0089] in Indicates the first Video modal information for each sample; For video frame rate, It is the first The first sample frame.

[0090] Formula 2: ;

[0091] in Indicates the first Title information for each sample; For the number of words in the title, It is the first The first sample A token.

[0092] Formula 3: ;

[0093] in Indicates the first Comment information for each sample; It represents splicing; For the number of comments, It is the first The first sample 1 comment.

[0094] Formula 4: ;

[0095] in Indicates the first Cover information for each sample; Represents the set of real numbers; Represents the height of the image; Represents the width of the image; This represents the number of channels in the image.

[0096] Formula 5: ;

[0097] in Indicates the first The text modal features generated by modal amplification of the cover features of each sample using the llava model; The total number of words in the current sample. It is the first The first sample One token, It refers to the position of the word in the current sample. indivual.

[0098] Formula 6: ;

[0099] Formula 7: ;

[0100] Formula 8: ;

[0101] Formula 9: ;

[0102] Formula 10: ;

[0103] in , , , , They are respectively , , , , Embedding; This is a text feature embedding function based on the Jina-Clip-V2 model; This is an image feature embedding function based on the jina-clip-v2 model.

[0104] Step 4: Constructing a UGC knowledge vector library;

[0105] The UCG knowledge vector library is built for text modalities, and the specific process is as follows:

[0106] To reduce the computational cost of the final prediction model, all text features are first concatenated, and then the Jina-Clip-V2 model is used to generate long text embeddings. The relevant formulas are as follows:

[0107] Formula 11: ;

[0108] Formula 12: ;

[0109] in Text embedding for UGC samples (used for training and retrieval queries); For text embedding in the concatenated UCG knowledge vector library; , , These correspond to the first in the UCG knowledge vector base. The title text features, cover text features, and comment text features of each sample are modally amplified.

[0110] Step 5: Dual-path retrieval and rearrangement recall;

[0111] Text embedding for UGC samples (in Indicates the first One sample, (The dimension of the dense vector), in the UGC knowledge vector base ( To perform the following retrieval and fusion operations on the total number of documents in the UGC knowledge vector base (for intuitive representation), Figure 3 (Displayed in human-readable style)

[0112] (1) Dense vector similarity retrieval ( Figure 3 (Mid-Retriever side story)

[0113] calculate The cosine similarity with all vectors in the UCG knowledge vector base is calculated using the following formula:

[0114] Formula 13: ;

[0115] in express With the UCG knowledge vector base j vectors The cosine similarity.

[0116] After sorting by cosine similarity in descending order, the Top-J results are retrieved and denoted as a set. .

[0117] (2) BM25 retrieval based on term frequency ( Figure 3 (Middle BM25 branch line)

[0118] The relevance score is calculated using the BM25 algorithm based on word frequency statistics, and the relevant formula is as follows:

[0119] Formula 14: ;

[0120] Formula 15: ;

[0121] in Score for BM25; For the UCG knowledge vector base The text content of the document; For query Terms in; Inverse document frequency; For terms exist The frequency in; and To adjust parameters (usually) , ); For document Length; The average length of all documents in the UCG knowledge vector library; The total number of documents in the UGC knowledge vector library; For included terms The number of documents.

[0122] After sorting the results in descending order of BM25 score, the Top-J results are retrieved and denoted as the set. .

[0123] (3) RFF (Rank Fusion Function) algorithm dual-path fusion recall ( Figure 3 (Middle RFF branch)

[0124] The dense vector similarity retrieval results and the BM25 retrieval results are input into the RFF algorithm to calculate the fusion score, as shown in the following formula:

[0125] Formula 16: ;

[0126] in For the fusion score; For the UCG knowledge vector base k One vector; For retrieval methods, For dense vector similarity retrieval, This is a BM25 search based on term frequency. For smoothing hyperparameters; for Ranking in the corresponding search results: for Dense vector similarity search results The ranking in (set to when not appearing) ), for Corresponding document BM25 results The ranking in (set to when not appearing) ).

[0127] according to After sorting in descending order, select the Top-K highest-scoring items. The final recall set is obtained:

[0128] Formula 17: ;

[0129] in For the recall set (i.e., the RAP results); These are the top K vectors sorted from highest to lowest score. .

[0130] Step 6: Retrieve Enhanced Prediction Results (RAP results) to assist in predicting the popularity of UGC in downstream networks;

[0131] Combining a UGC popularity prediction method based on dual-path retrieval and reordering recall architecture, and utilizing existing deep learning network models suitable for UGC popularity prediction (such as attention mechanism networks, multimodal fusion networks) or existing machine learning models (such as CatBoost, SVR), a more powerful end-to-end deep learning model (whose core is a downstream deep learning network module) is constructed for UGC popularity prediction. The specific process is as follows ( Figure 4 ):

[0132] Training phase: The model's input is the feature representation after multimodal fusion. As shown below:

[0133] For each video modality in the training set In UGC knowledge vector base A retrieval based on a dual-path retrieval and reordering recall architecture is performed to obtain the Top-K set. ; Find its characteristic mean is :

[0134] Formula 18: ;

[0135] For each title modality in the training set In UGC knowledge vector base A retrieval based on a dual-path retrieval and reordering recall architecture is performed to obtain the Top-K set. ; Find its characteristic mean is :

[0136] Formula 19: ;

[0137] For each comment modality in the training set In UGC knowledge vector base A retrieval based on a dual-path retrieval and reordering recall architecture is performed to obtain the Top-K set. ; Find its characteristic mean is :

[0138] Formula 20: ;

[0139] For each cover modality in the training set In UGC knowledge vector base A retrieval based on a dual-path retrieval and reordering recall architecture is performed to obtain the Top-K set. ; Find its characteristic mean is :

[0140] Equation 21: ;

[0141] For each comment modality in the training set In UGC knowledge vector base A retrieval based on a dual-path retrieval and reordering recall architecture is performed to obtain the Top-K set. ; Find its characteristic mean is :

[0142] Equation 22: ;

[0143] For general deep learning prediction networks The prediction result is:

[0144] Equation 23: ;

[0145] in ;

[0146] Output It is a continuous popularity score.

[0147] The training objective is to minimize the error between the predicted values ​​and the true values ​​(the number of likes, views, and comments for UGC content in step 1). The objective is:

[0148] Formula 24: ;

[0149] Where MSE is the mean squared error. nFor batch size, Let be the popularity label value of the i-th sample. Let be the predicted popularity value for the i-th sample.

[0150] A standard supervised learning paradigm is adopted, with AdamW selected as the optimizer, and a dynamic learning rate decay strategy is implemented to improve convergence stability. Through this process, a trained deep learning model is ultimately obtained.

[0151] Prediction phase: The trained deep learning model outputs accurate predictions of UGC content popularity based on the multimodal features of the input target samples.

[0152] By employing a UGC popularity prediction method based on a dual-path retrieval and reordering recall architecture, the number of network parameters in the deep learning system can be effectively reduced, resulting in more accurate predictions with less computation. The RAP method not only captures deep semantic relationships but also preserves the accuracy of keyword matching, enabling the prediction model to make more reliable inferences based on the true popularity of similar content. Simultaneously, the retrieved relevant cases provide intuitive explanations for the prediction results, enhancing the model's credibility. Furthermore, the RAP method requires the addition of a mechanism in the downstream deep learning network module to dynamically adjust the feature fusion weights based on the confidence level of the retrieval results, in order to avoid introducing noise—if the vector retrieval library is of poor quality and the Top-K information is irrelevant to the target UGC content, strong noise will be introduced. The dynamic adjustment of feature fusion weights based on the confidence level of the retrieval results... This is implemented in [the system / process]. The weight of each search library vector is [weight] relative to the current sample vector. Similarity: The greater the similarity, the higher the weight. For example, weight for:

[0153] Formula 25: ;

[0154] Other vector calculation methods are similar.

[0155] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0156] Example 1: Practical application verification of the present invention in a UGC popularity prediction scenario;

[0157] 1. Detection accuracy verification experiment;

[0158] To verify the advantage of the RFF algorithm's dual-path fusion recall in terms of retrieval accuracy, it was compared with retrieval based solely on dense vector similarity (ES) and retrieval based on word frequency (BM25). Evaluation metrics included Recall@K and MRR@K. Recall@K represents the percentage of relevant results among the top K UGC knowledge vector base retrieval results; MRR@K represents the average reciprocal rank, i.e., the reciprocal of the position of the first relevant result among the top K UGC knowledge vector base retrieval results. As shown in Equation 26:

[0159] Equation 26: ;

[0160] in It is a collection of UGC content to be retrieved from the UGC knowledge vector base; It is the number of the set; It is the first In the query results for each item to be retrieved, the position of the first relevant result (if not at the beginning) Then it is recorded as 0).

[0161] The results are shown in Table 1:

[0162] Table 1 Search Results

[0163]

[0164] As shown in Table 1, RFF (RFF algorithm dual-path fusion recall) is significantly better than ES and BM25 in all indicators. MRR@10 is 6.76% higher than ES, and Recall@1 is 6.96% higher than ES. It effectively solves the problem of poor accuracy of long text embedding and greatly enhances the retrieval accuracy.

[0165] 2. Comparison experiment based on calculation time;

[0166] To verify the computational efficiency advantage of the RFF algorithm's dual-path fusion recall, its computation time was compared with that of the large language model Qwen3-Beranker-0.6B. The experimental conditions were Top-J=1000 Recall (recalling the first 1000 results).

[0167] The results are shown in Table 2:

[0168] Table 2 Calculation Time Results

[0169]

[0170] As can be seen from Table 2, the computation time of RFF (RFF algorithm dual-path fusion recall) (1.2858s) is much lower than that of the large language model (34.3237s), which significantly reduces the computation cost and has a clear efficiency advantage.

[0171] 3. Downstream network prediction performance verification experiment;

[0172] To verify the performance improvement effect of the RAP method on the popularity prediction model, experiments were conducted on the MicroLens-100k-Dataset (training set to test set ratio 0.8:0.2). Comparison models included the TMALL model (a transductive model for multimodal domains), CatBoost (an algorithm based on Gradient Boosting Decision Trees (GBDT)), SVR (a regression version of Support Vector Machine (SVM), a support vector regression model; this experiment uses a Gaussian kernel function, suitable for nonlinear data), and versions combining RAP. Evaluation metrics were NMSE and SRC. NMSE is the normalized mean squared error, as shown in Equation 27:

[0173] Equation 27: ;

[0174] Where MSE is the mean squared error; It is the variance of the true value. , It is the mean of the true values; It is the actual value; It is a predicted value; That is the number of samples.

[0175] SRC is the Spearman rank correlation coefficient, used to measure the monotonic correlation between two variables. It is calculated based on the variable's rank. As shown in Equation 28:

[0176] Equation 28: ;

[0177] in , representing the difference between the actual ranking and the predicted ranking; That is the number of samples.

[0178] The results are shown in Table 3:

[0179] Table 3. Experimental Results of Downstream Networks

[0180]

[0181] As shown in Table 3, after adding the RAP method, the NMSE of all baseline models decreased significantly (prediction error decreased) and the SRC increased significantly (correlation between the true value and the predicted value increased), indicating that the RAP method can effectively assist downstream networks in improving the prediction performance of UGC popularity.

[0182] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for predicting the popularity of user-generated content based on retrieval-enhanced prediction, characterized in that, Includes the following steps: Step 1: Data Collection; Collect multimodal data from UGC websites, including video streams, titles, comment content, cover images, and popularity metrics such as likes, views, and comments. Step 2: Data preprocessing; The collected raw data is cleaned, duplicates are removed, modal missing items are processed, and text and images are standardized. The samples are then divided into a UGC knowledge vector base set, a training set, and a test set. Step 3: Feature construction; Multimodal embedding representations are generated by extracting features from three modalities: video, text, and image, and then performing modality augmentation and embedding transformation. Step 4: Constructing a UGC knowledge vector library; For text modalities, all text features are concatenated and long text embeddings are generated to form a UCG knowledge vector library; Step 5: Dual-path retrieval and rearrangement recall; For text embeddings of UGC samples, dense vector similarity retrieval and BM25 retrieval based on word frequency are performed sequentially in the UGC knowledge vector base. Then, the Top-K results are retrieved through a dual-path fusion using the RFF algorithm to obtain the final recall set. Step 6: Retrieve enhanced prediction results to assist in predicting the popularity of UGC on downstream networks; Combining deep learning models, model training and popularity prediction are performed based on the input multimodal features and recall set, and the downstream network has a mechanism to dynamically adjust the feature fusion weights according to the confidence of the retrieval results; In step 4, the construction of the UGC knowledge vector base includes: First, the text features of the title, cover text, and comments are concatenated. Then, the Jina-clip-v2 model is used to generate long text embeddings. The relevant formulas are as follows: ； ； in Text embedding for UGC samples; This is a text feature embedding function based on the Jina-Clip-V2 model; Indicates the first Title information for each sample; Indicates the first The text modal features generated by modal amplification of the cover features of each sample using the llava model; Indicates the first Comment information for each sample; It represents splicing; For text embedding in the concatenated UCG knowledge vector library; , , These correspond to the first in the UCG knowledge vector base. The title text features, cover text features, and comment text features of each sample are modally amplified.

2. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 1, characterized in that, In step 3, modality amplification includes: The Illava model is used to perform image-to-text processing on the video cover, generating text modal features: ； in Indicates the first The text modal features generated by modal amplification of the cover features of each sample using the llava model; This represents the total number of words in the current sample. It is the first The first sample One token, It refers to the position of the word in the current sample. indivual; For the video modality, take 10 frames at equal intervals and convert them into the image modality; The text and image modal information are transformed into embeddings using the jina-clip-v2 model: ； ； ； ； ； in , , , , They are respectively , , , , Embedded, Indicates the first Video modal information of each sample, Indicates the first Title information of each sample, Indicates the first Comment information for each sample Indicates the first Cover information for each sample; This is a text feature embedding function based on the Jina-Clip-V2 model; This is an image feature embedding function based on the jina-clip-v2 model.

3. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 1, characterized in that, In step 5, dense vector similarity retrieval is achieved by calculating cosine similarity, as shown in the following formula: ； in Text embedding representing UGC samples With the j-th vector in the UCG knowledge vector base Cosine similarity; The total number of documents in the UGC knowledge vector library; After sorting by cosine similarity in descending order, the Top-J results are retrieved and denoted as a set. .

4. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 3, characterized in that, In step 5, the relevance score of the BM25 retrieval based on word frequency is calculated using the following formula: ； ； in Score for BM25; For the UCG knowledge vector base The text content of the document; For query Terms in; Inverse document frequency; For terms exist The frequency in; and To adjust the parameters; For document Length; The average length of all documents in the UCG knowledge vector library; For included terms The number of documents; After sorting the results in descending order of BM25 score, the Top-J results are retrieved and denoted as the set. .

5. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 4, characterized in that, In step 5, the fusion score of the RFF algorithm is: ； in For the fusion score; This is the k-th vector in the UCG knowledge vector library; For retrieval methods, For dense vector similarity retrieval, This is a BM25 search based on term frequency. For smoothing hyperparameters; for Ranking in the corresponding search results: for Dense vector similarity search results The ranking in for Corresponding document BM25 results Ranking in; according to After sorting in descending order, select the Top-K highest-scoring items. The final recall set is obtained: ； in This refers to the recall set, i.e., the RAP results; These are the top K vectors sorted from highest to lowest score.

6. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 1, characterized in that, In step 6, the model training phase adopts a supervised learning paradigm, the optimizer is AdamW, and a dynamic learning rate decay strategy is set; the input of the model is the feature representation after multimodal fusion, the output is the continuous popularity score, and the training objective is to minimize the error between the predicted value and the true value.

7. The method for predicting the popularity of user-generated content based on retrieval enhancement prediction according to claim 1, characterized in that, The predictive performance of step 6 is evaluated using normalized mean square error and Spearman's rank correlation coefficient.

Images