Multimedia content popularity prediction method based on large language model text enhancement

Abstract

This invention discloses a multimedia content popularity prediction method based on large language model text enhancement. The deep learning model used includes a time-series encoder, a large language model with frozen parameters, a cross-attention module, and a linear layer. The method includes: preprocessing the time-series data of information forwarding volume; fusing segmented semantic descriptions for the time-series data through a segmented text prototype retrieval module; generating dynamic text descriptions for each time-series sample based on a sequence-level prompt generation module; employing a two-stage cross-attention mechanism to complete the fusion of time-series and textual dual-modal information; and finally outputting the prediction result after freezing the large language model and the linear mapping layer. This method enriches the time-series data with dual-scale semantic enhancement, improves prediction accuracy by leveraging the knowledge advantages of the large language model, and is particularly suitable for scenarios such as sudden hot topics and public opinion fermentation. At the same time, it relies on a lightweight architecture to ensure inference efficiency and meet the needs of large-scale business implementation.

Description

Technical Field

[0001] This invention belongs to the field of social media data mining and time series prediction technology, specifically a method for predicting the popularity of multimedia content. Background Technology

[0002] With the rapid development of mobile internet technology and the widespread adoption of smart terminals, social media has deeply penetrated all aspects of public life, evolving into a core carrier for people to obtain information, express opinions, and disseminate information. A deep understanding of the inherent laws of information dissemination and accurate prediction of its popularity at specific future points in time are of paramount practical significance for scenarios such as public opinion monitoring and crisis intervention by government departments, optimization of advertising strategies by commercial institutions, and content distribution and recommendation on social platforms. Social media platforms possess unique characteristics such as rapid information dissemination, dramatic fluctuations, strong interactivity, and rapid increases and decreases in popularity within a short period. The time-series popularity sequences formed by blog post reposts generally suffer from problems such as short observation periods, sparse early data, irregular fluctuations, and a disconnect between semantic information and numerical time-series characteristics. Accurately predicting the future popularity trends of topics and individual blog posts on social media has extremely strong practical application value for platforms to optimize content recommendation strategies, for regulatory departments to conduct real-time public opinion early warnings, for commercial entities to control the effectiveness of marketing and communication, and for emergency departments to handle sudden public opinion crises.

[0003] Existing methods for predicting the popularity of social media content are mainly divided into two categories: cascade prediction and time series prediction. Neither of these methods can meet the actual needs of Weibo. Traditional cascade prediction methods require the construction of a global user propagation graph. For Weibo's massive user base and high-concurrency propagation data, this can lead to problems such as excessive memory consumption, high computational complexity, and poor real-time performance. It cannot meet the business requirements of second-level updates and real-time prediction of post popularity. Conventional time series prediction methods are prone to "information poverty" when dealing with short observation sequences of post forwarding volume on multimedia platforms (such as 0.5 hours or 1 hour after posting). This results in low prediction accuracy and an inability to capture typical propagation scenarios such as sudden hot topics and reversals of public opinion.

[0004] Existing technologies related to time series prediction using large language models are mostly designed for general time series prediction. They are not specifically optimized for the time series characteristics and popularity fluctuation patterns of content dissemination on multimedia platforms. Their cross-modal fusion methods are rigid and cannot be efficiently adapted to popularity prediction scenarios. Moreover, most solutions require fine-tuning of the underlying parameters of large models, resulting in high computing costs and deployment difficulties, making it difficult to scale up in real business scenarios.

[0005] To address the pain points of predicting the popularity of multimedia content, existing technologies cannot simultaneously achieve real-time performance, accuracy, and feasibility. There is an urgent need for a specialized solution that is adapted to the characteristics of multimedia platform content dissemination data, lightweight in deployment, and highly accurate. Summary of the Invention

[0006] In response to existing technologies, this invention provides a multimedia content popularity prediction method based on large language model text enhancement. This method does not require the construction of complex user propagation graphs or fine-tuning of the underlying parameters of large models. It achieves lightweight, highly accurate, and real-time popularity prediction by relying on dual-scale semantic enhancement and cross-modal information fusion.

[0007] To address the practical challenges of predicting popularity in multimedia content—such as short time-series data, sparse information, fragmentation between text and time-series modal features, poor real-time performance of traditional methods, and high deployment costs of large models—this invention proposes a multimedia content popularity prediction method based on text enhancement using a large language model, comprising:

[0008] Acquire relevant data from multiple social media posts, including post timestamps, total number of reposts within the predicted time period, and timestamps of user reposts; input the preprocessed data into a trained multimedia content popularity prediction model, which includes a Transformer-based time-series encoder, a large language model with frozen parameters, two cross-attention modules, and two linear layers; the model outputs the final number of reposts for the predicted time.

[0009] The training of the multimedia content popularity prediction model includes: acquiring and preprocessing relevant data from multiple social media posts; performing dual-scale semantic enhancement using a block-level text prototype retrieval module and a sequence-level prompt generation module to enrich the time-series information of post repost volume; employing a two-stage cross-attention mechanism to fuse time-series and text representation information, wherein: the first-stage cross-attention mechanism uses time-series features to guide the improvement of text representation, thereby highlighting text segments semantically relevant to the target task; the second-stage cross-attention mechanism calibrates time-series features by combining text semantics, enhancing the capture of key time-series patterns consistent with semantic logic; and inputting the data after the above two-stage information fusion into a large language model with frozen parameters for prediction, leveraging the knowledge richness of the large language model to improve popularity prediction performance.

[0010] Furthermore, in the multimedia content popularity prediction method of the present invention:

[0011] The Transformer-based time-series encoder receives time-series input; the frozen-parameter large language model receives the time-series output by the block-level text prototype retrieval module and the text prompts output by the sequence-level prompt generation module, and outputs the feature encoding of the text prompts; of the two cross-attention modules, one cross-attention module receives the time-series fused with block-level semantic information and the text prompts describing the time-series, and the other cross-attention module receives the time-series fused with the first-stage information fusion and the time-series fused with the sequence-level semantic information; the frozen-parameter large language model receives the time-series fused with the second-stage information fusion and outputs the feature encoding of the fused time-series; of the two linear layers, one linear layer receives domain vocabulary input, and the other linear layer receives the output of the large language model.

[0012] The steps for preprocessing relevant data from multiple social media posts are as follows: First, read in the time-series dataset of repost volumes from the multiple social media posts. Then, filter data based on the posting timestamp and total repost volume, identifying posts published outside of normal times or with very few or very many reposts. Next, accumulate the repost count starting from the posting time, recording it every t minutes. up to the observation time Construct the input time series Each selected dataset is divided into a training set, a validation set, and a test set in a certain proportion for training, validation, and result testing, respectively.

[0013] The semantic enhancement steps of the block-level text prototype retrieval module are as follows: First, the constructed input time series is displayed using a sliding window technique. The sequence is divided into multiple time segments; the segment length is set to p, and the sliding step is s; for a sequence of length o, a set of segments is generated. , where the i-th segment ,and Secondly, segment feature mapping: An encoder based on a Transformer architecture is used to map each time segment... Mapping from the time dimension to the feature dimension yields the initial feature representation of the segment. Then, prototype retrieval and feature fusion: calculating fragment features. With text prototype library For each prototype vector, calculate the cosine similarity and select the prototype vector with the highest score. This serves as the semantic label for the segment; the semantic label is added to the initial features of the segment to obtain the semantically enhanced segment features. By concatenating the blocks that incorporate semantic information, a time series containing semantic information is obtained, denoted as a semantically enhanced time series feature. : .

[0014] The text prototype library The construction steps are as follows: Construct a domain-specific seed thesaurus and define a set of seed words that are strongly correlated with the popularity trend of time series data. This includes words describing trends, words describing quantities, and words describing tasks; vocabulary filtering and dimensionality reduction are performed to obtain the original vocabulary embedding matrix V of the pre-trained large language model, and the vector of each word in the vocabulary is calculated according to the following formula. With seed word set The cosine similarity of all word vectors in the dataset.

[0015]

[0016] Set similarity threshold =0.75, retain similarity greater than The words form the filter word list The filtering vocabulary was analyzed using principal component analysis algorithm. The word vectors in the text are subjected to dimensionality reduction processing, reducing the dimension from Dimensional reduction Set it to 500.

[0017] The semantic enhancement steps of the sequence-level prompt generation module are as follows: First, a prompt template is constructed, which includes a dataset description, a data description, and a trend description. The dataset description indicates that the input sequence is a time series of the number of times a post has been forwarded. The data description inputs the time series values ​​as additional text. The trend description pre-calculates the overall trend value and growth rate representing the overall change and inserts them into the prompt. A unique prompt text is generated for each input sample using the prompt template. The input parameters of the fully filled prompt text are frozen in a large language model to encode the features of the prompt text. The output of the last hidden state of the parameter-frozen large language model is extracted as the prompt feature representation. .

[0018] The step of fusing time series and text representation information using a two-stage cross-attention mechanism includes constructing a first-stage and a second-stage cross-attention mechanism. The first-stage cross-attention mechanism is constructed using semantically enhanced time series features obtained in the block-level text prototype retrieval step. As a query vector, the prompt features obtained in the sequence-level prompt generation step are used. As key and value vectors, intermediate features are calculated using a multi-head attention mechanism. The calculation formula is as follows: The second-stage cross-attention mechanism is constructed based on the aforementioned intermediate features. As a query vector, with the aforementioned semantically enhanced time series features The alignment feature Z is calculated as the key vector and value vector according to the following formula: .

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] (1) This invention effectively solves the problem of information scarcity caused by short observation sequences in social media content popularity prediction by designing a block-level text prototype module and a sequence-level prompt generation module for dual-scale information enhancement. Compared with traditional time series methods that rely solely on numerical fitting, this invention utilizes the rich semantic knowledge base of a large language model to map abstract numerical changes into specific semantic concepts, significantly improving the model's ability to capture sudden trends and long-tail effects.

[0021] (2) This invention adopts an architecture that combines pure time series data with text, avoiding the high computational cost of constructing and processing large-scale cascaded graphs. Compared with deep neural network methods based on cascaded graphs, this invention significantly reduces memory usage and computational complexity while ensuring prediction accuracy, and can be applied more efficiently to real-time prediction scenarios of large-scale social media data.

[0022] (3) This invention achieves deep alignment between continuous numerical modalities and discrete text modalities through a cross-modal attention mechanism, overcoming the modal gap problem that exists when directly inputting time series into a large language model. This design allows the model to retain the precise numerical characteristics of the time series while making full use of the general reasoning capabilities of the large language model, thereby achieving better prediction performance than existing technologies on multiple real datasets. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the multimedia content popularity prediction method described in this invention.

[0024] Figure 2 This is the present invention. Figure 1 The diagram shows the training process of the popularity prediction model. Detailed Implementation

[0025] The design concept of this invention is to provide a time series prediction framework based on LLM text enhancement for the field of popularity prediction. The deep learning model used includes a Transformer-based time series encoder, a large language model with frozen parameters, two cross-attention modules, and two linear layers. The method enriches the semantic information of time series features at two scales through block-level text prototype retrieval and sequence-level prompt generation modules, and fuses the information through a cross-attention mechanism before inputting it into the frozen LLM for prediction. The core method is reflected in three levels: First, by calculating the similarity between the block-level time series and the vectors in the LLM vocabulary, the semantic information that best describes the time series is obtained. Second, specific prompt descriptions are dynamically generated for each time series, providing sequence-level semantic information. Finally, information is fused through two cross-attention operations and input into the frozen LLM for prediction.

[0026] This method constructs a dual-scale information enhancement framework that combines a block-level text prototype retrieval mechanism with a sequence-level cue generation strategy. Specifically, the method mainly includes block-level text prototype retrieval, sequence-level cue generation, and cross-modal alignment and prediction.

[0027] I. Chunk-level Text Prototype Retrieval: To address the information scarcity caused by short time series sequences, a domain-specific filtering vocabulary is first constructed. This retains words from the large language model vocabulary related to descriptions of time trends and popularity predictions (such as "rising" and "surge"), and performs dimensionality reduction. Next, the input time series is divided into multiple time segments (Patches). The matching score between the features of each segment and the word vectors in the filtered vocabulary is calculated. The most matching text prototype vector is retrieved and fused with the original segment features. This mechanism aims to inject discrete semantic labels into continuous time series segments, enriching the feature representation of the time series at a micro-level.

[0028] II. Sequence-Level Prompt Generation: To guide the large language model in understanding the overall context of the time series and downstream tasks, a structured text prompt is dynamically generated for each prediction sample. This prompt template consists of three parts: a dataset description (indicating the task background), a data description (converting the numerical sequence into text form), and a trend description (including statistical indicators such as the total change and growth rate of the sequence). The generated prompt is input into the large language model with frozen parameters to obtain a prompt feature representation rich in macro-semantic information.

[0029] III. Cross-modal Alignment and Prediction: A bidirectional cross-attention mechanism is used to deeply fuse the semantically enhanced time-series features and cue text features. On the one hand, the time-series features guide the focus of textual information; on the other hand, the text semantics are used to calibrate the time-series features. Finally, the aligned features are input into a large language model (such as GPT-2) with frozen parameters, and the future popularity prediction is output by fine-tuning the linear layer.

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.

[0031] The problem defined in this invention is as follows: Posts are sampled after publication, and the number of reposts is accumulated starting from the publication time, with a record taken every t minutes. up to the observation time Construct the input time series The method of this invention performs multi-step prediction based on the time series within the observation period to obtain the predicted time. The predicted result Y within , That is, the number of forwards at the predicted time point.

[0032] Explanation of text prototypes: Pre-trained LLMs possess their own vocabularies, each containing corresponding embedding vectors, which are defined as text prototypes. Text prototypes retrieved from the vocabulary of a large language model provide high-level semantic anchors, complementing purely numerical temporal embeddings. While time-series representations primarily capture local dynamics, text prototypes encode abstract temporal concepts (such as growth or saturation) learned from large-scale corpora.

[0033] Regarding the datasets, this embodiment uses three publicly available datasets for the field of social media popularity to illustrate the time series prediction model proposed in this invention. The specific datasets used are as follows:

[0034] 1. Weibo2016 Dataset. Weibo is one of China's most popular social media platforms. This dataset contains the reposting data of posts published on July 1, 2016, during this period, including reposting user information and reposting timestamps, which can be used to obtain cascade graphs and time series of reposting volume.

[0035] 2. Twitter 2012 Dataset. This dataset contains retweets of tweets posted on Twitter between March 24 and April 25, 2012, covering retweet user information and retweet timestamps, and can be used to obtain cascade graphs and time series of retweet volume.

[0036] 3. Weibo2021 Dataset. This dataset contains the number of reposts of posts published on the Weibo platform on June 20, 2021, providing a time series of repost counts, but without cascading information.

[0037] For datasets with cascading data, the forwarding timestamps of forwarding users can be statistically analyzed to obtain a numerical time series of forwarding volume. For the Weibo2021 dataset, the time series can be directly partitioned. After obtaining the forwarding time series, analysis and prediction can be performed. The specific steps are as follows:

[0038] Step 1) Data preprocessing and sequence generation, including:

[0039] 1-1) Data Preprocessing: Read in existing time-series datasets of social media repost volumes, and filter data based on post timestamps and total repost volume, selecting data posted outside of normal times or with very few or very many reposts. Specifically, for the Weibo dataset, filter posts posted between 8:00 AM and 6:00 PM based on post timestamps, and control the total repost volume between 10 and 500.

[0040] 1-2) Time series generation: Accumulate the number of reposts starting from the publication time, and record once every t minutes. up to the observation time Construct the input time series For the Weibo dataset, observation times were set to 0.5 hours, 1 hour, and 2 hours, with a prediction time of 24 hours after publication. For the Twitter dataset, observation times were 1 day, 2 days, and 3 days, with a prediction time of 30 days after publication.

[0041] 1-3) Data Partitioning: Each filtered dataset is divided into a training set, a validation set, and a test set according to a certain ratio for training, validation, and result testing, respectively. In this embodiment, the method of the present invention uses the first 70% of instances as the training set for model training, the middle 15% as the validation set for model validation, and the last 15% as the test set for testing the model's performance, following the index order.

[0042] Step 2) Chunk-level text prototype retrieval, the specific steps are as follows:

[0043] 2-1) Divide the long sequence X generated in step 1-2) into multiple time segments using the sliding window technique. Set the segment length to p and the sliding step size to s. For a sequence of length o, generate a set of segments. , where the i-th segment ,and In this embodiment, the Weibo dataset is sampled every 5 minutes. Therefore, the time series length is 6 for an observation period of 0.5 hours, 12 for an observation period of 1 hour, and 24 for an observation period of 2 hours. The sliding step is set to 1, and the segment length is set to 3. For the Twitter dataset, sampling occurs every 3 hours. Therefore, the time series length is 8 for an observation period of 1 day, 16 for a time series of 2 days, and 24 for a time series of 3 days. The sliding step is set to 2, and the segment length is set to 5.

[0044] 2-3) Segment Feature Mapping: Construct a linear projection layer to map each time segment... Inputting this linear layer maps it from the time dimension to the feature dimension, yielding the initial feature representation of the segment. .

[0045] 2-4) Construct a domain-specific seed thesaurus: Define a set of seed words that are strongly correlated with the popularity trend of time series data. Specifically, this includes words describing trends (such as "rise" and "surge"), words describing quantities (such as "many" and "small"), and words describing tasks (such as "series" and "popularity").

[0046] 2-5) Lexicon Filtering and Dimensionality Reduction: Obtain the original vocabulary embedding matrix V of the pre-trained large language model (e.g., GPT-2). Calculate the vector for each word in the vocabulary. With seed word set The cosine similarity of all word vectors in the dataset. The calculation formula is as follows:

[0047]

[0048] Set similarity threshold =0.75, retain similarity greater than The words form the filter word list To reduce computational complexity, Principal Component Analysis (PCA) algorithm is used. The word vectors in the text are subjected to dimensionality reduction processing, reducing the dimension from Dimensional reduction Set it to 500.

[0049] 2-6) Prototype retrieval and feature fusion: Calculating fragment features With text prototype library For each prototype vector, calculate the cosine similarity and select the prototype vector with the highest score. This serves as the semantic label for the segment. The semantic label is then added to the initial features of the segment to obtain the semantically enhanced segment features. By concatenating the various blocks that incorporate semantic information, a time series containing semantic information is obtained, denoted as... .

[0050] Step 3) Sequence-level hint generation, the steps are as follows:

[0051] 3-1) Prompt Template Construction: To fully utilize the function of prompts, the method of this invention designs a unique prompt for each sample. The prompt template consists of three parts:

[0052] Dataset Description. This section is described as "the time series of the number of reposts of this Weibo post...". It indicates that the input sequence is a time series of the number of reposts of a Weibo post. This helps LLM better understand the downstream task, thereby improving prediction performance.

[0053] Data description. This section is described as "These values ​​are..." Samples are taken every f minutes. Here, the time series values ​​are input as additional text. f represents the sampling interval, supplementing the input with additional information.

[0054] Trend Description. This section states, "The overall trend value is ∆, and the growth rate is r." Here, ∆ represents the overall change, and r represents the growth rate. These two values ​​will be pre-calculated and inserted into the prompt.

[0055] Using the prompt template above, the system will generate a unique prompt for each input sample.

[0056] 3-2) Cue Encoding: The output of the last hidden state of the pre-trained large language model (GPT-2) with the input parameters frozen and the cue text filled in is extracted as the cue feature representation. .

[0057] Step 4) Cross-modal alignment and feature fusion

[0058] 4-1) Construct the first stage of cross-attention: using the semantically enhanced time series features obtained in steps 2-6) As the query vector, the hint features obtained in step 3-2) As key and value vectors, intermediate features are calculated using a multi-head attention mechanism. The calculation formula is as follows. The first-stage cross-attention mechanism uses time-series features to guide the improvement of text representation, thereby highlighting text segments that are semantically relevant to the target task.

[0059]

[0060] 4-2) Construct the second-stage cross-attention: using the intermediate features obtained in step 4-1). As a query vector, using the original time series features As key and value vectors, the final alignment feature Z is calculated using the following formula. The second-stage cross-attention mechanism calibrates time-series features by incorporating textual semantics, which enhances the capture of key time-series patterns consistent with semantic logic.

[0061]

[0062] Deep cross-modal alignment between text and time-series data was achieved through a two-stage cross-attention operation, which also facilitated mutual information complementarity between the two modalities.

[0063] Step 5) Model prediction and training optimization, including:

[0064] 5-1) Backbone Network Processing: Input the feature Z output from step 4-2) into the pre-trained GPT-2 backbone network. To ensure the general knowledge and computational efficiency of LLM, freeze the parameters of GPT-2 and do not participate in gradient updates.

[0065] 5-2) Linear Projection Prediction: Take the hidden state vector of the last time step of the backbone network output, input it into a trainable linear layer, and output the predicted popularity value for the future time step. .

[0066] 5-3) Loss function calculation: The mean squared logarithmic error is used as the objective function, and the calculation formula is as follows:

[0067]

[0068] 5-4) Parameter fine-tuning: Using the Adam optimizer, based on the calculated loss... The trainable parameters in the model are iteratively updated until the loss converges.

[0069] Once trained, the model can receive a time series of forwarding volume over a period of time as input, provide a prediction sequence for the corresponding prediction time, and select the last prediction time point as the final forwarding volume at the prediction time.

[0070] Research materials:

[0071] This study uses three datasets: Weibo2016, Twitter2012, and Weibo2021. These datasets record the reposting information of posts published on Weibo in 2016 and 2021, and on Twitter in 2012, respectively. The mean squared logarithmic error (MSLE) is used as the evaluation metric for assessing the model's predictive performance. Table 1 compares the prediction results of the proposed method with those of CasFlow (Xovee Xu, Fan Zhou, and Kunpeng Zhang et al., “Casflow: Exploring hierarchical structures and propagation uncertainty for cascade prediction,” IEEE Trans. Knowl. Data Eng., vol. 35, no. 4, pp. 3484–3499, 2023.) and AFRF (HaoyuWang, Zongxia Xie, and Meiyao Liu et al., “AFRF: angle feature retrieval based popularity forecasting,” in ACM, 2023, pp. 2606–2615.) based on cascade graphs and traditional time series prediction methods. On the Weibo dataset, with input lengths of 6, 12, and 24 time steps, the prediction lengths were 282, 276, and 264 time steps, respectively. On the Twitter dataset, with input lengths of 8, 16, and 24 time steps, the prediction lengths were 232, 224, and 216 time steps, respectively. Since the Weibo2021 dataset does not contain cascading information, the cascading method will produce an empty result on this dataset.

[0072] Table 1. MSLE evaluation results of our method compared with those of cascade-based and traditional timing methods.

[0073]

[0074] As can be seen from the results in Table 1, the method of this invention has significant advantages over cascade-based prediction methods and traditional time-series prediction methods, achieving lower MSLE loss on multiple datasets. This demonstrates the effectiveness of the proposed paradigm in the field of popularity prediction. The powerful reasoning ability and rich knowledge of large language models are leveraged to improve prediction performance. Simultaneously, a dual-scale semantic information enhancement module is introduced to inject richer feature descriptions into time-series information, effectively addressing the problems of low efficiency caused by the growth of cascaded information and the scarcity of information in short time-series data sequences in the popularity prediction field.

[0075] Furthermore, to demonstrate the effectiveness of introducing large language models to assist time series tasks, and to further prove the effectiveness of the dual-scale semantic enhancement module, the method of this invention was also evaluated against general time series large model methods, including: OFA (Tian Zhou, Peisong Niu, and Xue Wang et al., “One fits all: Power generaltime series analysis by pretrained LM,” in NeurIPS, 2023.) and Time-LLM (MingJin, Shiyu Wang, Lintao Ma et al., “Time-LLM: Time Series Forecasting by Reprogramming Large Language Models,” in ICLR, 2024.), and the evaluation results are shown in Table 2.

[0076] Table 2. MSLE evaluation results of our method and the general time series large model method.

[0077]

[0078] As can be seen from the results in Table 2, the large model-based method has significant advantages over cascade-based and traditional time series methods, proving the effectiveness of the LLM-based approach in this invention. Furthermore, this invention's method also achieves better performance than general time series LLM methods. This indicates that the two semantic enhancement modules proposed in this invention for the popularity prediction task provide richer information during the inference process, thereby improving prediction performance.

[0079] Using existing mature methods and datasets, the data in Tables 1 and 2 demonstrate that our proposed method achieves the minimum prediction error at different observation times in the popularity prediction scenario. Table 1 shows that it achieves the best prediction performance compared to cascaded methods and traditional time series prediction methods, while Table 2 demonstrates the effectiveness of our proposed method compared to general time series LLM methods, showcasing its superior performance in the field of popularity prediction.

[0080] This demonstrates the effectiveness of the dual-scale semantic enhancement framework based on a large language model used in this invention, showing broad application prospects and significant application value. Compared to methods based on cascaded graphs, the method of this invention has lower computational cost and higher computational efficiency. Compared to other time-series prediction methods, the method of this invention achieves better prediction performance.

[0081] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are preferred application examples that demonstrate the core technical ideas of the present invention, and are merely illustrative and not restrictive. Those skilled in the art can make many improvements and changes under the guidance of the present invention without departing from the spirit of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the popularity of multimedia content based on large language model text enhancement, characterized in that, include: Acquire relevant data from multiple social media posts, including the post timestamp, the total number of reposts within the predicted timeframe, and the timestamps of user reposts. The preprocessed data is input into the trained multimedia content popularity prediction model, which includes a time series encoder based on the Transformer structure, a large language model with frozen parameters, two cross-attention modules, and two linear layers; the model outputs the final number of reposts of a post at the predicted time. The training of the multimedia content popularity prediction model includes: Acquire relevant data from multiple social media posts and preprocess them; By utilizing a block-level text prototype retrieval module and a sequence-level prompt generation module, dual-scale semantic enhancement is performed to enrich the time-series information of the number of posts forwarded. A two-stage cross-attention mechanism is adopted to fuse information from time series and text representations. The first stage of the cross-attention mechanism uses time series features to guide the improvement of text representations, thereby highlighting text segments that are semantically relevant to the target task. The second stage of the cross-attention mechanism combines text semantics to calibrate time series features, enhancing the capture of key time series patterns that are consistent with semantic logic. The data after the information fusion of the above two stages is input into the large language model with frozen parameters for prediction, and the knowledge richness of the large language model is used to improve the popularity prediction performance. 2.The method of claim 1, wherein, The time series encoder based on the Transformer structure receives time series input; The large language model with frozen parameters receives the time series output by the block-level text prototype retrieval module and the text prompts output by the sequence-level prompt generation module, and outputs the feature encoding of the text prompts; In the two cross-attention modules, one cross-attention module receives the time series of fused block-level semantic information and the text prompts describing the time series, while the other cross-attention module receives the time series of information fusion in the first stage and the time series of fused sequence-level semantic information. The large language model with frozen parameters receives the time series of information fusion in the second stage and outputs the feature encoding of the fused time series. In the two linear layers, one linear layer receives domain vocabulary input, and the other linear layer receives the output of the large language model. 3.The method of claim 1, wherein, The steps for preprocessing relevant data from multiple social media posts are as follows: Read in the time series dataset of the number of reposts of multiple social media information, and filter the data based on the posting timestamp and the total number of reposts, and filter the data that was not posted at ordinary times or had very few or very many reposts. Cumulative number of retweets from time of publication, recorded every t minutes , up to observation time , construct input time series ; Each selected dataset is divided into a training set, a validation set, and a test set according to a certain ratio for use as training, validation, and result testing, respectively. 4.The method of claim 3, wherein, The semantic enhancement steps of the block-level text prototype retrieval module are as follows: Firstly, the constructed input time series is cut into multiple time segments by using sliding window technique; the segment length is set as p and the sliding step is set as s; for a sequence with length o, a segment set is generated, where the i-th segment , and ; Secondly, segment feature mapping: A Transformer-based encoder is used to map each time segment... Mapping from the time dimension to the feature dimension yields the initial feature representation of the segment. ; Then, prototype retrieval and feature fusion: calculating fragment features. With text prototype library For each prototype vector, calculate the cosine similarity and select the prototype vector with the highest score. This serves as the semantic label for the segment; the semantic label is added to the initial features of the segment to obtain the semantically enhanced segment features. By concatenating the blocks that incorporate semantic information, a time series containing semantic information is obtained, denoted as a semantically enhanced time series feature. : 。 5. The multimedia content popularity prediction method according to claim 4, characterized in that, The text prototype library The construction includes the following steps: Construct a domain-specific seed thesaurus and define a set of seed words that are strongly correlated with time series popularity trends. This includes words describing trends, words describing quantities, and words describing tasks; Lexicon filtering and dimensionality reduction yield the original vocabulary embedding matrix V of the pre-trained large language model. The vector for each word in the vocabulary is then calculated using the following formula. With seed word set The cosine similarity of all word vectors in the dataset. Set similarity threshold =0.75, retain similarity greater than The words constitute the filter vocabulary. ; The filtering vocabulary was analyzed using principal component analysis algorithm. The word vectors in the text are subjected to dimensionality reduction processing, reducing the dimension from Dimensional reduction Set it to 500.

6. The multimedia content popularity prediction method according to claim 1, characterized in that, The semantic enhancement steps performed by the sequence-level prompt generation module are as follows: First, a prompt template is constructed, which includes a dataset description, a data description, and a trend description. The dataset description indicates that the input sequence is a time series of the number of times a post has been forwarded. The data description allows the time series values ​​to be input as additional text. The trend description pre-calculates the overall trend value and growth rate representing the overall change and inserts them into the prompt. The aforementioned prompt template is used to generate a unique prompt text for each input sample; the input parameters of the large language model with the complete prompt text input parameter frozen are used for feature encoding of the prompt text, and the output of the last hidden state of the parameter frozen large language model is extracted as the prompt feature representation. .

7. The multimedia content popularity prediction method according to claim 6, characterized in that, The steps for fusing time series and text representation information using a two-stage cross-attention mechanism are as follows: A first-stage cross-attention mechanism is constructed to enhance the semantically augmented time-series features obtained in the block-level text prototype retrieval module. As a query vector, the prompt features obtained in the sequence-level prompt generation module are used. As key and value vectors, intermediate features are calculated using a multi-head attention mechanism. The calculation formula is as follows: Constructing a second-stage cross-attention mechanism: using the aforementioned intermediate features As a query vector, with the aforementioned semantically enhanced time series features The alignment feature Z is calculated as the key vector and value vector according to the following formula: 。

Images