Influencer matching evidence method and system based on cross-modal semantic decoupling and period-aware attention

Abstract

The present application relates to the technical field of marketing management, and discloses a net red matching evidence method and system based on cross-modal semantic decoupling and period perception attention, which comprises the following steps: unsupervised clustering of visual cross-modal features of the visual assets to obtain brand tone anchor point vectors, and dynamic fusion of content cross-modal features of the public content to obtain a net red content feature sequence; self-attention calculation is performed with the brand tone anchor point vectors as the query and the net red content feature sequence as the key to obtain initial attention weights; and according to the back reference strength of the net red content feature sequence of the target net red in the future period to the net red content feature sequence in the historical period, the present application can solve the problems that the semantic alignment between the brand tone anchor point and the net red content feature is not accurate enough, the interpretability of the matching score is insufficient, and it is difficult to form convincing multi-dimensional matching evidence.

Description

Technical Field

[0001] This invention relates to the field of marketing management technology, and in particular to a method and system for influencer matching evidence based on cross-modal semantic decoupling and time-period-aware attention. Background Technology

[0002] In the field of influencer marketing, precise matching between brands and influencers is a core element in improving marketing effectiveness. Traditional methods often rely on single-modal features, such as text tags or follower counts, for matching, making it difficult to comprehensively capture the cross-modal semantic relationships between brand visual assets and influencer public content. Although some studies have attempted to introduce multimodal feature fusion, they often use simple feature splicing or weighted averaging, failing to effectively decouple redundant and unique information between different modalities. At the same time, due to the inherent differences in information content and representational ability between visual and text modalities, modal bias is prone to occur during the fusion process. The model tends to over-rely on a certain dominant modality, resulting in a distortion of the brand's tone representation.

[0003] Meanwhile, existing methods typically treat influencer content as a static set of features, neglecting the dynamic nature of its evolution over time. Influencer content styles, topic preferences, and visual presentations change significantly over time, and traditional matching models lack the ability to model the temporal dependencies of content features. Furthermore, brand-influencer tone matching should not be based solely on static similarity calculations of historical content; it should also consider the potential alignment between the influencer's future content trends and the brand's tone. However, existing methods generally lack predictive perception mechanisms for future content, failing to incorporate the retrospective correlation between influencer's past and future content into the matching evaluation system.

[0004] The aforementioned deficiencies result in insufficient semantic alignment between brand tone anchors and influencer content characteristics, leading to inadequate interpretability of matching scores and difficulty in forming convincing multi-dimensional matching evidence. Summary of the Invention

[0005] This invention provides a method and system for matching evidence of online celebrities based on cross-modal semantic decoupling and time-period-aware attention, the main purpose of which is to address the problems raised in the background art above.

[0006] To achieve the above objectives, the present invention provides a method for influencer matching evidence based on cross-modal semantic decoupling and time-period-aware attention, comprising: S1: Acquire the target brand's visual assets and the target influencer's public content; S2: Perform unsupervised clustering on the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain a sequence of influencer content features; S3: Using the brand tone anchor vector as the query and the influencer content feature sequence as the key, perform self-attention calculation to obtain the initial attention weight; S4: Construct an echo enhancement time decay factor based on the retracement intensity of the target influencer's influencer content feature sequence in the future time period to the influencer content feature sequence in the historical time period; S5: The initial attention weight is corrected using the echo enhancement time decay factor to obtain the optimized contribution weight, and the tone matching score between the brand and the influencer is calculated based on the optimized contribution weight; S6: Based on the preset tone dimension library, and combined with the optimization contribution weight and the tone matching score, construct a multi-dimensional matching evidence package.

[0007] Preferably, the unsupervised clustering of the visual cross-modal features of the visual assets to obtain brand tone anchor vectors includes: Extract the brand's logo image, main color palette, and keyframes from the advertising video; The identification image, main color palette, and keyframes of the advertising video are input into a visual encoder to extract the corresponding 512-dimensional embedding vector. Density peak clustering is performed on the 512-dimensional embedding vectors to obtain the cluster centers; The brand tone anchor vector is obtained by calculating a weighted average of the cluster centers based on the cluster size of the cluster centers.

[0008] Preferably, the step of performing density peak clustering on the 512-dimensional embedding vectors to obtain cluster centers includes: Calculate the local density of each of the 512-dimensional embedding vectors, and calculate the target distance between each of the 512-dimensional embedding vectors and the nearest 512-dimensional embedding vector with a higher density point; The formula for calculating the local density is:

[0009] in: For the first The local density of the 512-dimensional embedding vectors, The width of the Gaussian kernel. For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, The index of the 512-dimensional embedding vector. The index of the 512-dimensional embedding vector; The formula for calculating the target distance is:

[0010] in: For the first Distance to each target For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, For the first The local density of the 512-dimensional embedding vectors, For the first The local density of the 512-dimensional embedding vector.

[0011] Preferably, the synchronous dynamic fusion of the cross-modal features of the publicly available content to obtain a feature sequence of influencer content includes: The first feature vector of the images and video frames in the publicly available content is extracted according to the preset CLIP visual encoder; When the text length of the image and text post in the disclosed content meets the preset condition, the second feature vector of the image and text post is extracted using the preset ViLT image and text joint encoder; The second feature vector is projected onto the same dimension as the first feature vector, and the projection result is dynamically adjusted according to the text length to obtain the fused feature. The formula for calculating the fusion feature is:

[0012] in: for The fusion characteristics of moments For dynamic adjustment of text length, The first eigenvector, This is the second feature vector. For time indexing, and , Total duration; The obtained fusion features are integrated into a feature sequence of influencer content.

[0013] Preferably, the step of using the brand tone anchor vector as the query and the influencer content feature sequence as the key to perform self-attention calculation to obtain the initial attention weight includes: The formula for calculating the initial attention weights is:

[0014] in: for The initial attention weights at time 1, It is a soft maximum function. This is the brand tone anchor vector. For the feature sequence of the internet celebrity content The fusion characteristics of moments The projection matrix, The projection matrix, For time indexing, and , Total duration It is a real number.

[0015] Preferably, the formula for calculating the echo enhancement time attenuation factor is:

[0016] in: for Echo enhancement time decay factor at time, It is an exponential function. The rate of decline of the basic exponent, Echo intensity function Echo amplification factor For time indexing, and , Total duration; The formula for calculating the echo intensity function is:

[0017] in: Echo intensity function For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. To control the attenuation of the echo with future distance, For the future time period, and , For the historical period, and .

[0018] Preferably, the The calculation formula is:

[0019]

[0020]

[0021]

[0022] in: For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. Here is the formula for calculating cosine similarity; For gating, To from the The echo components decomposed from it For the aforementioned future time period, For the aforementioned historical period, For the The newly added components decomposed from it Forward attention weights.

[0023] Preferably, the step of correcting the initial attention weights using the echo enhancement time decay factor to obtain optimized contribution weights includes:

[0024] in: Echo enhancement time decay factor To optimize contribution weights, These are the initial attention weights.

[0025] Preferably, the step of calculating the tone matching score between the brand and the influencer based on the optimized contribution weight includes:

[0026] in: for The tone matching score at any given moment. for The contribution weight of optimization at any time. The cosine similarity function is used. for The fusion characteristics of moments This is the brand tone anchor vector. For time indexing, and , Total duration.

[0027] A social media influencer matching evidence system based on cross-modal semantic decoupling and time-period-aware attention, the system comprising: Acquisition module: Used to acquire the visual assets of the target brand and the public content of the target influencer; Feature acquisition module: used to perform unsupervised clustering of the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain the influencer content feature sequence; Weight generation module: used to perform self-attention calculation with the brand tone anchor vector as the query and the influencer content feature sequence as the key to obtain the initial attention weight; The anaphora module is used to construct an echo enhancement time decay factor based on the anaphora intensity of the target influencer's content feature sequence in the future time period relative to the influencer's content feature sequence in the historical time period. The score generation module is used to correct the initial attention weight using the echo enhancement time decay factor to obtain an optimized contribution weight, and to calculate the tone matching score between the brand and the influencer based on the optimized contribution weight. Multi-dimensional matching evidence package generation module: used to construct a multi-dimensional matching evidence package based on a preset tone dimension library, combined with the optimization contribution weight and the tone matching score.

[0028] Compared with the prior art, the present invention has the following beneficial effects: This invention mines brand visual assets through unsupervised density peak clustering, adaptively aggregating multimodal visual features into brand tone anchor vectors. This avoids the tone representation bias caused by traditional methods relying on subjective labels or single-modal features. Simultaneously, for influencer-submitted content, it employs a CLIP visual encoder and a ViLT image-text joint encoder for dynamic fusion, adaptively adjusting the fusion weights based on text length. This preserves the global perception capability of the visual modality while also considering the fine-grained semantic information of the image-text joint modality. It achieves effective decoupling and complementarity of cross-modal features, reducing the impact of redundant information between modalities and dominant modality bias on matching accuracy. This makes the alignment of brand tone anchors with influencer content feature sequences in the semantic space more accurate and robust.

[0029] This invention introduces an echo-enhancing time decay factor. By calculating the indexation strength of future content features of influencers on historical content features, it explicitly models the temporal dependency of content evolution into attention weight calculation. This factor can adaptively decay the contribution of outdated historical content while enhancing historical echo information that is highly correlated with future trends. This allows the optimized contribution weight to not only reflect the similarity of current content but also incorporate the continuity and changing trends of content style, thereby dynamically adjusting the matching score and avoiding the inherent lag and one-sidedness of static matching. Furthermore, based on a preset tone dimension library, a multi-dimensional evidence package is constructed by combining optimized weights and matching scores to provide fine-grained and traceable semantic evidence for each matching score, improving the interpretability of the matching results and the credibility of the decision. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a method for matching evidence based on cross-modal semantic decoupling and time-period-aware attention, provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of an influencer matching evidence system based on cross-modal semantic decoupling and time-period-aware attention, provided in an embodiment of the present invention. The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0032] This application provides a method for providing evidence of influencer matching based on cross-modal semantic decoupling and time-period-aware attention. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data and artificial intelligence platforms.

[0033] Reference Figure 1 The diagram shown is a flowchart illustrating a method for providing evidence of influencer matching based on cross-modal semantic decoupling and time-period-aware attention, according to an embodiment of the present invention. In this embodiment, the method for providing evidence of influencer matching based on cross-modal semantic decoupling and time-period-aware attention includes: S1: Acquire the target brand's visual assets and the target influencer's public content; Specifically, the brand's logo, color palette, and three advertising videos with a duration of no more than 60 seconds were collected from official brand channels, such as the official website, official social media accounts, and advertising platforms.

[0034] At the same time, all recently posted images and texts, including original images, text, and short videos, are obtained from the target influencer's public social media platforms, such as Weibo, Douyin, and Instagram, as public content data.

[0035] Furthermore, the visual assets of a target brand refer to the collection of materials used to represent the brand's visual image, specifically including the brand logo, the primary color palette extracted from the brand's official visual materials, and the advertising videos released by the brand.

[0036] The public content of a target influencer refers to the multimodal content that the influencer publicly posts on social media platforms, including text and image posts and short videos.

[0037] The logo image is the brand logo, which is preprocessed to a resolution of 224×224 pixels in subsequent processing.

[0038] The primary color palette is derived from five dominant colors extracted through K-Means clustering from the brand's official visual assets, such as the logo, official website, and brochures. This results in a 128×128 pixel solid color mosaic image in RGB color space.

[0039] Keyframes in an advertising video are representative frames extracted from a brand's advertising video. The video length should not exceed 60 seconds. Use FFmpeg to sample 8 frames evenly, with the sampling interval being the total number of frames in the video divided by 8, ensuring coverage of the beginning, middle, and end.

[0040] Image and text posts are content formats posted by online celebrities that include both original images and accompanying text.

[0041] The short videos are video content posted by online celebrities. In the subsequent processing, FFmpeg is used to extract the frame sequence at a sampling rate of 1 frame / second, with a maximum of 50 frames sampled for each video.

[0042] S2: Perform unsupervised clustering on the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain a sequence of influencer content features; In this embodiment, the unsupervised clustering of the visual cross-modal features of the visual assets to obtain the brand tone anchor vector includes: Extract the brand's logo image, main color palette, and keyframes from the advertising video; The identification image, main color palette, and keyframes of the advertising video are input into a visual encoder to extract the corresponding 512-dimensional embedding vector. Density peak clustering is performed on the 512-dimensional embedding vectors to obtain the cluster centers; The brand tone anchor vector is obtained by calculating a weighted average of the cluster centers based on the cluster size of the cluster centers.

[0043] Specifically, including but not limited to: brand logos, which need to be preprocessed to a resolution of 224×224 pixels; The primary color palette is automatically extracted from the brand's official visual assets, such as the brand logo, official website, and official brochures, using K-Means clustering to extract five dominant colors and generate a 128×128 solid color mosaic, ensuring that the color space is RGB. Keyframes from 3 advertising videos, each no longer than 60 seconds, were used to sample 8 frames evenly using FFmpeg, with a sampling interval of 8 times the total number of video frames, ensuring coverage of the beginning, middle, and end to avoid sampling bias.

[0044] Furthermore, the collected brand logo images, solid color mosaics of the main color palette, and keyframes from the advertising video are input into a visual encoder to extract fixed-dimensional embedding vectors, thus mapping visual content from different sources to the same semantic vector space. Then, density peak clustering is applied to these embedding vectors to automatically identify multiple cluster centers by calculating the local density of each vector and the minimum distance between it and higher density points.

[0045] Finally, using the number of vectors contained in the cluster represented by each cluster center as the weight, a weighted average is performed on each cluster center to obtain a unique comprehensive vector, which is the brand tone anchor vector.

[0046] In this embodiment, the step of performing density peak clustering on the 512-dimensional embedding vector to obtain cluster centers includes: Calculate the local density of each of the 512-dimensional embedding vectors, and calculate the target distance between each of the 512-dimensional embedding vectors and the nearest 512-dimensional embedding vector with a higher density point; The formula for calculating the local density is:

[0047] in: For the first The local density of the 512-dimensional embedding vectors, The width of the Gaussian kernel. For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, The index of the 512-dimensional embedding vector. The index of the 512-dimensional embedding vector; Specifically, the width of the Gaussian kernel is set to 0.5 times the average Euclidean distance of the eigenvectors; The formula for calculating the target distance is:

[0048] in: For the first Distance to each target For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, For the first The local density of the 512-dimensional embedding vectors, For the first The local density of the 512-dimensional embedding vector.

[0049] In this embodiment, the synchronous dynamic fusion of the cross-modal features of the publicly available content to obtain the influencer content feature sequence includes: The first feature vector of the images and video frames in the publicly available content is extracted according to the preset CLIP visual encoder; When the text length of the image and text post in the disclosed content meets the preset condition, the second feature vector of the image and text post is extracted using the preset ViLT image and text joint encoder; The second feature vector is projected onto the same dimension as the first feature vector, and the projection result is dynamically adjusted according to the text length to obtain the fused feature. Specifically, for each point in time, such as daily or weekly content published by online celebrities, the images and video frames are first extracted and input into the CLIP visual encoder to obtain the first feature vector; Meanwhile, for image and text posts, when the text length exceeds a certain threshold, the ViLT image and text joint encoder is enabled to jointly encode the image and text in the image and text post into a second feature vector. This vector has a different dimension than the first feature vector and needs to be aligned to the same space as the first feature vector through linear projection.

[0050] Subsequently, based on the actual text length of the copy, a fusion weight is dynamically calculated, that is, when the text is long, the proportion of image and text vectors is increased, and when the text is short, the proportion of pure visual vectors is increased. The two feature vectors are weighted and summed to obtain the fusion feature at that moment. The fusion features of all moments are arranged in chronological order to form the feature sequence of popular content.

[0051] Specifically, the input data includes: recently published content, such as text and image posts: original image + text; short videos: using FFmpeg to extract frame sequences at a sampling rate of 1 frame / second, with a maximum sampling of 50 frames / video to avoid excessive computational overhead; The formula for calculating the fusion feature is:

[0052] in: for The fusion characteristics of moments For dynamic adjustment of text length, The first eigenvector, This is the second feature vector. For time indexing, and , Total duration; The obtained fusion features are integrated into a feature sequence of influencer content.

[0053] Specifically, pre-training on VQA tasks strengthens the semantic coupling between textual emotion and visual style, enabling the extraction of joint text-image embeddings. (768 dimensions, reduced to 512 dimensions by linear projection to match CLIP space); Dynamic text length adjustment is achieved by dynamically adjusting the length of the content text. This is done by calculating the length using the BERT tokenizer; if the length exceeds 20 characters, then... =0.6, increase the proportion of text-image coupling; otherwise =0.9, prioritizing visual features.

[0054] This strategy avoids introducing noise interference from ViLT in weak text scenarios, such as pure image posts. The ViLT branch is only enabled when the text length is ≥10 characters. The threshold is based on empirical testing to ensure that the signal-to-noise ratio is >1.5.

[0055] The aforementioned length thresholds of 10 characters and 20 characters, as well as the weighting coefficients (α=0.6 / 0.9), are all derived from statistical analysis and empirical testing of large-scale influencer content datasets. The aim is to ensure that the image-text coupling information is fully utilized when the text information is sufficient, and to avoid introducing noise when the text is sparse, thereby ensuring the optimal signal-to-noise ratio of the fused features.

[0056] S3: Using the brand tone anchor vector as the query and the influencer content feature sequence as the key, perform self-attention calculation to obtain the initial attention weight; Specifically, the brand tone anchor vector is used as the query, and the fused features of each moment in the influencer content feature sequence are used as the key. Scaling dot product self-attention calculation is performed. The query vector and each key vector are reduced in dimensionality by two different linear projection matrices, which reduces computational complexity while retaining key semantic information. Then, the dot product of the dimensionality-reduced query vector and each key vector is calculated to obtain the attention score. This score is then divided by a scaling factor to prevent gradient vanishing. Finally, the score is normalized to a probability distribution using the Softmax function, which is the initial attention weight for each time step.

[0057] In this embodiment, the step of using the brand tone anchor vector as the query and the influencer content feature sequence as the key to perform self-attention calculation to obtain the initial attention weight includes: The formula for calculating the initial attention weights is:

[0058] in: for The initial attention weights at time 1, It is a soft maximum function. This is the brand tone anchor vector. For the feature sequence of the internet celebrity content The fusion characteristics of moments The projection matrix, The projection matrix, For time indexing, and , Total duration It is a real number.

[0059] Specifically, Query via linear projection Dimensionality is reduced to 64 dimensions to improve computational efficiency; Key: Feature sequences of each influencer's content projection.

[0060] S4: Construct an echo enhancement time decay factor based on the retracement intensity of the target influencer's influencer content feature sequence in the future time period to the influencer content feature sequence in the historical time period; In this embodiment, the formula for calculating the echo enhancement time attenuation factor is:

[0061] in: for Echo enhancement time decay factor at time, It is an exponential function. The rate of decline of the basic exponent, Echo intensity function Echo amplification factor For time indexing, and , Total duration; Furthermore, the default value for the base exponential decay rate. =0.01, with a value range of 0.005 to 0.03, for example =8、 =1 base decay The decay is only 6.8% over seven time periods, ensuring that the early core tone is not excessively suppressed due to purely time-related factors. This is precisely the prerequisite for the echo-driven mechanism to function. If the fundamental decay is too rapid, such as... =0.05, decaying to 0.7 over 7 time periods, even with echo recovery, it is difficult to form a significant non-monotonic peak; =0.03 decays to 0.81 over 7 time periods, which is suitable for categories with rapidly changing content styles, such as beauty and fashion. Do not exceed One-third of the time, otherwise the basic attenuation dominates everything, the moderating effect of the echo mechanism is submerged, and the FMCG / fashion category takes... =0.02~0.03; Durable consumer goods / brand image category =0.005~0.01.

[0062] Default value of echo amplification factor =0.5, with a value range of 0.3 to 0.8. The value of should be chosen such that the weight of earlier content with an echo can exceed that of more recent content without an echo; by For example, with a value of 0.01, the early content ( =1, =8) The base weight is 0.932, and it needs to rise to >1.0 to reflect that the core tone surpasses recent noise, that is... .when When the typical value is 0.2 to 0.4, A value of 0.5 can ensure a rebound of 10% to 20%, reliably breaking through the threshold.

[0063] When =0.8, if =0.5 (high echo), with a recovery factor of 1.40, which may overemphasize the impact of a single period; When the value is 0.3, the rebound is only 6% to 12%, and the non-monotonicity is not obvious. The content sequence is relatively short. When ≤10, take =0.5~0.6; longer sequences ( >20) take =0.3~0.4.

[0064] Specifically, Using the "echoes of subsequent content on history" as the driving force for the recovery of historical weight, a more accurate weight of works can be calculated.

[0065] The formula for calculating the echo intensity function is:

[0066] in: Echo intensity function For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. To control the attenuation of the echo with future distance, For the future time period, and , For the historical period, and .

[0067] More specifically, for each moment in the historical period, we first iterate through the feature vectors of all its future moments and calculate the anaphoric strength of the future vectors to the historical vectors. This strength is obtained by using a forward attention mechanism to decompose the echo components related to the historical period from the features of the future period and to calculate the similarity between the echo components and the features of the historical period. A gating value is used to filter out non-continuous coincidental similarities to obtain the final anaphoric strength.

[0068] Then, the echo intensity of all future time periods to the same historical time period is exponentially decayed over time and accumulated to obtain the echo intensity of that historical time period. Finally, a base time decay value is calculated for the historical time period and enhanced by the echo intensity to obtain the final echo enhancement time decay factor.

[0069] Specifically, in the evolution of influencer styles, an early core tone may not necessarily disappear. If it is repeatedly echoed and continued in subsequent content, its weight should increase rather than continuously decrease. However, most influencers have a wide range of content coverage, so it is necessary to identify which tones and styles are being echoed. The echo of subsequent content to the past serves as the driving force for the historical weight recovery. A backward attention mechanism is introduced, where future queries refer back to historical keys, and the strength of this referencing is used as the basis for historical weight recovery, distinct from the forward direction of all existing attention mechanisms that refer to current queries to historical keys.

[0070] Furthermore, To control the attenuation of the echo with future distance, the default value is... =0.1, with a value range of 0.05 to 0.2.

[0071] It should be significantly greater than Because echoes are a short-term phenomenon at the semantic level, the reliability of echoes decreases more rapidly over longer distances. For example, consider distances of 2, 4, and 7: When the value is 0.1, the attenuation weights are 0.819, 0.67, and 0.497, respectively. The contribution of the nearest neighbor echo is about 1.6 times that of the far distance echo, which preserves the influence of the far distance core tone while avoiding noise accumulation. =0.05, the long-distance (7) decay only to 0.705, and coincidences such as riding the wave of popularity are difficult to be fully suppressed; =0.2, the value decays to 0.247 over long distances, and the continuation of long-term styles will be excessively suppressed. For influencers with high content update frequency (daily / weekly), the upper limit is 0.15~0.2; for influencers with low update frequency (monthly), the lower limit is 0.05~0.08.

[0072] In this embodiment, the The calculation formula is:

[0073]

[0074]

[0075]

[0076] in: For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. Here is the formula for calculating cosine similarity; For gating, To from the The echo components decomposed from it For the aforementioned future time period, For the aforementioned historical period, For the The newly added components decomposed from it Forward attention weights.

[0077] Specifically, and All belong to This is to emphasize the relationship between history, the present, and the future. It is relative to The history of, and It is relative to The future.

[0078] Specifically, By distinguishing between continuation and coincidental similarity, two periods may appear similar due to riding the same trending topic, but this does not necessarily indicate a continuation of style. Therefore, each period is decomposed into echo components and new components, where the echo components are obtained by weighted aggregation of historical periods through forward attention. Overall similarity × echo component percentage: First, examine how similar the two time periods are overall. Then, determine how much of this similarity is truly derived from historical continuity rather than coincidence. This gating ensures that only genuine historical continuity is recognized. Only the continuation of the trend is included in the echo; coincidental similarities that are merely riding the wave of a trend are filtered out.

[0079] S5: The initial attention weight is corrected using the echo enhancement time decay factor to obtain the optimized contribution weight, and the tone matching score between the brand and the influencer is calculated based on the optimized contribution weight; In this embodiment, the step of correcting the initial attention weights using the echo enhancement time decay factor to obtain optimized contribution weights includes:

[0080] in: Echo enhancement time decay factor To optimize contribution weights, These are the initial attention weights.

[0081] Specifically, the optimization contribution weight for each time step is obtained by multiplying the initial attention weight at each moment by the corresponding echo enhancement time decay factor, and then dividing the whole by the sum of all products.

[0082] This operation allows historical moments that are repeatedly echoed by subsequent content to regain their weight, even if they are from a long time ago; conversely, recent content that lacks echoes will not be overestimated. Then, for each moment, the cosine similarity between the brand tone anchor vector and the fused features of that moment is calculated, and then multiplied by the aforementioned optimization contribution weight to obtain the weighted tone matching score for that moment.

[0083] The score takes into account both the stylistic similarity of the content and its representative contribution to the overall evolution of internet celebrities.

[0084] In this embodiment, calculating the tone matching score between the brand and the influencer based on the optimized contribution weight includes:

[0085] in: for The tone matching score at any given moment. for The contribution weight of optimization at any time. The cosine similarity function is used. for The fusion characteristics of moments This is the brand tone anchor vector. For time indexing, and , Total duration.

[0086] Specifically, suppose a certain internet celebrity is in a certain time period =1 forms the core tone of minimalist black and white photography, and the content feature vector for this period is: In subsequent periods In points 3, 5, and 8, the influencer posted content continuing this style, and these feature vectors for future time periods... , , All contain pointers The echo components. The specific calculation process is as follows: First step: Echo component decomposition: Taking 3 as an example, Decomposed into echo components (from , (Weighted aggregation) and new ingredients (New stylistic elements added to period 3 itself, such as newly experimented filter tones); Forward attention weights softmax( · / )measure right The level of attention, among which =1 indicates the index of the historical period being monitored. =3 represents the index for the future time period to initiate the follow-up.

[0087] Step 2 Gating Filtering: Calculation ,Right now What percentage of the echo components point to ; like and Similarities arise simply from riding the same trending topic, and the echo component points to... The proportion is very low, and the threshold value is close to 0. The value is suppressed.

[0088] Third step: Echo intensity accumulation:

[0089]

[0090] It is important to note that: It is the distance decay term. =3 distance =1 is closer and makes a greater contribution; =8 distance =1 is relatively far away, and its contribution is relatively small (being) (attenuation), eventually >0 indicates that the core tone of period 1 was repeatedly echoed in the subsequent period.

[0091] Step 4: Calculate non-monotonic weights: Let =8 represents the latest time period. =0.01、 =0.5. Without an echo mechanism, the base attenuation weight for time period 1 is 0.932, exhibiting monotonic attenuation. With the introduction of echo, .like =0.3, then =0.932×1.15≈1.072, the weight increased instead of decreased, exceeding the weight of the latest period, forming a non-monotonic local peak. This indicates that although minimalist black and white photography was formed in the early period, it has been repeatedly echoed, and its representativeness of the overall style of internet celebrities is higher than some recent but non-continuous content.

[0092] S6: Based on the preset tone dimension library, and combined with the optimization contribution weight and the tone matching score, construct a multi-dimensional matching evidence package.

[0093] Specifically, the pre-defined tone dimension library is used for calculation. , Similarity to preset tones in the dimension library is used to provide a basis for matching tones with online celebrities.

[0094] In zero-shot scenarios, CLIP's text encoder finds the most effective cues to be natural language descriptive phrases that emphasize visual attributes, mood, and stylistic features rather than abstract labels.

[0095] Furthermore, the prompts must meet the following key specifications: highly descriptive and visually oriented: each prompt contains a core adjective, such as vibrant, warm and comfortable, or high-end and luxurious, which can effectively capture implicit tones, such as color mood, compositional rhythm, and atmosphere transmission, in line with the typical usage of CLIP in aesthetic / style matching tasks.

[0096] The length needs to be controlled at around 8-15 words, close to the caption style of CLIP training data, to avoid excessive length leading to token truncation. It should meet the requirements of being unsupervised, unlabeled, and a pure text prompt library without manually preset category labels. In brand-influencer tone matching, this type of prompt has been widely used in similar multimodal semantic alignment tasks, and can generate discriminative cosine similarity vectors.

[0097] Furthermore, firstly, prepare a pre-defined tone dimension library, which consists of a large number of descriptive natural language phrases, such as vibrant, warm and comfortable, high-end and luxurious, etc. Each phrase focuses on describing visual attributes, emotional atmosphere and style characteristics.

[0098] For each tone dimension phrase, CLIP's text encoder is used to encode it into a text feature vector. Then, the similarity between the brand tone anchor vector and the text vector of each tone dimension is calculated to obtain the brand tone tendency distribution. Simultaneously, the similarity of the influencer's fusion features at each moment is calculated to obtain the influencer's tone tendency distribution at each time point. Subsequently, combined with the calculated optimization contribution weights, the influencer's scores on each tone dimension are weighted and summed to obtain the influencer's comprehensive score on each tone dimension.

[0099] Finally, the brand's score on each dimension, the influencer's comprehensive score on the corresponding dimension, and the similarity or matching degree between the two, along with the time period index and feature fragments that contributed the most, are packaged into a structured multi-dimensional matching evidence package for subsequent decision-making presentation.

[0100] More specifically, the tone dimension library is a collection of natural language descriptive phrases, each describing a visual style or mood, used to provide interpretable matching dimensions.

[0101] The CLIP text encoder is the text branch of the CLIP model, used to encode natural language cues into text feature vectors aligned with visual features.

[0102] The multi-dimensional matching evidence package is the final output structured data, which includes the brand's scores on multiple tone dimensions, the influencer's comprehensive scores on the corresponding dimensions, the matching degree of each dimension, and the highest contribution period and its content identifier as support, which is used to improve the interpretability of the matching results and the credibility of the decision.

[0103] like Figure 2 The diagram shown is a functional block diagram of an influencer matching evidence system based on cross-modal semantic decoupling and time-period awareness attention, provided in an embodiment of the present invention.

[0104] The influencer matching evidence system 100 based on cross-modal semantic decoupling and time-period-aware attention described in this invention can be installed in an electronic device. Depending on the functions implemented, the influencer matching evidence system 100 may include a data acquisition module 101, a feature acquisition module 102, a weight generation module 103, a back-indexing module 104, a score generation module 105, and a multi-dimensional matching evidence package generation module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0105] In this embodiment, the functions of each module / unit are as follows: Acquisition Module 101: Used to acquire the visual assets of the target brand and the public content of the target influencer; Feature acquisition module 102: used to perform unsupervised clustering of the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain the influencer content feature sequence; Weight generation module 103: used to perform self-attention calculation with the brand tone anchor vector as the query and the influencer content feature sequence as the key to obtain the initial attention weight; The retracement module 104 is used to construct an echo enhancement time decay factor based on the retracement intensity of the target influencer's influencer content feature sequence in the future time period to the influencer content feature sequence in the historical time period; Score generation module 105: used to correct the initial attention weight using the echo enhancement time decay factor to obtain an optimized contribution weight, and to calculate the tone matching score between the brand and the influencer based on the optimized contribution weight; Multi-dimensional matching evidence package generation module 106: used to construct a multi-dimensional matching evidence package based on a preset tone dimension library, combined with the optimized contribution weight and the tone matching score.

[0106] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0107] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0109] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0110] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for influencer matching evidence based on cross-modal semantic decoupling and time-period-aware attention, characterized in that, The method includes: S1: Acquire the target brand's visual assets and the target influencer's public content; S2: Perform unsupervised clustering on the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain a sequence of influencer content features; S3: Using the brand tone anchor vector as the query and the influencer content feature sequence as the key, perform self-attention calculation to obtain the initial attention weight; S4: Construct an echo enhancement time decay factor based on the retracement intensity of the target influencer's influencer content feature sequence in the future time period to the influencer content feature sequence in the historical time period; S5: The initial attention weight is corrected using the echo enhancement time decay factor to obtain the optimized contribution weight, and the tone matching score between the brand and the influencer is calculated based on the optimized contribution weight; S6: Based on the preset tone dimension library, and combined with the optimization contribution weight and the tone matching score, construct a multi-dimensional matching evidence package.

2. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 1, characterized in that, The unsupervised clustering of the visual cross-modal features of the visual assets to obtain brand tone anchor vectors includes: Extract the brand's logo image, main color palette, and keyframes from the advertising video; The identification image, main color palette, and keyframes of the advertising video are input into a visual encoder to extract the corresponding 512-dimensional embedding vector. Density peak clustering is performed on the 512-dimensional embedding vectors to obtain the cluster centers; The brand tone anchor vector is obtained by calculating a weighted average of the cluster centers based on the cluster size of the cluster centers.

3. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 2, characterized in that, The step of performing density peak clustering on the 512-dimensional embedding vectors to obtain cluster centers includes: Calculate the local density of each of the 512-dimensional embedding vectors, and calculate the target distance between each of the 512-dimensional embedding vectors and the nearest 512-dimensional embedding vector with a higher density point; The formula for calculating the local density is: in: For the first The local density of the 512-dimensional embedding vectors, The width of the Gaussian kernel. For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, The index of the 512-dimensional embedding vector. The index of the 512-dimensional embedding vector; The formula for calculating the target distance is: in: For the first Distance to each target For the first The 512-dimensional embedding vector, For the first The 512-dimensional embedding vector, For the first The local density of the 512-dimensional embedding vectors, For the first The local density of the 512-dimensional embedding vector.

4. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 3, characterized in that, The synchronous dynamic fusion of the cross-modal features of the publicly available content yields a feature sequence of influencer content, including: The first feature vector of the images and video frames in the publicly available content is extracted according to the preset CLIP visual encoder; When the text length of the image and text post in the disclosed content meets the preset condition, the second feature vector of the image and text post is extracted using the preset ViLT image and text joint encoder; The second feature vector is projected onto the same dimension as the first feature vector, and the projection result is dynamically adjusted according to the text length to obtain the fused feature. The formula for calculating the fusion feature is as follows: in: for The fusion characteristics of moments For dynamic adjustment of text length, The first eigenvector, This is the second feature vector. For time indexing, and , Total duration; The obtained fusion features are integrated into a feature sequence of influencer content.

5. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 4, characterized in that, The process of using the brand tone anchor vector as the query and the influencer content feature sequence as the key to perform self-attention calculation to obtain initial attention weights includes: The formula for calculating the initial attention weights is: in: for The initial attention weights at time 1, It is a soft maximum function. This is the brand tone anchor vector. For the feature sequence of the internet celebrity content The fusion characteristics of moments The projection matrix, The projection matrix, For time indexing, and , Total duration It is a real number.

6. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 5, characterized in that, The formula for calculating the echo enhancement time attenuation factor is as follows: in: for Echo enhancement time decay factor at time, It is an exponential function. The rate of decline of the basic exponent, Echo intensity function Echo amplification factor For time indexing, and , Total duration; The formula for calculating the echo intensity function is: in: Echo intensity function For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. To control the attenuation of the echo with future distance, For the future time period, and , For the historical period, and .

7. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 6, characterized in that, The The calculation formula is: in: For reticular strength, This refers to the sequence of influencer content features within the historical period. This refers to the sequence of influencer content features within the future time period. Here is the formula for calculating cosine similarity; For gating, To from the The echo components decomposed from it For the aforementioned future time period, For the aforementioned historical period, For the The newly added components decomposed from it Forward attention weights.

8. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 6, characterized in that, The step of correcting the initial attention weights using the echo enhancement time decay factor to obtain optimized contribution weights includes: in: Echo enhancement time decay factor To optimize contribution weights, These are the initial attention weights.

9. The influencer matching evidence method based on cross-modal semantic decoupling and time-period-aware attention as described in claim 8, characterized in that, The process of calculating the tone matching score between the brand and the influencer based on the optimized contribution weight includes: in: for The tone matching score at any given moment. for The contribution weight of optimization at any time. The cosine similarity function is used. for The fusion characteristics of moments This is the brand tone anchor vector. For time indexing, and , Total duration.

10. A system for providing evidence of influencer matching based on cross-modal semantic decoupling and time-period-aware attention, used to implement the method for providing evidence of influencer matching based on cross-modal semantic decoupling and time-period-aware attention as described in any one of claims 1-9, characterized in that, The system includes: Acquisition module: Used to acquire the visual assets of the target brand and the public content of the target influencer; Feature acquisition module: used to perform unsupervised clustering of the visual cross-modal features of the visual assets to obtain brand tone anchor vectors, and simultaneously dynamically fuse the content cross-modal features of the public content to obtain the influencer content feature sequence; Weight generation module: used to perform self-attention calculation with the brand tone anchor vector as the query and the influencer content feature sequence as the key to obtain the initial attention weight; The anaphora module is used to construct an echo enhancement time decay factor based on the anaphora intensity of the target influencer's content feature sequence in the future time period relative to the influencer's content feature sequence in the historical time period. The score generation module is used to correct the initial attention weight using the echo enhancement time decay factor to obtain an optimized contribution weight, and to calculate the tone matching score between the brand and the influencer based on the optimized contribution weight. Multi-dimensional matching evidence package generation module: used to construct a multi-dimensional matching evidence package based on a preset tone dimension library, combined with the optimization contribution weight and the tone matching score.

Images