AI intelligent marketing content publishing subject delivery effect prediction method

Abstract

The application discloses an AI intelligent marketing content publishing subject putting effect prediction method, and particularly relates to the technical field of artificial intelligence and marketing data analysis; a target product label feature vector group and a content publishing subject micro-reaction feature group are constructed, a rhythm density matrix is generated based on the content rhythm of the content publishing subject, and the matching degree score of the content publishing subject and the product is obtained through coupling analysis; then the content publishing subject with the lowest style resonance degree is identified in combination with rhythm clustering, and feature inversion analysis is performed to extract key influence factors; finally, a lightweight sub-model is constructed based on the key dimension to locally optimize the original model, and the prediction of new content publishing subjects or small sample content publishing subjects is enhanced, the application can improve the accuracy and robustness of content publishing subject putting matching, and is particularly suitable for marketing strategy optimization in cold start or data sparse scenarios.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and marketing data analysis technology, specifically to a method for predicting the effectiveness of AI-powered marketing content publishing. Background Technology

[0002] With the development of social media and e-commerce platforms, content publisher marketing has become an important way for brands to promote themselves. However, due to the significant differences in audience characteristics, activity times, content styles, and historical performance among different content publishers, traditional advertising strategies based on full historical data or platform recommendation rules often lead to large deviations in results. This is especially true when facing small-sample emerging content publishers or categories with limited reach, as the lack of accurate and effective advertising prediction mechanisms results in wasted marketing resources and unstable ROI.

[0003] In certain niche scenarios (such as medical devices and cosmetic products), due to the highly concentrated focus of users and the long consumption decision-making cycle, existing methods struggle to accurately determine the potential fit between the content style and the target user's psychology when matching content publishers, thus failing to predict the actual conversion effect.

[0004] Furthermore, current ad placement predictions mostly ignore the coupling relationship between "user micro-reaction characteristics" and "content rhythm distribution of content publishers," relying solely on explicit indicators such as likes and shares for evaluation, making it difficult to capture the true value of high-quality but non-trending content publishers. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the effectiveness of AI-powered marketing content publishing, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent marketing content publishing entity ad placement effect prediction method, comprising:

[0007] S100, Obtain the tag feature vector group P of the target marketing product;

[0008] S200. Extract the micro-reaction feature group R of the corresponding content publisher from the historical delivery data of the content publisher;

[0009] S300. Based on the past content generation rhythm of the content publishing entity, establish a rhythm density matrix T, and perform feature coupling analysis with the micro-reaction feature group R to obtain the content publishing entity content-user micro-reaction coupling index group C.

[0010] S400. Input the coupling index group C and the tag feature vector group P of the target product into the trained AI prediction model, and output the matching degree score vector M of the content publishing subject to the target product.

[0011] S500: Select content publishers whose matching score vector M is higher than the preset threshold from the candidate content publisher set, and construct the target product's delivery combination D;

[0012] S600. Compare and cluster the content rhythm density matrix T of each content publisher in the ad mix D to identify the subset W of content publishers with the weakest content style resonance in the ad mix.

[0013] S700. Perform feature importance inversion analysis on the content publishing subject micro-reaction feature group R in subset W and the target product tag feature vector group P to identify the key influence dimension F;

[0014] S800 performs local optimization of the parameters in the AI ​​prediction model based on the key influence dimension F, thereby enhancing the prediction of new content publishers or small sample content publishers.

[0015] Preferably, the tag feature vector group P includes category tags, price sensitivity, content preference style tags, and conversion path feature parameters.

[0016] Preferably, the micro-reaction feature group R includes the average browsing dwell time, the content bounce time series, and implicit user behavior signals related to the repost rate within 3 seconds after liking.

[0017] Preferably, S300 includes:

[0018] S301. Based on multiple pieces of content published by the content publishing entity in history, extract the key publishing parameters of each piece of content in chronological order, and use the fixed window segmentation method to classify and quantify them to generate a time series rhythm vector.

[0019] S302. The rhythm vectors are arranged bidirectionally according to the content dimension and the time dimension, and stacked in the form of a two-dimensional matrix to obtain the rhythm density matrix T; where the rows of the matrix represent time segments, the columns represent interaction frequency and peak density index, and each cell of the matrix is ​​the standardized interaction intensity value within the corresponding time period.

[0020] The rhythm density matrix T and the micro-response feature group R of the content publishing subject are coupled dimension by dimension. A two-dimensional convolutional neural network model with a convolution kernel size of 3×3 is used to extract the high-order correlation features between the two. The high-correlation region is given high weight through the attention weighting mechanism, thereby generating the intermediate coupling tensor.

[0021] The intermediate coupling tensor is fused with mean pooling and max pooling to compress it into a fixed-length vector. After normalizing each dimension, the output content publishing main content - user micro-response coupling index group C is produced.

[0022] Preferably, S400 includes:

[0023] S401. Concatenate and splice the content publishing main body content-user micro-response coupling index group C with the target product tag feature vector group P to construct a joint input vector U, where C and P are respectively processed by independent normalization, and the dimension of the spliced ​​vector is the sum of the original dimensions.

[0024] S402. Embed the joint input vector U into features, perform non-linear mapping through the activation function, and output the hidden layer feature vector H. The dimension of the reduced H vector is smaller than that of U.

[0025] S403. Input the hidden layer feature vector H into the trained AI prediction model. The AI ​​prediction model is a deep neural network built on a multi-task learning framework, which can simultaneously output multiple dimensions of content publishing subject matching scores, including content fit, style similarity and behavioral guidance scores, and finally merge them into a matching score vector M.

[0026] Preferably, S500 includes:

[0027] S501. Based on the category characteristics of the target product and the marketing budget strategy, set the minimum entry threshold θ for the content publishing subject matching score;

[0028] S502. From the set of all content publishing entities that have obtained matching score vectors M, select content publishing entity samples with a comprehensive score not lower than the set threshold θ to form a candidate content publishing entity set D′, where each content publishing entity must satisfy: S_total ≥ θ, where S_total is the weighted total score of the content publishing entity matching score vector M;

[0029] S503. Detect the historical sample size of each content publisher in the candidate content publisher set D′, and remove content publishers whose actual exposure or conversion sample count is lower than the set lower limit S_min.

[0030] S504. Under the premise of meeting the scoring threshold and sample credibility conditions, the content publishers that meet the preset combination size upper limit N_max are selected to construct the target product's delivery combination D.

[0031] Preferably, S600 includes:

[0032] S601. For each content publisher in the distribution combination D, obtain its rhythm density matrix T based on its historical content publishing and user interaction feedback, and form a rhythm matrix set T_set, where the rows of each matrix represent time segments and the columns represent standardized interaction frequency indicators.

[0033] S602. For each pair of content publishing entities in T_set, perform similarity evaluation on their corresponding matrices. After flattening the structure tensor, perform cosine similarity calculation, or use the dynamic time warping algorithm to calculate the alignment distance of the content rhythm sequence, and generate the rhythm distance matrix D_matrix between the content publishing entity pairs.

[0034] S603. Using D_matrix as input, a hierarchical clustering algorithm is used to cluster the content publishing entities. The maximum inter-class distance threshold δ is set to automatically identify content publishing entity groups with similar rhythm and style, and generate a rhythm and style cluster identifier for each content publishing entity.

[0035] S604. Based on the clustering results, count the number of members in each cluster and the variance of the distance between clusters, identify the content publishing entities of the individual clusters or edge nodes with the highest rhythm and style deviation, and define the set formed by them as the subset W of content publishing entities with the weakest content style resonance.

[0036] Preferably, S700 includes:

[0037] S701. Combine the micro-reaction feature group R of each content publisher in the subset W of content publishers with the weakest content style resonance with the tag feature vector group P of the target product to construct the micro-reaction-product tag pair {R, P}, which serves as the input sample for the inversion analysis.

[0038] S702. The input sample is fed into the trained matching score model, and the sensitivity of each feature dimension in the input vector to the final score result is calculated by using the backpropagation mechanism combined with the gradient interpretability algorithm.

[0039] S703. Output the importance weight of each input feature dimension according to the gradient response value or feature contribution, and normalize it to the [0,1] interval to form the joint importance weight vector F_weight of micro-response dimension and label feature dimension.

[0040] S704. Select feature dimensions whose weight values ​​are greater than the preset threshold τ from the importance weight vector F_weight, and form the key influence dimension set F by combining the corresponding feature names or index numbers.

[0041] Preferably, S800 includes:

[0042] S801. Extract the corresponding feature dimensions from the set of key influence dimensions F identified from the original input feature space, construct a lightweight sub-model with a single hidden layer neural network structure, and use it to fit the mapping relationship between the features represented by F and the scoring output.

[0043] S802. Freeze the weights of non-critical parameters in the original trained AI prediction model, and only keep the input layer parameters corresponding to F and the associated intermediate network structure in the updated state.

[0044] S803. Select content publishing subject samples similar to the current target product or platform scenario, construct a micro training set, use the micro training set to perform local backpropagation training on the key feature pathways in the lightweight sub-model, and adopt the adjustment strategy in transfer learning.

[0045] S804. After completing the local parameter tuning, the fine-tuned lightweight sub-model and the AI ​​prediction model are fused together and output. The tuning score and the AI ​​prediction model score are fused together using a weighted integration method to generate the final matching degree prediction result for the new content publishing subject or the small sample content publishing subject, and the prediction is enhanced for low data volume samples.

[0046] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0047] 1. This invention constructs a multi-dimensional feature system that integrates the micro-reaction characteristics of content publishers, content rhythm density, and product tag semantics. By introducing coupling analysis and interpretable AI mechanisms, it can accurately identify the content publisher's suitability for dissemination of specific products, achieving a comprehensive matching evaluation from "behavior-driven" to "style resonance." Compared to traditional placement methods that rely on platform tags or single-dimensional indicators, this invention significantly improves the content synergy and user reach efficiency of placement combinations, especially in handling cold-start content publishers, new product categories, or small sample environments, exhibiting higher predictive accuracy and stability.

[0048] 2. This invention extracts key influencing dimensions and constructs lightweight sub-models, achieving targeted optimization of local feature paths without affecting the main model structure. This effectively solves the problem of weak model generalization ability under low data volume. This mechanism supports dynamic adjustment, rapid deployment, and multi-scenario migration, exhibiting good practicality and scalability, and is suitable for various types of content marketing, content publishing entity promotion, and intelligent delivery systems. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0050] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] For examples, please refer to Figure 1 As shown in this embodiment, the AI-powered intelligent marketing content publishing entity's method for predicting the effectiveness of content delivery includes:

[0053] S100. Obtain the tag feature vector group P of the target marketing product, wherein the tag features include category tags, price sensitivity, content preference style tags, and conversion path feature parameters;

[0054] S200. Extract the micro-reaction feature group R of the corresponding content publisher from the historical delivery data of the content publisher. The micro-reaction features include the average browsing dwell time, the content bounce time series, and the implicit user behavior signals related to the forwarding rate within 3 seconds after liking.

[0055] S300. Based on the past content generation rhythm of the content publishing entity, establish a rhythm density matrix T, and perform feature coupling analysis with the micro-reaction feature group R obtained in step S200 to obtain the content publishing entity content-user micro-reaction coupling index group C.

[0056] S400. Input the coupling index group C and the tag feature vector group P of the target product into the trained AI prediction model, and output the matching degree score vector M of the content publishing subject to the target product.

[0057] S500: Select content publishers whose matching score vector M is higher than the preset threshold from the candidate content publisher set, and construct the target product's delivery combination D;

[0058] S600. Compare and cluster the content rhythm density matrix T of each content publisher in the ad mix D to identify the subset W of content publishers with the weakest content style resonance in the ad mix.

[0059] S700. Perform feature importance inversion analysis on the content publishing subject micro-reaction feature group R in subset W and the target product tag feature vector group P to identify the key influence dimension F;

[0060] S800 performs local optimization of the parameters in the AI ​​prediction model based on the key influence dimension F, thereby enhancing the prediction of new content publishers or small sample content publishers.

[0061] In the AI-powered intelligent marketing content publishing subject delivery effect prediction method proposed in this invention, a structured feature representation system for the target marketing product must first be established to facilitate subsequent content publishing subject matching, modeling analysis, and predictive inference. Therefore, in step S100, a feature tag semantic extraction and quantization encoding method is proposed to construct the tag feature vector group P of the target marketing product. This step is the starting point for the entire method's input and directly determines the directionality and accuracy of subsequent model learning.

[0062] The "tag feature vector group P" is a multi-dimensional real-number vector group, where each dimension represents a normalized product tag parameter, used to characterize the product's content attributes, price attributes, user perception path, and preference style characteristics in a marketing context. P can be defined as... ;in, Let be the feature value of the i-th type of label, and n be the total number of label dimensions. In this invention, n is usually set to be no less than 4, mainly including the following four categories of label features:

[0063] Category Tags (p1~pk): These tags indicate the industry and category level to which the product belongs, such as "Skincare - Serum" or "Food - Functional Beverage". Tag data can come from the structured classification table in the product's backend system, or it can be extracted after semantic parsing of product names and descriptions using Natural Language Processing (NLP) technology. Each category tag is quantified using one-hot encoding; if a product belongs to multiple categories, the corresponding dimension is assigned a value of 1, and the rest are assigned 0.

[0064] Price Sensitivity (p{k+1}): Price sensitivity reflects the degree to which a product's price range influences the behavior of potential consumers. This parameter is constructed through analysis of historical browsing and conversion data from e-commerce platforms, specifically by statistically analyzing the conversion rate trend per unit of exposure across multiple price ranges (e.g., 0-99 yuan, 100-199 yuan, and above 200 yuan). After fitting the relationship between price and conversion behavior to a regression model (e.g., linear regression or logistic regression), a normalized sensitivity score (between 0 and 1) is output as the feature value for this dimension.

[0065] Content Preference Style Tags (p{k+2}~p{k+m}): Content preference style tags refer to the types of content expression consumers prefer to accept, such as "emotional expression," "professional reviews," "unboxing demonstrations," and "narrative immersion." This feature analyzes the correlation between the target user group's past content interaction data (including likes, comments, and shares) and the content structure of the content publisher. It uses TF-IDF combined with topic models (such as LDA) to cluster user preferences, ultimately deriving the correlation coefficient between the product and each content style dimension, and normalizing it to the [0,1] interval to form a multi-dimensional content style tag vector.

[0066] Transformation path feature parameters (p{k+m+1}~p) n This type of tag is used to represent the typical path a user takes from accessing content to completing a purchase, such as path combinations like "content-favorite-add to cart-purchase" or "content-add to cart-purchase". By analyzing historical sales funnel data of the product, the proportion of different paths in the overall conversion is statistically analyzed, and a path weight matrix is ​​constructed. Then, based on the weight distribution of the typical paths of the product, principal components are extracted, and a stable path feature vector is formed after principal component analysis (PCA) compression.

[0067] The specific process for constructing the label feature vector group P is as follows:

[0068] Extract basic information such as name, category, description, price, and historical transaction data from the target product's SKU system;

[0069] The trained NLP model is called to perform word segmentation and keyword extraction on the name and description, and combined with the category tree matching in the background to generate category tags;

[0070] Collect user behavior data of similar products on the platform at different price ranges, fit a price-conversion curve, and output a sensitivity score;

[0071] Call the user content interaction clustering module based on LDA topic model to extract high-frequency preference style tags and associate them with product categories;

[0072] Historical user path data was statistically analyzed, a path frequency matrix was constructed, and the principal components were extracted using the PCA algorithm and then normalized.

[0073] By concatenating the above labels according to the defined dimensional order, a complete label feature vector group P is obtained.

[0074] The "micro-response feature set R" refers to a set of fine-grained response features that reflect a user's behavioral responses after interacting with the main content. R is a multi-dimensional real vector, defined as follows: ;in, Let be the i-th micro-reaction feature value, and m be the total number of dimensions of the micro-reaction features. In this invention, it is recommended to set m ≥ 6, including at least the following key feature dimensions:

[0075] Average browsing dwell time (r1): This represents the average time (in seconds) a user spends browsing the main content of the published content. It is calculated by dividing the total browsing time after the content is exposed by the number of users exposed. The specific formula is: Average browsing dwell time = Total browsing time ÷ Total number of viewers. This indicator reflects the attractiveness and information carrying capacity of the content; the higher the value, the more continuously the content can attract user attention.

[0076] Content Bounce Time Series (r2): Records the distribution of time points when users close, swipe away, or stop playback, used to identify the locations in the content where users most frequently terminate interaction. By dividing the content length into equal-length intervals (e.g., every 5 seconds), the bounce frequency of each interval is counted to form a time series vector, which is then reduced to a 1-dimensional feature using PCA. This metric is used to evaluate the continuity of the content structure's attractiveness to users.

[0077] Share rate within 3 seconds of liking (r3): This metric represents the percentage of users who liked the content and then shared it within 3 seconds. It is precisely calculated using user behavior sequence timestamp data: Share rate within 3 seconds of liking = Number of users who shared within 3 seconds ÷ Total number of users who liked the post. This indicator measures the "infectiousness" of the content and reflects its topicality and impulsive sharing attributes.

[0078] Content swipe pause ratio (r4): This is the probability that a user will "actively pause" when swiping to the main content of a content publisher while browsing the content stream: Swipe pause ratio = Number of users who paused ÷ Total number of users who viewed the content; It represents the attractiveness of the first frame of the content and is a core implicit indicator for short video platforms to measure the click potential of content.

[0079] Mean comment behavior delay (r5): This refers to the average time interval between a user viewing content and their first comment. It is calculated by subtracting the content playback start time from the comment time for each user, and then averaging this across all comment behaviors, in seconds. This metric reflects the relationship between the amount of information in the content and the user's thinking time, indirectly measuring "thought engagement."

[0080] Interactive behavior conversion chain length (r6): Records the path length from a user's initial contact with content to completing all possible interactions (likes, favorites, comments, shares). For example, if a user likes → comments → favorites → shares in sequence, the path length is 4. This dimension is averaged across the user group. This metric is used to measure the content's "interaction-level guidance capability".

[0081] The data for the content publisher's micro-response feature group R mainly comes from the platform-level content interaction database, including but not limited to:

[0082] Content exposure log (including timestamp and user ID);

[0083] User behavior logs (such as clicks, likes, comments, swipes, exits, pauses, shares, etc.);

[0084] Content playback data on the client (including progress bar dragging behavior).

[0085] The data collection period is recommended to be set to the past 30 days to ensure the timeliness and representativeness of behavioral characteristics; for content publishers that have not been active for a long time, the period can be extended to 90 days or a dynamic window can be set based on the total exposure of the content.

[0086] To ensure the standardization of the model input and the comparability of features, the extracted R vector will undergo the following processing:

[0087] For cases where certain dimensions are missing in the data of some content publishers (e.g., insufficient sample size), the industry average or the average value of similar content publishers is used to fill the gaps and prevent the model from becoming unstable.

[0088] All rᵢ features are normalized to the [0,1] interval using Min-Max Scaling to ensure a balance of influence from different dimensions in subsequent model calculations.

[0089] For sequential features (such as time series with breakout points), Savitzky-Golay filtering is first used to remove sharp fluctuations, and then principal component analysis (PCA) is performed to compress them to the principal dimension to reduce redundant interference.

[0090] In this invention, step S300 first involves systematically extracting features from the historical content published by the content publishing entity. Taking N historical content posts (N ≥ 10) as the object, the following key publishing parameters are extracted in chronological order:

[0091] Content publication timestamp;

[0092] Total content duration (seconds);

[0093] The peak time period for comments, likes, and shares (i.e., peak interaction time);

[0094] The distribution rhythm of interactions for each piece of content within 24 hours of its publication (discretely divided in 1-hour increments).

[0095] To capture the overall release and interaction rhythm patterns of content publishers, a fixed window segmentation method is adopted, dividing the time series into continuous, non-overlapping time windows, each with a length of 7 days. Within each time window, the aforementioned content parameters are categorized and aggregated to form a time series rhythm vector group. Each rhythm vector records the average interaction metrics (such as average peak latency, average interaction density change rate, etc.) of all content within the corresponding window.

[0096] Each rhythm vector takes the following form: ,in This represents the j-th rhythm indicator, and k is the total number of rhythm dimensions (it is recommended to set it to no less than 6 dimensions).

[0097] The multiple rhythm vectors obtained in step S301 The data is combined and stacked according to the time window order (rows) and the rhythm indicator dimension (columns) to generate a two-dimensional matrix structure, namely the rhythm density matrix T. The structure of matrix T is as follows:

[0098] Line: Represents a continuous time segment (e.g., 7 days per line);

[0099] Columns: Represent specific rhythm characteristics (such as peak interaction delay, first hour interaction ratio, average content length, etc.);

[0100] Matrix elements: represent the standardized interaction intensity values ​​of the corresponding indicators within the time period, which are uniformly scaled to the [0,1] interval using the max-min normalization method.

[0101] The construction method is as follows:

[0102] For each (Rhythm Index), finding its minimum value over its entire historical data. With the maximum value ; return the original value Normalized to: The final result is in the form of: ; where T is an n-row × k-column real matrix.

[0103] This matrix can not only depict the timing pattern of content release by the content publisher, but also represent the density variation of interactive behavior over time, which helps to identify the potential time period rhythm of the content publisher's "triggering behavioral response".

[0104] The aforementioned rhythm density matrix T is coupled with the content publishing subject micro-response feature group R obtained in step S200. R is a one-dimensional real vector representing the deep-level behavioral feedback characteristics exhibited by users after interacting with the content publishing subject, such as dwell time, forwarding delay, and swipe pause rate.

[0105] To fully explore the nonlinear higher-order coupling relationship between R and T, this invention constructs a coupling model in the following manner:

[0106] Construct a two-dimensional convolutional neural network (2D CNN) for feature extraction from the combined effects of T and R:

[0107] Use T as the input, which is a two-dimensional tensor (of shape n×k).

[0108] R is repeatedly broadcast into an auxiliary matrix R' (also of shape n×k) with the same dimension as T, representing the average micro-reaction level under each time window;

[0109] Combine T and R' into a single multi-channel tensor input to the CNN along the channel dimension;

[0110] The convolution kernel is set to 3×3 with a stride of 1, and the ReLU activation function is used.

[0111] After extraction, the CNN outputs a high-order correlation tensor F, where each cell represents the coupling strength between rhythm features and behavioral features in a local region.

[0112] To highlight the most critical coupling regions in the model, an attention mechanism is introduced:

[0113] For each value in F, calculate its deviation from the global average;

[0114] The Softmax function is used to normalize the deviation values ​​to obtain the weight distribution;

[0115] Applying this weight distribution to F generates a weighted coupling tensor F', which strengthens the rhythm-response region that is highly correlated with the transformation behavior.

[0116] To generate a structured set of indicators that can be used by subsequent prediction models, the weighted coupling tensor F' is further processed:

[0117] Perform mean pooling and max pooling respectively.

[0118] The results from the two methods are concatenated and merged to form a fixed-length vector V;

[0119] This vector represents the overall coupling between the content publisher's content publishing rhythm and user micro-reaction behavior. The number of dimensions can be set according to the model requirements (such as 128 dimensions or 256 dimensions).

[0120] L2 normalization is performed on each dimension of V to make it unit length, so as to prevent the weights of different dimensions from interfering with the model training.

[0121] The output is the content publishing entity content - user micro-response coupling index group C.

[0122] The final output C will be used as one of the input variables for the subsequent delivery matching scoring model to measure the adaptability of the content publisher in terms of content rhythm and style to "stimulate potential behavior".

[0123] To achieve data structure consistency, the two types of data features from different sources, namely the content publishing subject behavior coupling index group C and the target product tag feature vector group P, are first uniformly formatted.

[0124] To avoid model training bias caused by differences in the units of measurement between features, vectors C and P need to be independently normalized. Specifically, the min-max scaling method is used:

[0125] For each feature value x in each dimension, its normalized value is calculated as follows: ; where min and max are the minimum and maximum values ​​of the feature in the training sample set, respectively. After processing, all feature values ​​are mapped to the interval [0,1].

[0126] After normalization, the vectors C′ and P′ are merged into a joint input vector U through vector-level concatenation, with the following structure: ;in This represents a vector concatenation operation. Assuming C′ has dimension c and P′ has dimension p, the concatenated vector U has dimension (c + p).

[0127] This joint vector will serve as the main input to the deep model, containing information about the content-driven capabilities of the content publishing entity and preserving the structured semantic features of the product, thus providing a complete semantic foundation for subsequent model calculations.

[0128] The constructed joint input vector U is a high-dimensional dense feature representation. In order to extract its potential high-order semantic relations and compress redundant dimensions, this step introduces an embedding layer to achieve feature dimensionality reduction.

[0129] The embedding layers are one or more fully connected neural networks. Each layer performs a non-linear mapping through linear transformations and activation functions. Assuming the input is a vector U, the network structure is as follows: H = f(W × U + b); where: U is the input vector; W is the weight matrix; b is the bias term; f is the activation function, and ReLU (Rectified Linear Unit) is recommended, with the formula f(x) = max(0, x), which has good gradient propagation properties; the output vector H is the hidden layer feature vector after dimensionality reduction.

[0130] To improve model efficiency and suppress overfitting, the dimension of H is generally set to half or less of the original dimension U. For example, if U is 256 dimensions, then H can be set to 128 or 64 dimensions.

[0131] The hidden vector H preserves the deep semantic correlation between the behavioral features of the content publisher and the semantic features of the product, which is learned through model training. It serves as the input basis for the subsequent matching score model.

[0132] The matching degree prediction model used in this invention is a deep neural network (Multi-Task Deep Neural Network, MTL-DNN) based on a multi-task learning architecture, which is used to output multi-dimensional content publishing subject matching scores.

[0133] The model takes a hidden layer vector H as input and outputs multiple task heads, each of which is used to predict a specific matching dimension score.

[0134] Task 1: Content Affinity Prediction: Predict the degree of structural, rhythmic, and stylistic match between the content produced by the content publisher and the content required for the target product to be disseminated.

[0135] Task 2: Style Similarity Prediction: Measures the consistency between the overall tone of the content published by the main content provider and the tone of the target brand or the style preferred by users.

[0136] Task 3: Behavioral Influence Prediction: Predict the driving force of content publishing on users' deep behaviors (such as collection, forwarding, conversion), and model it in combination with micro-reaction indicators.

[0137] Each task branch structure is a two-layer fully connected network, and the output is a single scalar (score value) ranging from 0 to 1, representing the matching score of that dimension.

[0138] To obtain the comprehensive matching score vector M, this invention employs a weighted fusion strategy to combine the score vectors of the three sub-tasks. The specific method is as follows: ;in: Score the content relevance; Score the style similarity. Score for behavioral guidance;

[0139] Optionally, different weighting coefficients can be set according to specific use cases. This constitutes the weighted total score: The default weight is set to 1 / 3, but it can be dynamically adjusted based on business configuration.

[0140] The AI ​​prediction model can be trained based on the following process:

[0141] Construct a training set using known content publishing entities, product portfolios, and their actual campaign performance data;

[0142] The loss function adopts a combination of multi-task loss functions: ; each of them Let be the mean squared error loss or cross-entropy loss for this task, where i = 1, 2, 3;

[0143] Gradient descent training is performed using the Adam optimizer;

[0144] Configure an early stopping strategy to prevent overfitting.

[0145] After the model is trained, it can perform scoring inference on any input content publishing subject feature group C and product feature group P, and output a matching degree score vector M.

[0146] To filter out samples with poor quality or insufficient matching from the set of content publishers, this invention first introduces a rating filtering mechanism, setting a matching score threshold θ as the minimum screening threshold. θ is a real number between 0 and 1, used to define whether a content publisher possesses the basic matching ability to enter the candidate set; the default recommended range for θ is 0.65 to 0.85; it can be dynamically set according to the industry characteristics of the target product, the required accuracy of dissemination, and the platform's rating distribution.

[0147] The present invention sets the threshold θ using any of the following methods:

[0148] Method 1: Static settings: Manually entered by campaign operators, suitable for high-value or brand-sensitive products, such as pharmaceuticals and financial products;

[0149] Method 2: Dynamic Quantile Generation: Based on the distribution of the total rating of all content publishers (S_total), the upper quantile is selected as the threshold. For example, the threshold is set to the 80th percentile (P80) of the total rating of all content publishers, which automatically calculates the lowest score of the top 20% of content publishers as θ.

[0150] Method 3: Category Adaptive Adjustment: Set predefined baseline θ values ​​for different product categories, such as: default θ = 0.70 for lifestyle products; default θ = 0.80 for professional and technical products; further offset adjustments can be made based on the marketing budget.

[0151] After setting the threshold, content publishers with a comprehensive matching score S_total higher than or equal to θ are selected from the set of content publishers that have completed the scoring, forming an initial candidate set D′.

[0152] In the candidate set D′, some content publishers may lack sufficient historical content or interaction data, leading to high volatility or error risk in the scoring results. This step introduces a sample credibility filtering mechanism to eliminate content publishers with insufficient data support.

[0153] S_min represents the minimum amount of real user behavior data required for a content publishing entity to enter the candidate set within a fixed period (default 90 days).

[0154] Recommended settings: exposure ≥ 100, or conversion behavior (such as add to cart / share) ≥ 50; S_min can be automatically and dynamically adjusted according to the platform's traffic structure.

[0155] For each content publisher in the candidate set D′: obtain its content exposure records and conversion behavior logs for the past 90 days; if its total behavior sample size is lower than the set S_min, it is judged as "insufficient credibility" and removed from D′ to avoid the impact of scoring distortion on the final combination.

[0156] After screening based on both matching degree and credibility, the optimal content publishing entity is selected from the candidate set to construct the target product's distribution portfolio D. This step aims to maximize the overall dissemination effect of the portfolio and employs a multi-dimensional ranking and optimization strategy.

[0157] The entities publishing candidate content are ranked according to the following three core dimensions:

[0158] Score Priority: Sorted from highest to lowest according to S_total, ensuring that content authors with high relevance are given priority in the combination.

[0159] Content Diversity: The system statistically analyzes the distribution of candidate content publishers on content style tags, prioritizes retaining content publisher samples with different content structures, improves the coverage of combined content, and avoids homogenization.

[0160] Audience Reach Diversity: By combining the user profiles of the content publisher's target audience with the historical user profiles, overlap rate analysis is performed. Priority is given to content publishers whose target audiences do not highly overlap, thereby improving the overall audience breadth of the combination.

[0161] Set the maximum number of people in a combination N_max (e.g., 10, 20, 50 people) based on the campaign budget, expected exposure, or the platform's recommendation system capabilities.

[0162] Based on the above ranking results, the top N_max content publisher samples are extracted to form the final distribution combination D.

[0163] Step S600 aims to extract the rhythm density matrix of historical content from all content publishing entities in the delivery combination D, and construct the basic data set T_set for analysis.

[0164] Each content publisher has its own rhythm density matrix T, which is a two-dimensional matrix with the following structure:

[0165] Rows represent consecutive time segments (e.g., a time window of 7 days).

[0166] The columns represent standardized interaction density metrics, such as average content dwell time, like density, and forwarding response latency.

[0167] Each cell represents the normalized interaction intensity value for the corresponding time period and feature dimension, with a value range from 0 to 1.

[0168] The generation method of matrix T has been clarified in step S300, and its core feature is: the rhythmic expression structure of content behavior over time.

[0169] For each content publishing entity i in combination D, extract its matrix. Ultimately, this forms a set: Where n is the total number of content publishing entities in combination D.

[0170] To measure the similarity of content style and rhythm structure between content publishers, this step compares the rhythm density matrices of any two content publishers in T_set, calculates their content rhythm similarity, and generates the distance matrix D_matrix between the content publisher pairs.

[0171] Since the number of rows and columns may differ for each content publisher T (due to differences in content quantity or frequency), the matrix needs to be structurally aligned first. Methods include:

[0172] Row number normalization: uniformly set the number of time windows (e.g., the most recent 6 periods);

[0173] Column vector unification: ensures that each column is a standardized interactive feature dimension;

[0174] Each T is flattened, that is, expanded into a one-dimensional tensor vector by rows.

[0175] For example, if T is a 6-row × 8-column matrix, it becomes a 48-dimensional vector after flattening.

[0176] This invention provides two similarity calculation methods:

[0177] Method 1: Cosine Similarity: Calculate the cosine of the angle between two flattened vectors. The formula is: Where "•" represents the vector dot product, Let be the vector magnitude.

[0178] Method 2: Dynamic Time Warping (DTW): This algorithm is used to process time series of different lengths, dynamically aligning two sets of rhythm data to obtain the shortest path distance. It is suitable for comparing interactive time series within a rhythm density matrix.

[0179] Based on the actual platform resources and computing needs, one method or a combination of both can be selected to generate the distance values ​​between each pair of content publishing entities.

[0180] The final distance matrix is ​​generated as follows: ;in, This represents the rhythmic distance value between content publisher i and content publisher j.

[0181] After obtaining the rhythm distance matrix between content publishers, the next step is to identify clusters of content publishers with similar styles and find the individuals with the greatest rhythm differences.

[0182] This invention employs a hierarchical agglomerative clustering algorithm, suitable for clustering tasks where the number of clusters is not predetermined. The method flow is as follows:

[0183] Each content publisher is initialized as an independent cluster;

[0184] Based on the distance matrix D_matrix, the most similar cluster pairs are merged step by step;

[0185] Until all content publishing entities are clustered into a tree structure (clustering tree);

[0186] The clustering tree is truncated according to the set maximum inter-class distance threshold δ (it is recommended to set it to 1.5 times the mean distance value in D_matrix) to form several final style clusters.

[0187] δ is a key parameter controlling the "granularity" of clustering; the method for setting it is as follows: Where μ_d is the average distance value of D_matrix, and α is an adjustment factor (α ∈ [1.2, 1.8] is recommended); this setting ensures that the clustering results can both preserve style consistency and effectively identify heterogeneous content publishers. The clustering results output the style cluster ID to which each content publisher belongs, which is used in subsequent identification steps.

[0188] Based on the clustering results, the internal consistency and marginal sample distribution of each style cluster are analyzed to identify the content publishing subject set W with the lowest resonance.

[0189] For each cluster, calculate the following metric: intra-cluster variance: representing the dispersion of rhythmic distances among members within the cluster;

[0190] Number of members within a class: reflects the representativeness of this style in the composition.

[0191] If the variance of a cluster is much higher than the mean, or the number of members is less than the set minimum cluster size (e.g., 2 people), it is considered a "low-resonance style cluster".

[0192] For nodes in the distance matrix that are significantly larger than the average distance of other content publishers (e.g., exceeding the mean plus 1.5 times the standard deviation), they are judged as rhythm abnormal individuals and classified into W.

[0193] The set of content publishers identified above is defined as: W = {content publisher i ∈ D | resonance degree is lower than the set threshold η}. This set will be used as the object to be optimized in subsequent model parameter tuning or combination adjustment.

[0194] In step S600, the subset W of content publishers with the lowest style resonance in the distribution combination has been identified. Although the content publishers in this subset have a certain score in the matching scoring model, their content rhythm deviates significantly from that of other members, belonging to potential "structural incoordination factors". In order to further clarify the internal factors that cause the weak fit between these content publishers and the product, this invention introduces a key influence dimension identification method based on interpretability inversion analysis in step S700.

[0195] For the subset W of content publishers with the weakest identified content style resonance, extract the following:

[0196] Micro-reaction feature group R: represents the implicit behavioral response triggered by the content published by the content provider at the user level;

[0197] Product tag feature vector group P: represents the structural features of the target product in terms of content semantics, category, price, conversion path, etc.

[0198] The R vector corresponding to each content publisher is concatenated with the P vector of the target product to form a joint input vector: ;in This indicates a splicing operation, where U is the feature vector input to the inversion analysis module.

[0199] Iterate through all content publishing entities in subset W and construct a joint dataset: Where n is the number of content publishers in subset W. This sample set will serve as the input source for interpretive analysis.

[0200] The matching score model trained by this invention is invoked (see step S400 for details). This model is a multi-task learning architecture based on deep neural networks. It accepts an input vector U and outputs a matching score vector M or its weighted total score S_total.

[0201] To achieve feature importance inversion analysis, we introduce current mainstream interpretable neural network interpretation algorithms, such as Grad-CAM (Gradient-weighted Class Activation Mapping): by calculating the gradient of the output relative to the activation map of the intermediate convolutional layer, we evaluate the impact of input features on the output.

[0202] Integrated Gradients: Integrate the gradient of the input vector from the baseline to the current value to calculate the importance contribution of the feature;

[0203] SHAP (SHapley Additive exPlanations): Based on the Shapley value principle of game theory, it evaluates the marginal contribution of each feature in multiple combinations.

[0204] The above algorithms can be flexibly selected according to the model structure and platform performance to ensure that the output has sufficient stability and interpretability.

[0205] For each input sample Run the above interpretation algorithm to obtain the sensitivity or gradient contribution values ​​corresponding to each dimension of its input features (i.e., each dimension of R and P). Let these be: ; where k is the total number of dimensions of the joint input features.

[0206] To ensure comparability, the gradient contribution values ​​are normalized. Methods include:

[0207] Mini-Max Normalization: ;

[0208] Absolute value normalization method (used when the gradient contains positive or negative signs): Finally, the standardized feature importance vector is obtained: Each of them .

[0209] For multiple content publishing entity samples in subset W, after calculating their F_weight vectors respectively, perform per-dimensional average aggregation operation: .

[0210] The aggregated F_avg vector reflects the sensitivity of the overall subset of content publishing entities to the scoring of each dimension feature.

[0211] Set the threshold τ for feature importance extraction, which is used to filter out the dimensions that have a significant impact on the scoring. The recommended threshold range is 0.3 to 0.5, and the preferred value is: τ = 0.4.

[0212] Traverse all dimension indices j in the aggregated vector F_avg, and filter out the feature dimensions that meet the following conditions: If , then include the dimension index j in the set F.

[0213] Output the set: This set can be transformed into corresponding feature names, such as "forwarding delay rate", "average comment stay time", "category sensitivity", etc.

[0214] According to the set F of key impact dimensions obtained from the inversion analysis in step S700, in the original input feature space (formed by splicing the micro-reaction feature group R of content publishing entities and the product label feature vector group P), extract the subset of feature dimensions corresponding to F to form a new reduced feature subspace Uf.

[0215] Let the original input vector be U, with dimension d, and the overall feature set be: ; If , then: ; where k < d, and Uf is the key feature input for constructing the sub-model.

[0216] For the key feature subspace Uf, construct a lightweight neural network sub-model, whose structure is:

[0217] Input layer dimension: k;

[0218] Hidden layer structure: single hidden layer, with the number of fully connected neurons being h (recommended h = 2k);

[0219] Activation function: ReLU;

[0220] Output layer dimension: 1 (representing the matching degree prediction score);

[0221] The sub-model aims to specifically learn the non-linear mapping relationship between the key features F and the scoring output, and perform parameter training independently of the original main model.

[0222] To avoid overfitting or performance degradation of the overall model structure due to small sample training, this step freezes the parameters of the trained AI prediction master model, allowing only the partial paths corresponding to the key feature F to be updated.

[0223] Set the parameters in the main model that do not belong to the corresponding input channel of F (i.e., columns in the input weight matrix that are not associated with F) to a frozen state and prevent them from being updated;

[0224] At the same time, freeze the intermediate hidden layer weights and bias parameters associated with these input channels;

[0225] The parameters in the path corresponding to F are kept in a trainable state.

[0226] This strategy ensures that the tuning process is performed only on key feature dimensions and does not affect other input paths, thus improving tuning stability and efficiency.

[0227] Using the neural network graph structure analysis module, the path corresponding to key dimensions is automatically identified from the input layer to the output layer, and an updatable parameter set W_trainable is constructed.

[0228] Since there is insufficient data on new content publishers or small sample content publishers, it is necessary to select a sample of content publishers that are similar to the current target product and platform scenario from historical data to construct a training set.

[0229] The selection criteria for entities publishing similar content include:

[0230] The product categories are the same or similar (based on the cosine similarity of the product tag vector P).

[0231] The distributions of microreaction characteristics are similar (the statistical distribution of R is approximate).

[0232] It has real campaign performance and a sample size greater than the preset threshold (e.g., exposure ≥ 100).

[0233] The above-mentioned similar content publishing entities are used to construct a mini training set D_small, retaining only the key feature dimension U_f and the corresponding rating results as training targets.

[0234] The sub-model is trained using a fine-tuning strategy from transfer learning:

[0235] Loss function: Mean Squared Error (MSE) loss or Huber loss;

[0236] Optimizer: Adam optimizer, with a learning rate set to 0.001 to 0.0001;

[0237] Early stopping mechanism: If the loss on the validation set does not decrease after 5 consecutive training epochs, the training will stop.

[0238] Maximum training steps: Default is no more than 1000 steps.

[0239] This process achieves local adaptation and fine-tuning of the critical path without disrupting the overall structure of the main model.

[0240] After the sub-model training is completed, in order to generate more accurate matching prediction results, the optimized sub-model and the main model need to be weighted and fused to form the final output.

[0241] Let S_main be the original matching score output by the main model; S_fine be the score output by the sub-model (tuned model); α be the fusion coefficient, representing the weight of the sub-model's score (recommended α ∈ [0.2, 0.5]); then the final fusion score is: The fusion results are more targeted, especially suitable for revising the scores of content publishers with low sample size and high uncertainty.

[0242] The fusion coefficient α can be dynamically adjusted based on the sparsity of the samples:

[0243] If the sample size of content publishers is less than 50, let α = 0.5;

[0244] If the sample size is between 50 and 100, let α = 0.3;

[0245] If the sample size is >100, set α = 0.2 or abandon fusion and directly use the results of the master model.

[0246] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the effectiveness of AI-powered intelligent marketing content distribution, characterized by: include: S100. Obtain the tag feature vector group P of the target marketing product, wherein the tag feature vector group P includes category tags, price sensitivity, content preference style tags and conversion path feature parameters; S200. Extract the micro-reaction feature group R of the corresponding content publisher from the historical delivery data of the content publisher. The micro-reaction feature group R includes the average browsing dwell time, the content bounce time series, and the implicit user behavior signal related to the forwarding rate within 3 seconds after liking. S300. Based on the past content generation rhythm of the content publishing entity, establish a rhythm density matrix T, and perform feature coupling analysis with the micro-reaction feature group R to obtain the content publishing entity content-user micro-reaction coupling index group C. S400. Input the coupling index group C and the tag feature vector group P of the target product into the trained AI prediction model, and output the matching degree score vector M of the content publishing subject to the target product. S500: Select content publishers whose matching score vector M is higher than the preset threshold from the candidate content publisher set, and construct the target product's delivery combination D; S600. Compare and cluster the content rhythm density matrix T of each content publisher in the ad mix D to identify the subset W of content publishers with the weakest content style resonance in the ad mix. S700. Perform feature importance inversion analysis on the content publishing subject micro-reaction feature group R and the target product tag feature vector group P in subset W to identify the key influence dimension set F; S800: Based on the key influence dimension set F, the parameters in the AI ​​prediction model are locally fine-tuned to enhance the prediction of new content publishers or small sample content publishers. Specifically, this includes: The feature dimensions corresponding to the identified key influence dimension set F are extracted from the original input feature space to construct a lightweight sub-model with a single hidden layer neural network structure, used to fit the mapping relationship between the features represented by F and the scoring output. The non-key parameter weights in the original trained AI prediction model are frozen, and only the input layer parameters corresponding to F and the associated intermediate network structure are kept in the updated state for local parameter tuning. Samples of content publishing entities similar to the current target product or platform scenario are selected to construct a micro training set. The key feature pathways in the lightweight sub-model are trained locally using the micro training set, and the adjustment strategy in transfer learning is adopted. The AI ​​prediction model with local parameter tuning is fused with the lightweight sub-model, and the scores of the lightweight sub-model and the AI ​​prediction model are fused using a weighted ensemble method to generate the final matching degree prediction result for new content publishing entities or small sample content publishing entities, enhancing the prediction of low data volume samples.

2. The AI-powered intelligent marketing content publishing entity's method for predicting the delivery effect according to claim 1, characterized in that: The S300 includes: S301. Based on multiple pieces of content published by the content publishing entity in history, extract the key publishing parameters of each piece of content in chronological order, and use the fixed window segmentation method to classify and quantify them to generate a time series rhythm vector. S302. The rhythm vectors are arranged bidirectionally according to the content dimension and the time dimension, and stacked in the form of a two-dimensional matrix to obtain the rhythm density matrix T; where the rows of the matrix represent time segments, the columns represent interaction frequency and peak density index, and each cell of the matrix is ​​the standardized interaction intensity value within the corresponding time period. The rhythm density matrix T and the micro-response feature group R of the content publishing subject are coupled dimension by dimension. A two-dimensional convolutional neural network model with a convolution kernel size of 3×3 is used to extract the high-order correlation features between the two. The high-correlation region is given high weight through the attention weighting mechanism, thereby generating the intermediate coupling tensor. The intermediate coupling tensor is fused with mean pooling and max pooling to compress it into a fixed-length vector. After normalizing each dimension, the output content publishing main content - user micro-response coupling index group C is produced.

3. The AI-powered intelligent marketing content publishing entity placement effect prediction method according to claim 2, characterized in that: The S400 includes: S401. Concatenate and splice the content publishing main body content-user micro-response coupling index group C with the target product tag feature vector group P to construct a joint input vector U, where C and P are respectively processed by independent normalization, and the dimension of the spliced ​​vector is the sum of the original dimensions. S402. Embed the joint input vector U into features, perform non-linear mapping through the activation function, and output the hidden layer feature vector H. The dimension of the reduced H vector is smaller than that of U. S403. Input the hidden layer feature vector H into the trained AI prediction model. The AI ​​prediction model is a deep neural network built on a multi-task learning framework, which can simultaneously output multiple dimensions of content publishing subject matching scores, including content fit, style similarity and behavioral guidance scores, and finally merge them into a matching score vector M.

4. The AI-powered intelligent marketing content publishing entity placement effect prediction method according to claim 3, characterized in that: The S500 includes: S501. Based on the category characteristics of the target product and the marketing budget strategy, set the minimum entry threshold θ for the content publishing subject matching score; S502. From the set of all content publishing entities that have obtained matching score vectors M, select content publishing entity samples with a comprehensive score not lower than the set threshold θ to form a candidate content publishing entity set D′, where each content publishing entity must satisfy: S_total≥θ, where S_total is the weighted total score of the content publishing entity matching score vector M; S503. Detect the historical sample size of each content publisher in the candidate content publisher set D′, and remove content publishers whose actual exposure or conversion sample count is lower than the set lower limit S_min. S504. Under the premise of meeting the scoring threshold and sample credibility conditions, the content publishers that meet the preset combination size upper limit N_max are selected to construct the target product's delivery combination D.

5. The AI-powered intelligent marketing content publishing entity placement effect prediction method according to claim 4, characterized in that: The S600 includes: S601. For each content publisher in the distribution combination D, obtain its rhythm density matrix T based on its historical content publishing and user interaction feedback, and form a rhythm matrix set T_set, where the rows of each matrix represent time segments and the columns represent standardized interaction frequency indicators. S602. For each pair of content publishing entities in T_set, perform similarity evaluation on their corresponding matrices. After flattening the structure tensor, perform cosine similarity calculation, or use the dynamic time warping algorithm to calculate the alignment distance of the content rhythm sequence, and generate the rhythm distance matrix D_matrix between the content publishing entity pairs. S603. Using D_matrix as input, a hierarchical clustering algorithm is used to cluster the content publishing entities. The maximum inter-class distance threshold δ is set to automatically identify content publishing entity groups with similar rhythm and style, and generate a rhythm and style cluster identifier for each content publishing entity. S604. Based on the clustering results, count the number of members in each cluster and the variance of the distance between clusters, identify the content publishing entities of the individual clusters or edge nodes with the highest rhythm and style deviation, and define the set formed by them as the subset W of content publishing entities with the weakest content style resonance.

6. The AI-powered intelligent marketing content publishing entity placement effect prediction method according to claim 5, characterized in that: The S700 includes: S701. Combine the micro-reaction feature group R of each content publisher in the subset W of content publishers with the weakest content style resonance with the tag feature vector group P of the target product to construct the micro-reaction-product tag pair {R,P}, which serves as the input sample for the inversion analysis. S702. The input sample is fed into the trained matching score model, and the sensitivity of each feature dimension in the input vector to the final score result is calculated by using the backpropagation mechanism combined with the gradient interpretability algorithm. S703. Output the importance weight of each input feature dimension according to the gradient response value or feature contribution, and normalize it to the [0,1] interval to form the joint importance weight vector F_weight of micro-response dimension and label feature dimension. S704. Select feature dimensions with weight values ​​greater than the preset threshold τ from the importance weight vector F_weight, and form the key influence dimension set F by combining the corresponding feature names or index numbers.

Images