A fact-checking article natural search traffic flow prediction and speech step contribution analysis method

Abstract

The application discloses a kind of fact checking article natural search traffic prediction and speech step contribution analysis method, collects the article data containing sentence sequence, rhetoric label, statistical characteristics and search traffic, constructs prediction target and static control variable and standardizes flow data;Through pre-training model, sentence semantic coding is completed, and rhetoric structure sequence of misplacement input is constructed;Build double-branch rhetoric structure causal transformer, fuse text, rhetoric and control variable characteristics;Sentence-level fusion representation is constructed and rhetoric classification is completed, and the search traffic of predicted article is embedded in combination with representation and real rhetoric label;Two-stage alternating loss optimization is used, and classification, prediction and counterfactual field confusion loss training are combined;Through rhetoric label counterfactual replacement, the marginal contribution of different speech steps is analyzed, and the output traffic prediction value and speech step analysis report are output.The application provides quantitative basis for the organization of fact checking article, evidence presentation sequence and rhetoric strategy adjustment.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, machine learning, causal inference, discourse structure analysis, fact-checking text analysis, and prediction of online dissemination effects, and in particular to a method for predicting natural search traffic and analyzing the contribution of fact-checking articles to online discourse. Background Technology

[0002] With the development of internet search engines, social media, and content platforms, fact-checking articles have become an important textual form for correcting misinformation and reducing the impact of misleading dissemination. Fact-checking articles typically present the statement being verified, the source of evidence, factual reasoning, authoritative citations, and conclusions at considerable length. Their dissemination effectiveness is not only related to external factors such as topic popularity, basic website traffic, external links, and keyword coverage, but may also be influenced by the article's internal discourse organization.

[0003] Existing fact-checking text analysis technologies mostly focus on truth / falsehood identification, opinion extraction, evidence retrieval, or automatic sentence-level phrasing identification, rarely addressing further questions such as "Does rhetorical structure affect subsequent search performance of an article?" and "Which phrasing is more critical for predicting results?" Traditional regression or classification models can use rhetorical tags as ordinary feature inputs, but text semantics, article themes, and basic traffic variables are often highly coupled with rhetorical structure. Direct modeling can easily misinterpret semantic content, topic popularity, or basic platform performance as the role of rhetorical structure itself.

[0004] In the field of causal inference, handling confounding factors and estimating counterfactual outcomes are core issues. Existing time-series causal transformer methods can interactively model treatment variables, state variables, and static features, and estimate effect differences by replacing treatment variables with counterfactual ones. However, fact-checking articles are non-temporally ordered discourse sequences composed of multiple sentences, where consecutive sentences may jointly determine the article's semantics and argumentation structure. Directly applying time-series causal models makes it difficult to simultaneously address global inter-sentence dependencies, suppress rhetorical tag leakage, and perform counterfactual analysis of the phrasing sequence.

[0005] Furthermore, in actual projects, the article performance metrics can be set according to the application objectives. For example, the original solution could use the change in Google Trends popularity before and after the publication date as the variable for the debunking effect; this embodiment uses the article's Uniform Resource Locator (URL) organic search traffic within the observation window as the prediction target, and removes this field from the static control variables to avoid target leakage.

[0006] Therefore, a causal transformer method for fact-checking article discourse sequences is needed to predict continuous propagation performance such as natural search traffic by integrating text semantics, rhetorical structure and article-level static variables, and to provide interpretable step contribution analysis. Summary of the Invention

[0007] Purpose of the invention: This invention proposes a method for predicting natural search traffic and analyzing step contribution of fact-checking articles based on the Rhetorical Structure Causal Transformer (RSCT). This method aims to solve the problems in existing technologies, such as the difficulty in separating confounding effects among sentence-level discourse structure, text semantics, and article-level control variables, the difficulty in performing counterfactual substitution analysis on rhetorical structures, and the difficulty in outputting step contribution explanations for writing optimization.

[0008] Technical Solution: The present invention provides a method for predicting natural search traffic and analyzing the contribution of fact-checking articles, comprising the following steps:

[0009] Step S1: Data Acquisition and Preprocessing: Acquire fact-checking article data, which includes at least the sentence sequence of the article text, sentence-level rhetorical structure tags, article-level statistical features, and the natural search traffic of the article's corresponding Uniform Resource Locator;

[0010] Step S2: Construction of prediction target and control variables: The natural search traffic is used as the prediction target, and the article-level features other than natural search traffic are used as static control variables. The prediction target is then standardized.

[0011] Step S3: Sentence-level semantic encoding: Organize each fact-checking article into a sentence sequence according to the sentence order, and use a pre-trained text embedding model to encode each sentence into a fixed-dimensional sentence semantic vector to generate its text representation;

[0012] Step S4: Construction of Rhetorical Structure Sequence: Convert the rhetorical structure label of each sentence into a category index, form a rhetorical structure sequence according to the order in which the sentences appear in the article, and use staggered rhetorical input on the encoder input side so that the current position receives the rhetorical label embedding of the previous position, and the first position receives the start mark;

[0013] Step S5: Construction of Rhetorical Structure Causal Transformer and Multi-Input Encoding: Construct a rhetorical structure causal transformer including a multi-input encoder, a rhetorical classification head, and a result prediction head. The multi-input encoder receives sentence-level text representations, misplaced rhetorical representations, and article-level static control variables. It adopts a dual-branch structure with parallel text and rhetorical branches, and is composed of multiple stacked multi-input blocks. Each multi-input block performs global multi-head self-attention on the text and rhetorical branches respectively, and then realizes the information interaction between text and rhetoric through bidirectional cross-attention. The article-level static control variables are projected and broadcast to each sentence position to obtain the final encoded text hidden state and misplaced rhetorical hidden state, which are used for subsequent fusion representation construction and natural search traffic prediction encoding results.

[0014] Step S6: Fusion Representation Construction and Rhetorical Classification: After the last multi-input block, the final text hidden state of the current sentence output in step S5 is added to the corresponding misaligned rhetorical hidden state, and the initial sentence-level fusion representation is obtained through linear mapping, activation, random deactivation and layer normalization; a rhetorical classification head is set, and the probability distribution of the current rhetorical label is output according to the sentence-level balanced representation.

[0015] Step S7: Natural search traffic prediction: Set the result prediction head, concatenate the sentence-level fusion representation with the current position real rhetorical tag embedding and input it into the result prediction head, output the sentence-level local prediction value, and then perform masked average pooling and global mapping on the local prediction values ​​of all valid sentences to obtain the article-level natural search traffic prediction value in the standardized space.

[0016] Step S8: Alternating Loss Optimization: Calculate the rhetoric classification loss based on the current rhetoric tag probability distribution output in Step S6, calculate the result prediction loss based on the standardized prediction value output in Step S7, and calculate the counterfactual domain confusion loss with the goal of uniform distribution of rhetoric categories; In the first optimization stage, fix the multi-input encoder, sentence-level representation construction parameters, and result prediction head, and update the rhetoric classification head to minimize the rhetoric classification loss; In the second optimization stage, fix the rhetoric classification head, update the multi-input encoder, sentence-level representation construction parameters, and result prediction head to minimize the sum of the result prediction loss and the weighted counterfactual domain confusion loss; In the second optimization stage, the sentence-level fusion representation reduces the discriminability of the current rhetoric tag while retaining relevant information of natural search traffic prediction, forming a sentence-level balanced representation for subsequent counterfactual analysis and model prediction;

[0017] Step S9: Counterfactual Substitution and Rhetorical Contribution Analysis: Using the model optimized in Step S8, keeping the model parameters, factual text representation, factual misalignment rhetorical representation, attention mask, and static control variables unchanged, only replace the current actual rhetorical tag embeddings at one or more positions with candidate tag embeddings on the result prediction head input side; after each replacement, re-perform the forward computation to obtain a balanced representation consistent with the factual computation under the condition that the encoder input and model parameters remain unchanged, and calculate the difference between factual prediction and counterfactual prediction to measure the marginal contribution of candidate rhetorical tags to the natural search traffic prediction results relative to factual tags;

[0018] Step S10: Output of prediction results and explanation report: Using the model optimized in step S8, obtain standardized prediction values ​​according to the prediction process described in step S7, and denormalize the obtained standardized prediction values ​​to the original scale of the prediction target, and output the natural search traffic prediction value and word contribution analysis report for each article.

[0019] Further, the static control variables in step S2 include one or more of the following: Backlinks, Keywords (natural search keywords), Article Length, Sentiment, Modal (multimodal display features), and Topic (topic variable), expressed as follows:

[0020]

[0021] In the formula, Indicates an article index; Indicates the first The predictive objectives of this article; This represents the observed organic search traffic for the corresponding article; Represents a vector of static control variables; Indicates the number of topic variables, multimodal display features Used to indicate whether an article contains images, videos, charts, or other non-plain text display elements;

[0022] The prediction target is standardized, and the standardized training target is represented as follows:

[0023]

[0024] In the formula, Indicates the first The standardized training objectives for this article; and represent the mean and standard deviation of the predicted targets in the training set, respectively; This represents a positive smoothing constant to prevent the denominator from being zero.

[0025] Furthermore, step S3 specifically includes:

[0026] Each sentence is encoded into a fixed-dimensional sentence vector using a pre-trained text embedding model, and each article is padded or truncated to a preset maximum number of sentences. Attention masks are generated to identify valid sentences and padded positions.

[0027]

[0028]

[0029] In the formula, Indicates the sentence position index; Indicates the first The first article One sentence; This represents the semantic vector of the corresponding sentence; The parameter is Pre-trained text embedding models; Represents the dimension of the sentence vector; Indicates an attention mask; Indicates the first The number of effective sentences in an article.

[0030] Furthermore, step S4 specifically includes:

[0031] The rhetorical structure tags adopt a step-level tagging system: macro steps include summary, introduction of false information, refutation of false information, conclusion, and supplementary reading materials; among them, the introduction of false information is further subdivided into three levels: the spread of false information, the description of the content of false information, and the introduction of relevant information; the refutation of false information is further subdivided into three levels: source verification and content comparison, logical reasoning or factual statement, and authoritative citation; when used for model coding, the summary, conclusion, supplementary reading materials, and the above six levels together constitute nine mutually exclusive sentence-level rhetorical structure categories;

[0032] The initial text hiding state and the initial misaligned rhetorical hiding state are represented as follows:

[0033]

[0034]

[0035]

[0036] In the formula, and They represent the first Current sentence position in the article The initial hidden text state, and the source index paired with the sentence. The initial misplaced rhetoric is hidden; This represents the initial misaligned rhetorical hidden state obtained by using the start marker at the beginning of the sequence; exist The time is from the previous position label structure; and These represent the text projection matrix and the bias, respectively. This represents the rhetorical tag embedding function on the encoder input side; the start tag START indicates the special tag used at the beginning of the sequence.

[0037] Furthermore, in step S5, each multi-input block of the rhetorical structure causal transformer performs the following operations sequentially:

[0038] Multi-head self-attention is performed on both the textual and rhetorical hidden states, and a learnable relative position encoding is introduced to preserve the relative order and proximity relationships between sentences.

[0039]

[0040]

[0041] In the formula, This indicates that multiple block layer numbers can be entered. and These represent the hidden state of the previous layer of text and the hidden state of rhetoric, respectively. and These represent text self-attention output and rhetorical self-attention output, respectively. Indicates an attention mask; Represents relative position encoding; Indicates multi-headed attention;

[0042] Secondly, perform bidirectional cross-attention, so that the text representation query rhetorical context and the rhetorical representation query text context are:

[0043]

[0044]

[0045] In the formula, The text represents the cross-attention output obtained from the rhetorical context of the query; The output represents the cross-attention response obtained from the context of the query text;

[0046] Then, the article-level static control variables are projected onto the model dimension and broadcast to each sentence position, merging with the text and rhetorical representations:

[0047]

[0048]

[0049]

[0050] In the formula, Indicates the first In the article, the first The sentence is processed by the first The hidden text state obtained after processing multiple input blocks; This indicates the source index that is paired with this sentence. The misplaced rhetorical label after the first The rhetorical hidden state obtained after processing multiple input blocks; Indicates the sequence number of the multi-input block; This represents the article-level static control variable that is projected and broadcast to each sentence position; and These represent the static feature projection matrix and the bias, respectively. Indicates a feedforward network; Representation layer normalization.

[0051] Further, step S6 specifically includes:

[0052] Sentence-level fusion representations are constructed according to the following formula:

[0053]

[0054] In the formula, This represents the sentence-level fusion representation obtained by fusing the final text hidden state and the final misaligned rhetorical hidden state; Indicates the total number of multiple input blocks; Indicates the first In the article, the first The final hidden text state obtained after processing the last multi-input block of each sentence; This indicates the source index that is paired with this sentence. The rhetorical hidden state obtained after the misaligned rhetorical tags are processed by the last multi-input block; and These represent the sentence-level fusion representation mapping matrix and the bias, respectively. Represents the activation function of the exponential linear unit; Indicates a random deactivation operation; Representation layer normalization;

[0055] The rhetoric classification head predicts the probability distribution of the current rhetoric tag based on the fusion representation:

[0056]

[0057]

[0058] In the formula, The rhetorical tag prediction probability vector represents the current position; This indicates the category in the vector. The probability of; Indicates the heading of a rhetorical device category; and These represent the weight and bias of the classification head, respectively; This represents the normalized exponential function.

[0059] Furthermore, step S7 specifically includes:

[0060] Sentence-level local prediction values ​​are calculated using the following formula:

[0061]

[0062] In the formula, Indicates the first Article No. Local predicted values ​​for each sentence; This represents a local outcome prediction network; This indicates the current rhetorical tag embedding function used by the result prediction head; The actual rhetorical tag indicates the current location; the semicolon in square brackets indicates vector concatenation;

[0063] The predicted article-level organic search traffic in the normalized space is obtained by masking and averaging the local predicted values ​​of all valid sentences, followed by global mapping:

[0064]

[0065] In the formula, Indicates the first The predicted value of this article in the standardized space; This represents the sentence-level fusion representation sequence of the article; This indicates a sequence of factual and rhetorical tags in the article; This indicates the prediction header, generated by the current label embedding function. Local outcome prediction network Masked average pooling and global result mapping network Consisting in sequence; This represents the positive smoothing constant.

[0066] Furthermore, step S8 specifically includes:

[0067] The rhetorical classification loss, outcome prediction loss, and counterfactual domain confusion loss are respectively expressed as follows:

[0068]

[0069]

[0070]

[0071] The alternating optimization objective is expressed as:

[0072]

[0073]

[0074] In the formula, Indicates loss in rhetorical classification; Indicates the predicted loss; This indicates confusion and loss in the counterfactual domain; Indicates the rhetorical category header parameter; The parameters represent the encoding part, including multi-input encoder parameters and sentence-level representation construction parameters; This indicates the result prediction header parameters, including the parameters of the current rhetorical tag embedding function; Indicates the weight of the confusion loss; This represents the total number of rhetorical categories, with the internal category index ranging from 0 to... .

[0075] Further, step S9 specifically includes:

[0076] Counterfactual predictions are calculated using the following formula:

[0077]

[0078] In the formula, This represents the fact balance representation sequence calculated while keeping model parameters, fact text, fact misalignment rhetoric sequence, attention mask, and static control variables constant. This indicates a candidate counterfactual rhetorical sequence used only on the result prediction header input side; if no replacement is specified, the position remains the same as the factual rhetorical sequence. Indicates the result prediction header; This represents the standardized counterfactual prediction value obtained solely from the effect of candidate rhetorical labels, assuming that the factual balance representation remains unchanged.

[0079] The single-sample counterfactual effect is calculated using the following formula:

[0080]

[0081] In the formula, Indicates the first The article's factual rhetorical sequence; Indicates the first The change in standardized predictions when an article is replaced by a factual rhetorical sequence with a candidate counterfactual rhetorical sequence; its absolute value represents the marginal impact of the candidate replacement.

[0082] Calculate the average embedding vector of all rhetorical categories in the result prediction head:

[0083]

[0084] In the formula, This represents the average vector of all rhetorical category embeddings in the prediction header of the result; This represents the internal rhetoric category index, with values ​​ranging from 0 to... ; This indicates the rhetorical tag embedding function used in the result prediction head;

[0085] The marginal contribution of rhetorical structure categories is calculated using label-by-label masking according to the following formula:

[0086]

[0087] In the formula, Indicates rhetorical category The average absolute conditional marginal contribution; This represents a collection of sample articles containing this rhetorical category; This indicates that the category in the prediction header will be used to predict the result. The embedding at the corresponding position is replaced with the standardized prediction value obtained after average embedding; This represents the standardized predicted value corresponding to the factual rhetorical sequence.

[0088] Further, the anti-normalization described in step S10 is represented as:

[0089]

[0090] In the formula, This represents the predicted natural search traffic value restored to the original scale of the predicted target. This represents the standardized predicted value.

[0091] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: This invention forms a complete end-to-end computational process from fact-checking articles to natural search traffic prediction values; it reduces the confounding between textual semantics and rhetorical structure, improving the interpretability of counterfactual analysis; it outputs overall rhetorical structure ablation results, label-by-label masking results, and sample-level prediction perturbations, providing a quantitative basis for evaluating the rhetorical organization of fact-checking articles; compared with schemes that only use textual semantics or only treat rhetorical tags as ordinary features, this invention models rhetorical structure as an interventionist variable and provides a unified process for predicting search traffic and interpreting discourse structure of fact-checking articles through adversarial confounding suppression and counterfactual substitution mechanisms with fixed representations, providing a quantitative basis for adjusting the chapter organization, evidence presentation order, and rhetorical strategies of fact-checking articles. Attached Figure Description

[0092] Figure 1 This is a schematic diagram of the Rhetorical Structure Causal Transformer (RSCT) model architecture. Detailed Implementation

[0093] The present invention will now be described in further detail with reference to the accompanying drawings.

[0094] This invention provides a method for predicting organic search traffic and analyzing the contribution of rhetoric to fact-checking articles. The method includes data acquisition and preprocessing, construction of prediction targets and control variables, sentence-level semantic encoding, construction of rhetorical structure sequences, construction of a rhetorical structure causal transformer and multi-input encoding, construction of fusion representations and rhetorical classification, organic search traffic prediction, alternation loss optimization, counterfactual substitution and rhetoric contribution analysis, and output of prediction results and an explanation report. The specific implementation process is as follows:

[0095] Suppose the dataset contains several fact-checking articles, each article consisting of a semantic sequence of sentences, a rhetorical structure sequence, article-level static control variables, and an organic search traffic target:

[0096]

[0097]

[0098] In the formula, Represents a dataset; Indicates the number of articles; Indicates the first The semantic sequence of sentences in the article; This indicates the corresponding rhetorical structure sequence; This represents an article-level static control variable; Indicates the goal of organic search traffic; Indicates the number of valid sentences.

[0099] Step S1: Obtain the fact-checking articles and their structured data.

[0100] Article data is obtained from fact-checking websites, news organizations' fact-checking sections, third-party fact-checking platforms, or databases. Each article includes at least an article number, main text sentences, sentence order within the article, sentence-level rhetorical structure tags, a Uniform Resource Locator (URL), publication date, and article-level statistical features. Sentence-level rhetorical structure tags can adopt a pre-defined cadence / scale system, such as abstract, introduction of false information, dissemination of false information, related information introduction, rebuttal of false information, factual statements or logical reasoning, authoritative citations, conclusion, supplementary reading materials, etc., and can also be expanded based on the specific corpus.

[0101] Step S2: Construct the prediction target and control variables.

[0102] In this current embodiment, the organic search traffic field in the control feature file (control_features.csv) is used as the prediction target. To avoid target leakage, this field is no longer used as a static feature input. Static features include at least one or more of the following: backlinks, keywords, article length, sentiment, modal presentation features, and topic variables; wherein, the modal presentation features are used to characterize whether the article contains images, videos, charts, or other non-plain text presentation elements.

[0103] The prediction target and static control variables are expressed as follows:

[0104]

[0105]

[0106] In the formula, Indicates an article index; Indicates the first The predictive objectives of this article; This represents the observed organic search traffic for the corresponding article; Represents a vector of static control variables; Indicates the number of subject variables.

[0107] The prediction target is standardized before training to reduce the impact of different dimensions and extreme values ​​on the training process; the standardized training target is expressed as:

[0108]

[0109] In the formula, Indicates the first The standardized training objectives for this article; and represent the mean and standard deviation of the predicted targets in the training set, respectively; This represents a positive smoothing constant to prevent the denominator from being zero.

[0110] Step S3: Construct a sentence-level semantic vector sequence.

[0111] Each fact-checking article is organized into a sentence sequence in sentence order and encoded using a pre-trained text embedding model. In one implementation, a SentenceTransformer-like model is used to generate 384-dimensional sentence vectors, which are then used to pad or truncate the articles; the system also generates an attention mask to identify valid sentences and padding locations.

[0112] The sentence semantic vector and the corresponding attention mask are represented as follows:

[0113]

[0114]

[0115] In the formula, Indicates the sentence position index; Indicates the first The first article One sentence; This represents the semantic vector of the corresponding sentence; The parameter is Pre-trained text embedding models; Represents the dimension of the sentence vector; Indicates an attention mask; Indicates the first The number of effective sentences in an article.

[0116] Step S4: Construct the rhetorical structure sequence A.

[0117] In this embodiment, the original rhetorical tags 1 to 9 of each sentence are converted into internal category indices 0 to 8, and a rhetorical structure sequence is formed according to the sentence order. A staggered rhetorical input is used on the encoder input side, so that the current position receives the rhetorical tag embedding from the previous position during input, instead of directly receiving the current rhetorical tag embedding from the same position; the first position receives the start marker.

[0118] The rhetorical structure of fact-checking articles adopts a move-step hierarchical labeling system: Move 1 is the summary; Move 2 introduces false information, and its steps include the spread of false information, a description of the false information content, and related information introduction; Move 3 refutes false information, and its steps include source verification and content comparison, logical reasoning or factual statements, and authoritative citations; Move 4 draws a conclusion; and Move 5 provides supplementary reading material. When used for model coding, the summary, conclusion, supplementary reading material, and the above six steps together constitute nine mutually exclusive sentence-level rhetorical categories.

[0119] The aforementioned tagging system can be obtained through manual annotation or generated by an automatic speech step recognition model. This hierarchical tagging system enables the discourse structure of an article to be encoded into a sequence of nine sentence-level rhetorical tags; when expanding the categories in other corpora, the total number of rhetorical categories and the embedded vocabulary are adjusted accordingly.

[0120] Convert the original rhetorical tags 1 to 9 into internal category indices from 0 to 8, and construct the misaligned rhetorical input according to the following formula:

[0121]

[0122]

[0123]

[0124] In the formula, and They represent the first Current sentence position in the article The initial hidden text state, and the source index paired with the sentence. The initial misplaced rhetoric is hidden; This represents the initial misaligned rhetorical hidden state obtained by using the start marker at the beginning of the sequence; exist The time is from the previous position label structure; and These represent the text projection matrix and the bias, respectively. This represents the rhetorical tag embedding function on the encoder input side; the start tag START indicates the special tag used at the beginning of the sequence.

[0125] Step S5: Construct as follows Figure 1 The rhetorical structure shown is a causal transformer, and multi-input encoding is performed.

[0126] like Figure 1As shown, the overall architecture of the modified Rhetorical Structure Causal Transformer (RSCT) includes a multi-input encoder, a rhetorical classification head, and a result prediction head. The multi-input encoder is constructed and run, receiving the initial text hiding state, the initial misaligned rhetorical hiding state, the attention mask, the relative position encoding, and the article-level static control variables, and jointly encoding the textual information, rhetorical information, and article-level control information.

[0127] The multi-input encoder employs a text-rhetoric dual-branch structure that allows global inter-sentence interaction and is composed of a preset number of multi-input blocks stacked together. Each multi-input block first performs global multi-head self-attention on the text branch and the rhetoric branch respectively, then performs bidirectional cross-attention from text to rhetoric and from rhetoric to text, and finally broadcasts and merges the projected article-level static control variables to each sentence position.

[0128] The above modifications enable any two valid sentences to establish a global dependency; at the same time, they preserve learnable relative positional encodings to represent the order and proximity relationships between sentences.

[0129] Multi-head self-attention is performed on both the textual and rhetorical hidden states, and a learnable relative position encoding is introduced to preserve the relative order and proximity relationships between sentences.

[0130]

[0131]

[0132] In the formula, This indicates that multiple block layer numbers can be entered. and These represent the hidden state of the previous layer of text and the hidden state of rhetoric, respectively. and These represent text self-attention output and rhetorical self-attention output, respectively. Indicates an attention mask; Represents relative position encoding; This refers to multi-head attention.

[0133] Secondly, perform bidirectional cross-attention, so that the text representation query rhetorical context and the rhetorical representation query text context are:

[0134]

[0135]

[0136] In the formula, The text represents the cross-attention output obtained from the rhetorical context of the query; This indicates the cross-attention output obtained from the context of the query text.

[0137] Then, the article-level static control variables are projected onto the model dimension and broadcast to each sentence position, merging with the text and rhetorical representations:

[0138]

[0139]

[0140]

[0141] In the formula, Indicates the first In the article, the first The sentence is processed by the first The hidden text state obtained after processing multiple input blocks; This indicates the source index that is paired with this sentence. The misplaced rhetorical label after the first The rhetorical hidden state obtained after processing multiple input blocks; Indicates the sequence number of the multi-input block. This represents the article-level static control variable that is projected and broadcast to each sentence position; and These represent the static feature projection matrix and the bias, respectively. This refers to a feed-forward network. Layer normalization.

[0142] Step S6: Construct sentence-level fusion representation and rhetorical classification head.

[0143] Sentence-level fusion representation refers to a sentence-level representation obtained by fusing textual hidden states, misplaced rhetorical hidden states, and article-level static control information. After the last multi-input block, the system adds the final textual hidden state of the current sentence to the corresponding final misplaced rhetorical hidden state, and then performs linear mapping, exponential linear unit (ELU) activation, random dropout, and layer normalization to obtain a balanced sentence-level representation. The system sets a rhetorical classification head and outputs the probability distribution of the current rhetorical label based on the sentence-level fusion representation.

[0144] Sentence-level fusion representations are calculated according to the following formula:

[0145]

[0146] In the formula, This represents the sentence-level fusion representation obtained by fusing the final text hidden state and the final misaligned rhetorical hidden state; Indicates the total number of multiple input blocks; Indicates the first In the article, the first The final hidden text state obtained after processing the last multi-input block of each sentence; This indicates the source index that is paired with this sentence. The rhetorical hidden state obtained after the misaligned rhetorical tags are processed by the last multi-input block; and These represent the sentence-level fusion representation mapping matrix and the bias, respectively. Represents the activation function of the exponential linear unit; Indicates a random deactivation operation; Representation layer normalization.

[0147] The rhetoric classification head predicts the probability distribution of the current rhetoric tag based on the fusion representation:

[0148]

[0149]

[0150] In the formula, The rhetorical tag prediction probability vector represents the current position; This indicates the category in the vector. The probability of; Indicates the heading of a rhetorical device category; and These represent the weight and bias of the classification head, respectively; This represents the normalized exponential function.

[0151] Step S7: Construct the outcome prediction head and predict organic search traffic.

[0152] The prediction head explicitly receives the concatenated vector of sentence-level fusion representation and the embedding of the real rhetorical tag at the current position, and outputs the local prediction value. Then, it performs masked average pooling and global mapping on the local prediction values ​​of all valid sentences to obtain the article-level natural search traffic prediction value in the normalized space.

[0153] Sentence-level local prediction values ​​are calculated using the following formula:

[0154]

[0155] In the formula, Indicates the first Article No. Local predicted values ​​for each sentence; This represents a local outcome prediction network; This indicates the current rhetorical tag embedding function used by the result prediction head; The rhetorical label indicates the current location; the semicolon in square brackets indicates vector concatenation.

[0156] The predicted article-level organic search traffic in the normalized space is obtained by masking and averaging the local predicted values ​​of all valid sentences, followed by global mapping:

[0157]

[0158] In the formula, Indicates the first The predicted value of this article in the standardized space; This represents the sentence-level fusion representation sequence of the article; This indicates a sequence of factual and rhetorical tags in the article; This indicates the prediction header, generated by the current label embedding function. Local outcome prediction network Masked average pooling and global result mapping network Consisting in sequence; This represents the positive smoothing constant.

[0159] Step S8: Perform alternating loss optimization based on the rhetoric classification head and the result prediction head.

[0160] The rhetoric classification loss is calculated based on the current rhetoric label probability distribution output in step S6, the prediction loss is calculated based on the standardized prediction values ​​output in step S7, and the counterfactual domain confusion loss is calculated with the goal of uniform distribution of rhetoric categories; the three losses are expressed as follows:

[0161]

[0162]

[0163]

[0164] The alternating optimization objective is expressed as:

[0165]

[0166] In the formula, Indicates loss in rhetorical classification; Indicates the predicted loss; This indicates confusion and loss in the counterfactual domain; Indicates the rhetorical category header parameter; The parameters represent the encoding part, including multi-input encoder parameters and sentence-level representation construction parameters; This indicates the result prediction header parameters, including the parameters of the current rhetorical tag embedding function; Indicates the weight of the confusion loss; This represents the total number of rhetorical categories, with the internal category index ranging from 0 to... .

[0167] During training, the first optimization phase fixes the multi-input encoder, sentence-level representation construction parameters, and result prediction head, updating only the rhetoric classification head to minimize the rhetoric classification loss. The second optimization phase fixes the rhetoric classification head and updates the multi-input encoder, sentence-level representation construction parameters, and result prediction head to minimize the sum of the result prediction loss and the weighted counterfactual domain confusion loss. In the second optimization phase, the counterfactual domain confusion loss is reduced by fixing the rhetoric classification head and updating the multi-input encoder and sentence-level representation construction parameters. This allows the sentence-level fusion representation to retain information relevant to natural search traffic prediction while reducing information that directly identifies the current rhetoric tag, forming a balanced sentence-level representation. After completing the alternating optimization, in a specific embodiment, exponential moving average (EMA) weights are used for model evaluation.

[0168] Step S9: Perform counterfactual substitution and verbal contribution analysis.

[0169] After the alternating optimization in step S8, keeping the model parameters, factual text representation, factual misalignment rhetorical representation, attention mask, and static control variables unchanged, the current rhetorical label embeddings at one or more positions are replaced with candidate label embeddings only on the input side of the result prediction head. A complete forward computation is re-executed after each replacement to obtain the counterfactual prediction value. Since the replacement operation does not change the input and model parameters of the multi-input encoder, the balanced representation obtained from each forward computation remains consistent with the factual computation under deterministic evaluation conditions. By calculating the difference or absolute difference between the factual prediction and the counterfactual prediction, the conditional marginal contribution of random replacement of all current position label embeddings, mean masking of specified labels, or replacement of specified sentence candidate labels to the prediction result can be measured.

[0170] With the model parameters and encoder inputs unchanged, and only the current rhetorical tag embedding used in the result prediction head replaced, the counterfactual prediction value is calculated according to the following formula:

[0171]

[0172] In the formula, This represents the fact balance representation sequence calculated while keeping model parameters, fact text, fact misalignment rhetoric sequence, attention mask, and static control variables constant. This indicates a candidate counterfactual rhetorical sequence used only on the result prediction header input side; if no replacement is specified, the position remains the same as the factual rhetorical sequence. This indicates the result prediction header defined in step S7; This represents the standardized counterfactual prediction value obtained solely from the action of candidate rhetorical labels, assuming the factual balance representation remains unchanged.

[0173] The single-sample effect between the factual predicted value and the counterfactual predicted value is:

[0174]

[0175] In the formula, Indicates the first The article's factual rhetorical sequence; Indicates the first When an article is replaced by a factual rhetorical sequence with a candidate counterfactual rhetorical sequence, the standardized predicted value shows a signed change; its absolute value represents the conditional marginal impact of the candidate replacement.

[0176] To construct a unified label masking benchmark, we first calculate the average embedding vector of all rhetorical categories in the result prediction head:

[0177]

[0178] In the formula, This represents the average vector of all rhetorical category embeddings in the prediction header of the result; This represents the internal rhetoric category index, with values ​​ranging from 0 to... ; This indicates the rhetorical tag embedding function used in the result prediction head.

[0179] After replacing the specified rhetorical category with the average embedding, its per-label masking contribution is:

[0180]

[0181] In the formula, Indicates rhetorical category The average absolute conditional marginal contribution; This represents a collection of sample articles containing this rhetorical category; This indicates that the category in the prediction header will be used to predict the result. The embedding at the corresponding position is replaced with the standardized prediction value obtained after average embedding; This represents the standardized predicted value corresponding to the factual rhetorical sequence.

[0182] In the overall rhetorical embedding ablation, by keeping the model parameters and encoder input unchanged and fixing the fact balance representation, all current rhetorical label embeddings on the outcome prediction head input side are replaced with random Gaussian vectors, and the complete forward computation is re-performed. The sample-by-sample average absolute prediction perturbation in the normalized space is approximately 0.0116. This result is used to quantify the overall conditional marginal effect of the rhetorical embeddings on the outcome prediction head input side relative to the fact rhetorical embeddings, and to provide a computable reference baseline for subsequent discourse structure adjustments.

[0183] Step S10: Destandardize the predicted values ​​and output the prediction results and a speech contribution analysis report.

[0184] Using the model optimized in step S8, standardized prediction values ​​are obtained according to the prediction process described in step S7, and the obtained standardized prediction values ​​are restored to the original scale of the prediction target:

[0185]

[0186] In the formula, This represents the predicted natural search traffic value restored to the original scale of the predicted target. This represents the standardized predicted value.

[0187] Output the predicted natural search traffic values ​​restored to the original scale of the prediction target and the sample-level verbal contribution analysis report for each article. At the test set or specified article set level, summarize the prediction error index, overall rhetoric ablation comparison, label-by-label masking results, sample-by-sample label perturbation matrix, and natural language analysis report. The verbal contribution analysis report should include at least one or more of the following: overall rhetoric ablation comparison, label-by-label masking summary, sample-by-sample label perturbation matrix, and natural language analysis report.

[0188] In one specific embodiment, the text embedding dimension is 384, the static feature dimension is 21, the number of rhetorical categories is 9, and the model hidden dimension is 128. The model uses multiple multi-input blocks to fuse sentence semantics, misaligned rhetorical structures, and static features, and employs exponential moving average (EMA) weights for evaluation. During the testing phase, natural search traffic prediction is first performed using the true rhetorical structure, and the model output is denormalized to the scale of the natural search traffic field in the data file according to step S10 before calculating the error metric. The test sample size is 380, the Pearson correlation coefficient is 0.2535, the mean absolute error (MAE) is 0.1178, and the root mean square error (RMSE) is 0.3161. Since the natural search traffic field in the current data file itself uses standardized values, the mean absolute percentage error (MAPE) is only used as an auxiliary indicator and is interpreted in conjunction with the symmetric mean absolute percentage error (sMAPE). To further solidify the factual balance representation, we maintain the model parameters, factual text representation, factual misalignment rhetorical representation, attention mask, and static control variables. We then replace all current rhetorical label embeddings on the result prediction head input side with random Gaussian vectors and re-perform the complete forward computation to obtain the overall rhetorical embedding ablation result. In this embodiment, the sample-by-sample average absolute prediction perturbation in the normalized space is approximately 0.0116. This result quantifies the overall conditional marginal impact of the rhetorical embeddings on the result prediction head input side relative to the factual rhetorical embeddings, and provides a calculable baseline for subsequent discourse structure adjustments. When masking each rhetorical label one by one, we also keep the model parameters and encoder input unchanged, replacing all occurrences of that label in the result prediction head with the mean vector of the rhetorical embedding vocabulary. We observe the changes in the normalized prediction values ​​and convert the scale transformation to the original scale of the prediction target. This result provides the differences in conditional marginal contributions and sensitivity ranking for different steps or levels of discourse, which can be used to identify the rhetorical structure positions and candidate step types that require priority adjustment.

[0189] In practical applications, fact-checking organizations or researchers can input a fact-checking article to be evaluated, its automatically identified or manually annotated rhetorical structure sequence, and available article-level control variables to output predicted organic search traffic and explanations of verbal contributions. Organic search traffic can serve as an example indicator of the dissemination or reach of debunking articles; it can also be replaced with target variables such as debunking effectiveness index, reader error correction effectiveness, click-through rate, dwell time, or number of shares.

[0190] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for predicting natural search traffic and analyzing the contribution of fact-checking articles, characterized in that, Includes the following steps: Step S1: Data Acquisition and Preprocessing: Acquire fact-checking article data, which includes at least the sentence sequence of the article text, sentence-level rhetorical structure tags, article-level statistical features, and the natural search traffic of the article's corresponding Uniform Resource Locator; Step S2: Construction of prediction target and control variables: The natural search traffic is used as the prediction target, and the article-level features other than natural search traffic are used as static control variables. The prediction target is then standardized. Step S3: Sentence-level semantic encoding: Organize each fact-checking article into a sentence sequence according to the sentence order, and use a pre-trained text embedding model to encode each sentence into a fixed-dimensional sentence semantic vector to generate its text representation; Step S4: Construction of Rhetorical Structure Sequence: Convert the rhetorical structure label of each sentence into a category index, form a rhetorical structure sequence according to the order in which the sentences appear in the article, and use staggered rhetorical input on the encoder input side so that the current position receives the rhetorical label embedding of the previous position, and the first position receives the start mark; Step S5: Construction of Rhetorical Structure Causal Transformer and Multi-Input Encoding: Construct a rhetorical structure causal transformer including a multi-input encoder, a rhetorical classification head, and a result prediction head. The multi-input encoder receives sentence-level text representations, misplaced rhetorical representations, and article-level static control variables. It adopts a dual-branch structure with parallel text and rhetorical branches, and is composed of multiple stacked multi-input blocks. Each multi-input block performs global multi-head self-attention on the text and rhetorical branches respectively. Then, it realizes the information interaction between text and rhetoric through bidirectional cross-attention. The article-level static control variables are projected and broadcast to each sentence position to obtain the final encoded text hidden state and misplaced rhetorical hidden state, which are used for subsequent fusion representation construction and natural search traffic prediction encoding results. Step S6: Fusion Representation Construction and Rhetorical Classification: After the last multi-input block, the final text hiding state of the current sentence output in step S5 is added to the corresponding misaligned rhetorical hiding state, and the initial sentence-level fusion representation of the integrated text and rhetorical information is obtained through linear mapping, activation, random deactivation and layer normalization; a rhetorical classification head is set, and the probability distribution of the current rhetorical label is output according to the sentence-level fusion representation. Step S7: Natural search traffic prediction: Set the result prediction head, concatenate the sentence-level fusion representation with the current position real rhetorical tag embedding and input it into the result prediction head, output the sentence-level local prediction value, and then perform masked average pooling and global mapping on the local prediction values ​​of all valid sentences to obtain the article-level natural search traffic prediction value in the standardized space. Step S8: Alternating Loss Optimization: Calculate the rhetoric classification loss based on the current rhetoric tag probability distribution output in Step S6, calculate the result prediction loss based on the standardized prediction value output in Step S7, and calculate the counterfactual domain confusion loss with the goal of uniform distribution of rhetoric categories; In the first optimization stage, fix the multi-input encoder, sentence-level representation construction parameters, and result prediction head, and update the rhetoric classification head to minimize the rhetoric classification loss; In the second optimization stage, fix the rhetoric classification head, update the multi-input encoder, sentence-level representation construction parameters, and result prediction head to minimize the sum of the result prediction loss and the weighted counterfactual domain confusion loss; In the second optimization stage, the sentence-level fusion representation reduces the discriminability of the current rhetoric tag while retaining relevant information of natural search traffic prediction, forming a sentence-level balanced representation for subsequent counterfactual analysis and model prediction; Step S9: Counterfactual Substitution and Rhetorical Contribution Analysis: Using the model optimized in Step S8, keeping the model parameters, factual text representation, factual misalignment rhetorical representation, attention mask, and static control variables unchanged, only replace the current actual rhetorical tag embeddings at one or more positions with candidate tag embeddings on the result prediction head input side; after each replacement, re-perform the forward computation to obtain a balanced representation consistent with the factual computation under the condition that the encoder input and model parameters remain unchanged, and calculate the difference between factual prediction and counterfactual prediction to measure the marginal contribution of candidate rhetorical tags to the natural search traffic prediction results relative to factual tags; Step S10: Output of prediction results and explanation report: Using the model optimized in step S8, obtain standardized prediction values ​​according to the prediction process described in step S7, and denormalize the obtained standardized prediction values ​​to the original scale of the prediction target, and output the natural search traffic prediction value and word contribution analysis report for each article.

2. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles in accordance with claim 1, characterized in that, The static control variables in step S2 include one or more of the following: Backlinks, Keywords (organic search keywords), Article Length, Sentiment, Modal (multimodal display features), and Topic (topic variable), expressed as follows: In the formula, Indicates an article index; Indicates the first The predictive objectives of this article; This represents the observed organic search traffic for the corresponding article; Represents a vector of static control variables; Indicates the number of topic variables, multimodal display features Used to indicate whether an article contains images, videos, charts, or other non-plain text display elements; The prediction target is standardized, and the standardized training target is represented as follows: In the formula, Indicates the first The standardized training objectives for this article; and represent the mean and standard deviation of the predicted targets in the training set, respectively; This represents a positive smoothing constant to prevent the denominator from being zero.

3. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles in accordance with claim 1, characterized in that, Step S3 specifically includes: Each sentence is encoded into a fixed-dimensional sentence vector using a pre-trained text embedding model, and each article is padded or truncated to a preset maximum number of sentences. Attention masks are generated to identify valid sentences and padded positions. In the formula, Indicates the sentence position index; Indicates the first The first article One sentence; This represents the semantic vector of the corresponding sentence; The parameter is Pre-trained text embedding models; Represents the dimension of the sentence vector; Indicates an attention mask; Indicates the first The number of effective sentences in an article.

4. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles according to claim 1, characterized in that, Step S4 specifically includes: The rhetorical structure tags adopt a step-level tagging system: macro steps include summary, introduction of false information, refutation of false information, conclusion, and supplementary reading materials; among them, the introduction of false information is further subdivided into three levels: the spread of false information, the description of the content of false information, and the introduction of relevant information; the refutation of false information is further subdivided into three levels: source verification and content comparison, logical reasoning or factual statement, and authoritative citation; when used for model coding, the summary, conclusion, supplementary reading materials, and the above six levels together constitute nine mutually exclusive sentence-level rhetorical structure categories; The initial text hiding state and the initial misaligned rhetorical hiding state are represented as follows: In the formula, and They represent the first Current sentence position in the article The initial hidden text state, and the source index paired with the sentence. The initial misplaced rhetoric is hidden; This represents the initial misaligned rhetorical hidden state obtained by using the start marker at the beginning of the sequence; exist The time is from the previous position label structure; and These represent the text projection matrix and the bias, respectively. This represents the rhetorical tag embedding function on the encoder input side; the start tag START indicates the special tag used at the beginning of the sequence.

5. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles according to claim 1, characterized in that, In step S5, each multi-input block of the rhetorical structure causal transformer performs the following operations sequentially: Multi-head self-attention is performed on both the textual and rhetorical hidden states, and a learnable relative position encoding is introduced to preserve the relative order and proximity relationships between sentences. In the formula, This indicates that multiple block layer numbers can be entered. and These represent the hidden state of the previous layer of text and the hidden state of rhetoric, respectively. and These represent text self-attention output and rhetorical self-attention output, respectively. Indicates an attention mask; Represents relative position encoding; Indicates multi-headed attention; Secondly, perform bidirectional cross-attention, so that the text representation query rhetorical context and the rhetorical representation query text context are: In the formula, The text represents the cross-attention output obtained from the rhetorical context of the query; The output represents the cross-attention response obtained from the context of the query text; Then, the article-level static control variables are projected onto the model dimension and broadcast to each sentence position, merging with the text and rhetorical representations: In the formula, Indicates the first In the article, the first The sentence is processed by the first The hidden text state obtained after processing multiple input blocks; This indicates the source index that is paired with this sentence. The misplaced rhetorical label after the first The rhetorical hidden state obtained after processing multiple input blocks; Indicates the sequence number of the multi-input block; This represents the article-level static control variable that is projected and broadcast to each sentence position; and These represent the static feature projection matrix and the bias, respectively. Indicates a feedforward network; Representation layer normalization.

6. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles according to claim 1, characterized in that, Step S6 specifically includes: Sentence-level fusion representations are constructed according to the following formula: In the formula, This represents the sentence-level fusion representation obtained by fusing the final text hidden state and the final misaligned rhetorical hidden state; Indicates the total number of multiple input blocks; Indicates the first In the article, the first The final hidden text state obtained after processing the last multi-input block of each sentence; This indicates the source index that is paired with this sentence. The rhetorical hidden state obtained after the misaligned rhetorical tags are processed by the last multi-input block; and These represent the sentence-level fusion representation mapping matrix and the bias, respectively. Represents the activation function of the exponential linear unit; Indicates a random deactivation operation; Representation layer normalization; The rhetoric classification head predicts the probability distribution of the current rhetoric tag based on the fusion representation: In the formula, The rhetorical tag prediction probability vector represents the current position; This indicates the category in the vector. The probability of; Indicates the heading of a rhetorical device category; and These represent the weight and bias of the classification head, respectively; This represents the normalized exponential function.

7. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles in accordance with claim 1, characterized in that, Step S7 specifically includes: Sentence-level local prediction values ​​are calculated using the following formula: In the formula, Indicates the first Article No. Local predicted values ​​for each sentence; This represents a local outcome prediction network; This indicates the current rhetorical tag embedding function used by the result prediction head; The actual rhetorical tag indicates the current location; the semicolon in square brackets indicates vector concatenation; The predicted article-level organic search traffic in the normalized space is obtained by masking and averaging the local predicted values ​​of all valid sentences, followed by global mapping: In the formula, Indicates the first The predicted value of this article in the standardized space; This represents the sentence-level fusion representation sequence of the article; This indicates a sequence of factual and rhetorical tags in the article; This indicates the prediction header, generated by the current label embedding function. Local outcome prediction network Masked average pooling and global result mapping network Consisting in sequence; This represents the positive smoothing constant.

8. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles in accordance with claim 1, characterized in that, Step S8 specifically includes: The rhetorical classification loss, outcome prediction loss, and counterfactual domain confusion loss are respectively expressed as follows: The alternating optimization objective is expressed as: In the formula, Indicates loss in rhetorical classification; Indicates the predicted loss; This indicates confusion and loss in the counterfactual domain; Indicates the rhetorical category header parameter; The parameters represent the encoding part, including multi-input encoder parameters and sentence-level representation construction parameters; This indicates the result prediction header parameters, including the parameters of the current rhetorical tag embedding function; Indicates the weight of the confusion loss; This represents the total number of rhetorical categories, with the internal category index ranging from 0 to... .

9. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles in accordance with claim 1, characterized in that, Step S9 specifically includes: Counterfactual predictions are calculated using the following formula: In the formula, This represents the fact balance representation sequence calculated while keeping model parameters, fact text, fact misalignment rhetoric sequence, attention mask, and static control variables constant. This indicates a candidate counterfactual rhetorical sequence used only on the result prediction header input side; if no replacement is specified, the position remains the same as the factual rhetorical sequence. Indicates the result prediction header; This represents the standardized counterfactual prediction value obtained solely from the effect of candidate rhetorical labels, assuming that the factual balance representation remains unchanged. The single-sample counterfactual effect is calculated using the following formula: In the formula, Indicates the first The article's factual rhetorical sequence; Indicates the first The change in standardized predictions when an article is replaced by a factual rhetorical sequence with a candidate counterfactual rhetorical sequence; its absolute value represents the marginal impact of the candidate replacement. Calculate the average embedding vector of all rhetorical categories in the result prediction head: In the formula, This represents the average vector of all rhetorical category embeddings in the prediction header of the result; This represents the internal rhetoric category index, with values ​​ranging from 0 to... ; This indicates the rhetorical tag embedding function used in the result prediction head; The marginal contribution of rhetorical structure categories is calculated using label-by-label masking according to the following formula: In the formula, Indicates rhetorical category The average absolute conditional marginal contribution; This represents a collection of sample articles containing this rhetorical category; This indicates that the category in the prediction header will be used to predict the result. The embedding at the corresponding position is replaced with the standardized prediction value obtained after average embedding; This represents the standardized predicted value corresponding to the factual rhetorical sequence.

10. The method for predicting natural search traffic and analyzing the contribution of fact-checking articles according to claim 1, characterized in that, The anti-normalization mentioned in step S10 is represented as follows: In the formula, This represents the predicted natural search traffic value restored to the original scale of the predicted target. This represents the standardized predicted value.

Images