An AI big data-based college propaganda operation analysis method

Abstract

The application relates to the technical field of data processing of college propaganda operation analysis, and discloses a college propaganda operation analysis method based on AI big data, which comprises the following steps: obtaining audience interactive behavior flow records; analyzing interactive texts to calculate text feature quantities, extracting sliding pause time series in front-end behavior time series data, calculating second-order central moments to generate interactive feature quantities; determining content complexity weights based on display content information entropy, dynamically adjusting the modal trust weights of the text feature quantities and the interactive feature quantities according to the content complexity weights; performing weighted calculation on each feature quantity by using the modal trust weights, and aggregating to generate a comprehensive participation index; inputting the index into an analysis model to generate a propagation coverage depth value; and dynamically guiding the modal weights by the content complexity, solving the evaluation deviation caused by the lack of explicit feedback when the audience processes high-complexity information, and improving the attribution accuracy of the propaganda effect evaluation.

Description

Technical Field

[0001] This invention relates to a method for analyzing university publicity and operations based on AI big data, belonging to the field of data processing technology for university publicity and operations analysis. Background Technology

[0002] Currently, operational analysis for university promotion scenarios generally adopts a data processing method based on traffic indicators. This method evaluates the effectiveness of promotion by counting webpage views, interaction likes, and information forwards. This method is based on the traditional marketing funnel model, which assumes that the frequency of interaction feedback directly represents the accumulation of brand value.

[0003] However, the research achievements or disciplinary advantages carried by university publicity content have a high information density. When faced with such high semantic entropy content, the audience will have a specific cognitive resistance, which will prevent the information from being transformed into explicit interactive behavior in a short period of time. The existing monitoring schemes lack a quantitative hedging mechanism for the information potential energy of the content itself, which will judge high-value signals as low-popularity content, thus causing the evaluation results to shift towards shallow traffic data. Existing data monitoring methods have shortcomings in the front-end interactive recognition hardware perception dimension and the back-end analysis logic software control method. For example, Chinese invention patent application with publication number CN120450787A discloses a method and system for enterprise operation analysis and supervision for enterprise management. It extracts the features of publicity and quality inspection data through 3D convolutional neural network and language processing model, and performs consumer review clustering supervision based on cosine similarity feature matching and K-Means algorithm.

[0004] Therefore, the technical problem to be solved by this invention is how to construct an analytical mechanism that can objectively characterize the impedance of content information and quantify cross-modal cognitive work, so as to solve the technical problem that the value assessment of high-potential publicity content in universities is blocked due to the distortion of traffic data in the evaluation system. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A method for analyzing university publicity and operation based on AI big data, comprising the following steps:

[0006] Step 101: Obtain the interaction behavior log of the audience within the display interface. The interaction behavior log includes the interaction text and the front-end behavior time sequence data associated with the interaction text.

[0007] Step 102: Analyze the frequency of logical connectors and the proportion of unemotional words in the interactive text, and calculate the text feature quantity that reflects the semantic and logical complexity of the text.

[0008] Step 103: Extract the time series of the audience's swiping pauses in the display interface from the front-end behavior time series data, calculate the second central moment of the swiping pause time series, and generate interactive feature quantities that reflect the dispersion of the audience's reading rhythm.

[0009] Step 104: Extract the display content within the display interface and calculate the information entropy to determine the content complexity weight that reflects the difficulty of understanding the display content;

[0010] Step 105: Based on the content complexity weight, dynamically adjust the first modality trust weight corresponding to the text feature quantity and the second modality trust weight corresponding to the interaction feature quantity; wherein, the content complexity weight is negatively correlated with the first modality trust weight and positively correlated with the second modality trust weight.

[0011] Step 106: Use the first modality trust weight to perform weighted calculation on the text feature quantity, and use the second modality trust weight to perform weighted calculation on the interaction feature quantity, and aggregate to generate a comprehensive participation index;

[0012] Step 107: Input the comprehensive participation indicators into the preset analysis model to generate a dissemination coverage depth value to evaluate the publicity effect, and output the audience's deep interaction status label in the display interface based on the dissemination coverage depth value.

[0013] Preferably, the process of calculating the second-order central moment of the sliding pause time series in step 103 includes the following steps: Step 201, cleaning the front-end behavior time series data, removing invalid interference data with a dwell time of less than 10ms and static abnormal data with a dwell time of more than 3600s, and obtaining the effective dwell time of the audience in each content block within the display interface; Step 202, reconstructing the effective dwell time into a sliding pause time series according to the temporal logic of the audience's swiping operation; Step 203, calculating the mean of the sliding pause time series, and calculating the second-order central moment based on the squared difference between the mean and each effective dwell time.

[0014] Preferably, step 104 includes the following steps: step 301, extracting the term frequency and the proportion of logical guide words in the displayed content; step 302, calculating the text information entropy corresponding to the displayed content based on the term frequency and the proportion of logical guide words; step 303, obtaining the cognitive difficulty benchmark of the domain to which the displayed content belongs, and performing normalization processing in combination with the text information entropy to generate content complexity weights.

[0015] Preferably, step 106 includes: calculating a comprehensive participation index based on a linear weighted algorithm using the first modality trust weight, the second modality trust weight, text features, and interaction features. The calculation process follows the formula: ,in, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, Let be the interactive feature quantity, and satisfy α+β=1.

[0016] Preferably, the logic for dynamically adjusting the first modality trust weight and the second modality trust weight in step 105 includes: when the content complexity weight is greater than a preset complexity threshold, a weight offset operation is performed; the weight offset operation includes: lowering the set value of the first modality trust weight to a first preset range, and simultaneously raising the set value of the second modality trust weight to a second preset range.

[0017] Preferably, after generating the dissemination coverage depth value in step 107, the following steps are also included: Step 601, establishing a publicity evaluation space that includes professional value dimension, brand value dimension and cultural value dimension; Step 602, performing vectorization processing on the displayed content and projecting it into the publicity evaluation space to determine the attribute components of the displayed content in different dimensions.

[0018] Preferably, the method further includes the following steps: Step 701, configuring differentiated time decay parameters for each dimension in the publicity evaluation space, wherein the decay parameter for the professional value dimension is less than the decay parameter for the cultural value dimension; Step 702, performing time-series decay correction on the dissemination coverage depth value generated by a single interaction based on the differentiated time decay parameters.

[0019] Preferably, the method further includes the following steps: Step 801, performing time-series integral calculation on the dissemination coverage depth value within a preset time window to generate a cumulative effectiveness index for the corresponding publicity event; Step 802, constructing an evaluation threshold that dynamically adjusts with the content complexity weight, and comparing the cumulative effectiveness index with the evaluation threshold.

[0020] Preferably, the method further includes the following steps: Step 901, when the cumulative performance index exceeds the evaluation threshold and the character length of the corresponding interactive text is less than 5 characters, the corresponding audience is marked as a deep engagement target; Step 902, the conversion probability of the deep engagement target is calculated, and a digital asset accumulation index for optimizing resource allocation strategy is generated.

[0021] Preferably, after generating the propagation coverage depth value in step 107, the method further includes: adjusting the weight distribution of subsequent push content according to the magnitude of the propagation coverage depth value; and increasing the push frequency for similar audience feature tags for display pages whose propagation coverage depth value exceeds a preset depth threshold.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. In the analysis of university publicity and operation using AI big data, the system utilizes content potential coefficient and impedance compensation mechanism to address the value shielding problem of high information entropy content during marketing data processing. By analyzing the density of scarce terms and the depth of syntactic structure in the publicity content, the system quantifies the information entropy impedance required for the audience to understand the content. Based on this, the compensation gain factor increases non-linearly with the increase of content potential coefficient, resulting in a weighted amplification effect on the basic resonance. This mechanism compensates for the traffic attenuation caused by the cognitive threshold during the dissemination of hard-core academic content at the underlying data processing logic, avoiding the brand evaluation bias caused by simply relying on click volume in existing technologies, and enabling the data analysis results to truly restore the value of publicity events with profound influence.

[0024] 2. By analyzing interactive work and aggregating impedance compensation effectiveness, the system captures high-value feedback signals in a silent state. Addressing the aphasia experienced by audiences reading high-density information, the system no longer relies solely on the semantic richness of the interactive text. Instead, it extracts the variance of the audience's scrolling pause time series on the promotional page, generating a behavioral work index reflecting physical reading impedance. Based on the content potential coefficient, the system dynamically adjusts the modal trust weights of text and behavior. When dealing with content that is difficult to understand, the evaluation focus shifts to the audience's physical reading rhythm. This cross-modal work quantification method breaks the dependence of data monitoring on explicit expression, identifies deep cognitive resonance through the dispersion of physical behavior, and improves the attribution accuracy of marketing effectiveness monitoring in complex decision-making scenarios.

[0025] 3. By combining resonance filtering and time-series cumulative effectiveness calculation, the system achieves accurate mapping of long-cycle conversion processes. Based on the preset publicity value space, the system performs vector mapping on unstructured content and introduces differentiated time decay parameters corresponding to specific dimensions. The parameter configuration of long half-life dimensions such as academic reputation makes the evaluation of long-tail influence more durable. By accumulating the corrected effectiveness value of a single interaction over time and comparing the accumulated value with the dynamic threshold that changes with the content's potential energy, the system identifies high-quality content with low instantaneous traffic but strong brand asset accumulation effect. This strategy generation logic avoids evaluation distortion caused by time window truncation, provides management decisions with technical indicators that conform to the lagging characteristics of university publicity, and optimizes the allocation path of publicity resources. Attached Figure Description

[0026] Figure 1 This is a flowchart of the multimodal features of university publicity data processing and comprehensive evaluation according to the present invention;

[0027] Figure 2 This is a diagram illustrating the multidimensional projection and long-term digital asset accumulation analysis mechanism of the present invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0030] A method for analyzing university publicity and operations based on AI and big data includes the following steps:

[0031] Step 101: Obtain the interaction behavior log of the audience within the display interface. The interaction behavior log includes the interaction text and the front-end behavior time sequence data associated with the interaction text.

[0032] Step 102: Analyze the frequency of logical connectors and the proportion of unemotional words in the interactive text, and calculate the text feature quantity that reflects the semantic and logical complexity of the text.

[0033] Step 103: Extract the time series of the audience's swiping pauses in the display interface from the front-end behavior time series data, calculate the second central moment of the swiping pause time series, and generate interactive feature quantities that reflect the dispersion of the audience's reading rhythm.

[0034] Step 104: Extract the display content within the display interface and calculate the information entropy to determine the content complexity weight that reflects the difficulty of understanding the display content;

[0035] Step 105: Based on the content complexity weight, dynamically adjust the first modality trust weight corresponding to the text feature quantity and the second modality trust weight corresponding to the interaction feature quantity; wherein, the content complexity weight is negatively correlated with the first modality trust weight and positively correlated with the second modality trust weight.

[0036] Step 106: Use the first modality trust weight to perform weighted calculation on the text feature quantity, and use the second modality trust weight to perform weighted calculation on the interaction feature quantity, and aggregate to generate a comprehensive participation index;

[0037] Step 107: Input the comprehensive participation indicators into the preset analysis model to generate a dissemination coverage depth value to evaluate the publicity effect, and output the audience's deep interaction status label in the display interface based on the dissemination coverage depth value.

[0038] Preferably, the process of calculating the second-order central moment of the sliding pause time series in step 103 includes the following steps: Step 201, cleaning the front-end behavior time series data, removing invalid interference data with a dwell time of less than 10ms and static abnormal data with a dwell time of more than 3600s, and obtaining the effective dwell time of the audience in each content block within the display interface; Step 202, reconstructing the effective dwell time into a sliding pause time series according to the temporal logic of the audience's swiping operation; Step 203, calculating the mean of the sliding pause time series, and calculating the second-order central moment based on the squared difference between the mean and each effective dwell time.

[0039] Preferably, step 104 includes the following steps: step 301, extracting the term frequency and the proportion of logical guide words in the displayed content; step 302, calculating the text information entropy corresponding to the displayed content based on the term frequency and the proportion of logical guide words; step 303, obtaining the cognitive difficulty benchmark of the domain to which the displayed content belongs, and performing normalization processing in combination with the text information entropy to generate content complexity weights.

[0040] Preferably, step 106 includes: calculating a comprehensive participation index based on a linear weighted algorithm using the first modality trust weight, the second modality trust weight, text features, and interaction features. The calculation process follows the formula: ,in, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, Let be the interactive feature quantity, and satisfy α+β=1.

[0041] Preferably, the logic for dynamically adjusting the first modality trust weight and the second modality trust weight in step 105 includes: when the content complexity weight is greater than a preset complexity threshold, a weight offset operation is performed; the weight offset operation includes: lowering the set value of the first modality trust weight to a first preset range, and simultaneously raising the set value of the second modality trust weight to a second preset range.

[0042] Preferably, after generating the dissemination coverage depth value in step 107, the following steps are also included: Step 601, establishing a publicity evaluation space that includes professional value dimension, brand value dimension and cultural value dimension; Step 602, performing vectorization processing on the displayed content and projecting it into the publicity evaluation space to determine the attribute components of the displayed content in different dimensions.

[0043] Preferably, the method further includes the following steps: Step 701, configuring differentiated time decay parameters for each dimension in the publicity evaluation space, wherein the decay parameter for the professional value dimension is less than the decay parameter for the cultural value dimension; Step 702, performing time-series decay correction on the dissemination coverage depth value generated by a single interaction based on the differentiated time decay parameters.

[0044] Preferably, the method further includes the following steps: Step 801, performing time-series integral calculation on the dissemination coverage depth value within a preset time window to generate a cumulative effectiveness index for the corresponding publicity event; Step 802, constructing an evaluation threshold that dynamically adjusts with the content complexity weight, and comparing the cumulative effectiveness index with the evaluation threshold.

[0045] Preferably, the method further includes the following steps: Step 901, when the cumulative performance index exceeds the evaluation threshold and the character length of the corresponding interactive text is less than 5 characters, the corresponding audience is marked as a deep engagement target; Step 902, the conversion probability of the deep engagement target is calculated, and a digital asset accumulation index for optimizing resource allocation strategy is generated.

[0046] Preferably, after generating the propagation coverage depth value in step 107, the method further includes: adjusting the weight distribution of subsequent push content according to the magnitude of the propagation coverage depth value; and increasing the push frequency for similar audience feature tags for display pages whose propagation coverage depth value exceeds a preset depth threshold.

[0047] Example 1: In a promotional campaign targeting the annual major research achievements of research universities, the system receives display content containing in-depth academic analysis and the resulting stream of audience interaction behavior. When audiences encounter such content with numerous technical terms and complex syntactic structures, their cognitive load increases, leading to a decrease in the frequency of explicit interactive behaviors such as long text comments. Audiences enter a state of focused, silent reading, with output interactive text shorter than 5 characters and lacking logical connectors. If conventional data statistics logic equates this silent state with low engagement, the weight of this academic promotional content in the recommendation algorithm diminishes, resulting in marginalization in the promotional resource allocation strategy. The system extracts the frequency of terms and the proportion of logical connectors in the displayed content to calculate text information entropy. This information is then normalized using a cognitive difficulty benchmark for the relevant domain to determine the content complexity weight, reflecting the difficulty of understanding the displayed content. When the content complexity weight exceeds a preset complexity threshold, the system triggers a weight shift operation, lowering the first modality trust weight α (reflecting the semantic and logical complexity of the text) to a first preset range, while simultaneously raising the second modality trust weight β (reflecting the interaction features) to a second preset range. Under this processing logic, the system analyzes the frequency of logical connectors and the proportion of non-emotional words in the interactive text to calculate text features. The associated front-end behavior time-series data is cleaned and reconstructed into a time series of audience swiping and pausing within the display interface. By calculating the second-order central moments of the mean of this swiping and pausing time series and the squared differences of each effective dwell time, an interactive feature quantity reflecting the dispersion of the audience's reading rhythm is generated. The linear weighted algorithm follows the formula Calculate the comprehensive participation index ,in, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, The system converts static front-end time-series data into quantified behavioral features based on content complexity weights and modal trust weights that are higher than reading behavior features, providing data compensation input for semantic features.

[0048] The system inputs the aggregated comprehensive participation indicators into a preset analysis model to generate a dissemination coverage depth value to evaluate the promotional effect, and outputs the audience's deep interaction status tags within the display interface based on the dissemination coverage depth value. Display content with high content complexity weights obtains a dissemination coverage depth value exceeding the preset depth threshold based on its high dispersion score in interaction feature quantities, thereby adjusting the weight distribution of subsequent push content and increasing the push frequency for similar audience feature tags. In the established promotional evaluation space that includes professional value dimension, brand value dimension, and cultural value dimension, the system performs vectorization processing on the display content and projects it to determine each dimension. The system adjusts the depth of dissemination coverage generated by a single interaction by adjusting the attenuation parameter configuration based on the professional value dimension being less than the cultural value dimension. It also calculates the cumulative effectiveness index of the corresponding publicity event by performing time-series integral calculation within a preset time window. The system compares the cumulative effectiveness index with the evaluation threshold that is dynamically adjusted according to the content complexity weight. When the index exceeds the threshold and the length of the interactive text is less than 5 characters, the depth participation target is marked and its conversion probability is calculated. The system generates a digital asset accumulation index to optimize resource allocation strategy. The resource allocation weight of a specific display interface is maintained in a targeted digital asset accumulation state based on this index.

[0049] Example 2: In a real-world test scenario evaluating the promotional effectiveness of academic display content, the test environment accessed an anonymized front-end behavior time-series data set recorded continuously for 30 natural days by an internal server. To construct a perturbation environment equivalent to real mobile reading conditions, a network jitter parameter with a random delay of 50ms to 100ms, accounting for 15% of the time-series data set, was injected, along with high-frequency short-duration sliding signals caused by physical touches at the screen edge. The system received the aforementioned time-series data set containing composite perturbations, extracted the interactive text and information entropy of the corresponding display interface, and, for the cleaning stage of the front-end behavior time-series data, determined that the effective dwell time time boundary parameter needed to achieve an optimized balance between filtering out physical accidental touches and retaining fragmented reading characteristics. Based on the audience's visual processing... Based on the minimum physiological response cycle of information and the physical extreme value of the capacitance sampling frequency of the smart terminal touch screen, the system establishes a judgment model for time boundaries. It stipulates that when the captured single screen contact time is lower than the minimum threshold for visual information acquisition, it is classified as invalid physical swipe; when the single dwell time exceeds the default sleep time of the terminal screen hardware, it is classified as device-unattended static. Based on the judgment model, the system establishes a standardized cleaning procedure to judge data with a dwell time of less than 10ms as invalid interference data and data with a dwell time of more than 3600s as static abnormal data. It extracts the effective dwell time of the audience in each content block in the display interface and reconstructs it into a swipe pause time sequence. This procedure eliminates 14.7% of pseudo-interaction noise in the above-mentioned test conditions containing high-frequency short-time swipe signals.

[0050] To obtain quantitative evidence for the calculation logic of the comprehensive engagement index, a comparative analysis was conducted between a control group and an experimental group in the test environment. The control group used a single evaluation dimension—calculating text feature quantities based on the frequency of logical connectives in the interactive text—while the experimental group implemented a cross-modal processing step to extract sliding pause time series and calculate their second-order central moments to generate interaction feature quantities. When processing target display content with a complexity weight exceeding a preset threshold, the control group achieved a recall rate of 41.2% for deeply engaged targets, while the experimental group achieved a higher recall rate based on the calculated text feature quantities. Interaction feature quantity Under the dynamic weighted logic of lowering the first modality trust weight α and raising the second modality trust weight β, following the formula... Calculate the comprehensive participation index Its identification and recall rate climbed to 93.5%, of which, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, The experimental environment simultaneously introduced a gradient validation group targeting the lower time limit parameter. When the lower limit parameter shrank to 5ms, high-frequency false touch noise penetrated, causing distortion of the calculated second-order central moment, and the recognition recall rate dropped to 76.8%. When the lower limit parameter widened to 20ms, some effective information from rapid scanning was hard-truncated, and the recognition recall rate deteriorated non-linearly, falling to 82.1%. The above quantitative trends confirmed that 10ms is the working extreme point for suppressing physical disturbances and ensuring the integrity of the time sequence, and parameter deviation induces the decay of evaluation accuracy. The test data verified the dynamic adjustment model of content complexity weights. The system's logic for processing state confidence and aggregating second-order central moments to quantify values ​​improves engineering usability in suppressing disturbances in front-end time-series data. In silent reading scenarios where the interactive text length is less than 5 characters, the system transforms the discrete physical characteristics of sliding pauses into quantified comprehensive participation indicators, providing data compensation input for dimensionality loss caused by calculations based on single text semantic logic. The aforementioned execution mechanism adjusts the resource allocation weights of specific display interfaces based on the aggregated propagation coverage depth value, avoiding the imbalance in digital asset allocation faced by complex promotional content due to explicit interaction decay under conventional text frequency statistics algorithms.

[0051] Example 3: In the deployment and calibration scenario of the backend analysis algorithm of the university publicity system, before processing the audience interaction behavior log, the system performs the calibration of the judgment threshold and the setting of the internal network architecture of the analysis model. It obtains the set of display content within the past 30 natural days to construct the calibration data source, extracts the frequency of professional terms and the proportion of logical guiding words in each display content; calculates the product of the term frequency and the proportion of logical guiding words, and calculates the absolute value of the natural logarithm of the product result to determine the text information entropy; the system extracts the average text information entropy of popular science texts in the relevant field as the cognitive difficulty benchmark, calculates the ratio of the text information entropy of the display content to the cognitive difficulty benchmark to generate the content complexity weight; the system arranges the content complexity weights of all display content in the calibration data source according to their numerical size, and extracts the content complexity weight at the 85th percentile as the preset complexity threshold to trigger the weight offset operation.

[0052] The system inputs the aggregated comprehensive participation index into the analysis model, which is configured as a multilayer perceptron architecture containing one input layer, two fully connected hidden layers, and one output layer. After the comprehensive participation index enters the input layer, the activation function in the first fully connected hidden layer maps the comprehensive participation index to a high-dimensional feature space, extracting the nonlinear coupling component reflecting the superposition state of text feature quantities and interaction feature quantities. In the second fully connected hidden layer, a linear correction unit is used to project the nonlinear coupling component to the dimension-reduced quantization space. The output layer receives the dimension-reduced scalar value and outputs the propagation coverage depth value. The system compares the propagation coverage depth value with the evaluation threshold that is dynamically adjusted according to the content complexity weight, and outputs the depth interaction status label of the audience in the display interface based on the comparison result.

[0053] Example 4: In the on-site deployment and initial debugging of the system for newly established university publicity nodes, the system initiates the baseline calibration procedure and the underlying data desensitization mechanism without accumulating local interaction history data; the front-end acquisition module intercepts the incoming raw device fingerprint parameters, uses a secure hash algorithm to unidirectionally map the physical identity parameters into an encrypted string, and severs the association link between the physical identity and the front-end behavior time-series data; it imports an offline test corpus containing standard literature of specific sub-disciplines, extracts the frequency of professional terms and the distribution data of logical guiding words in the literature, and generates an initial cognitive difficulty benchmark based on the text information entropy calculation logic; the system writes the initial cognitive difficulty benchmark into the cache register as a cold start comparison parameter for real-time data stream feature aggregation.

[0054] The system extracts simulated dwell time series associated with the offline test corpus and a preset expected propagation depth, transforming these parameters into a standard input tensor containing initial text features and initial interaction features. This standard input tensor is then injected into the analysis model of a multilayer perceptron architecture. The mean squared error between the output propagation coverage depth and the preset expected propagation depth is calculated using a loss function. Based on this mean squared error, the activation node weights and linear correction unit parameters are updated in reverse along the fully connected hidden layer topology until the mean squared error decreases and converges to a set lower limit interval for the loss. The specific configuration of this multilayer perceptron architecture is as follows: the input layer has 2 nodes, the first of which is a fully connected hidden layer... The first fully connected hidden layer contains 16 neurons, the second fully connected hidden layer contains 8 neurons, and the output layer has 1 node. During training, the initial step size for weight updates is set to 0.01, and the target threshold for the loss function is set to 0.0001. If the error descent gradient is less than 0.000001 for 50 consecutive iterations, the model is considered to have converged, and the weight matrix is ​​fixed. The system extracts the numerical boundary point corresponding to the model's output layer as the preset complexity threshold and calibration parameters for determining the depth interaction state. The updated network node weight matrix is ​​fixed, and the analysis model is switched to a real-time evaluation state that receives anonymized front-end behavioral time-series data.

[0055] Example 5: Targeting Text Feature Quantities The calculation process involves analyzing the frequency of transition words and the proportion of unemotional technical terms in the interactive text, and introducing a logical coherence correction coefficient determined based on the statistical distribution patterns of audience feedback grammar in an offline corpus; targeting interactive features... The system calculates by performing differential operations on the time-series data of front-end behavior, extracting the time sequence of the audience's swiping pauses within the display interface, and then using the formula... Calculate the dispersion of the audience's reading rhythm, where, Here, n represents the number of sliding pauses, and n is the interaction feature. For the i-th valid stay, The average effective dwell time is used as the threshold for determining data as interference noise during the real-time operation phase. The system opens the front-end data cleaning logic gate and sets the physical boundary threshold for data with a dwell time of less than 10ms or more than 3600s.

[0056] The system uses a linear weighted algorithm following the formula Generate comprehensive participation indicators ,in, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, The system uses interactive features as a parameter. To address fluctuations in content complexity weights, the system retrieves the corresponding weight allocation scheme from the preset feature weight matrix based on the content complexity weight. When the content complexity weight is between 0.8 and 1.2, the system selects a balanced weight configuration. When it exceeds 1.5, the system increases the second modality trust weight β to 0.7 and simultaneously decreases the first modality trust weight α to 0.3. The system performs online alignment correction between the generated dissemination coverage depth value and the average effectiveness value of similar historical topics, and adjusts the resource allocation weight for specific promotional content based on the corrected difference.

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing university publicity and operations based on AI big data, characterized in that, Includes the following steps: Step 101: Obtain the interaction behavior log of the audience within the display interface. The interaction behavior log includes the interaction text and the front-end behavior time sequence data associated with the interaction text. Step 102: Analyze the frequency of logical connectors and the proportion of unemotional words in the interactive text, and calculate the text feature quantity that reflects the semantic and logical complexity of the text. Step 103: Extract the time series of the audience's swiping pauses in the display interface from the front-end behavior time series data, calculate the second central moment of the swiping pause time series, and generate interactive feature quantities that reflect the dispersion of the audience's reading rhythm. Step 104: Extract the display content within the display interface and calculate the information entropy to determine the content complexity weight that reflects the difficulty of understanding the display content; Step 105: Based on the content complexity weight, dynamically adjust the first modality trust weight corresponding to the text feature quantity and the second modality trust weight corresponding to the interaction feature quantity; wherein, the content complexity weight is negatively correlated with the first modality trust weight and positively correlated with the second modality trust weight. Step 106: Use the first modality trust weight to perform weighted calculation on the text feature quantity, and use the second modality trust weight to perform weighted calculation on the interaction feature quantity, and aggregate to generate a comprehensive participation index; Step 107: Input the comprehensive participation indicators into the preset analysis model to generate a dissemination coverage depth value to evaluate the publicity effect, and output the audience's deep interaction status label in the display interface based on the dissemination coverage depth value. The process of calculating the second-order central moments of the sliding pause time series in step 103 includes: cleaning the front-end behavior time series data, removing invalid interference data with a dwell time of less than 10ms and static abnormal data with a dwell time of more than 3600s, obtaining the effective dwell time of the audience in each content block in the display interface; reconstructing the effective dwell time into a sliding pause time series according to the temporal logic of the audience's swiping operation; calculating the mean of the sliding pause time series, and calculating the second-order central moments based on the squared difference between the mean and each effective dwell time. Step 104 includes: extracting the term frequency and the proportion of logical guide words in the displayed content; calculating the text information entropy corresponding to the displayed content based on the term frequency and the proportion of logical guide words; obtaining the cognitive difficulty benchmark of the domain to which the displayed content belongs, and performing normalization processing in combination with the text information entropy to generate content complexity weights.

2. The method for analyzing university publicity and operations based on AI big data according to claim 1, characterized in that, Step 106 includes: calculating a comprehensive participation index based on a linear weighted algorithm, using the first modality trust weight, the second modality trust weight, text features, and interaction features. The calculation process follows the formula: ,in, To integrate participation metrics, α represents the first modality trust weight, and β represents the second modality trust weight. For text features, Let be the interactive feature quantity, and satisfy α+β=1.

3. The method for analyzing university publicity and operations based on AI big data according to claim 1, characterized in that, The logic for dynamically adjusting the first modality trust weight and the second modality trust weight in step 105 includes: when the content complexity weight is greater than the preset complexity threshold, a weight offset operation is performed; the weight offset operation includes: lowering the set value of the first modality trust weight to the first preset range, and simultaneously raising the set value of the second modality trust weight to the second preset range.

4. The method for analyzing university publicity and operations based on AI big data according to claim 1, characterized in that, After generating the dissemination coverage depth value in step 107, the following steps are also included: Step 401, establish a publicity evaluation space that includes professional value dimension, brand value dimension and cultural value dimension; Step 402, perform vectorization processing on the displayed content and project it into the publicity evaluation space to determine the attribute components of the displayed content in different dimensions.

5. The method for analyzing university publicity and operations based on AI big data according to claim 4, characterized in that, It also includes the following steps: Step 501: Configure differentiated time decay parameters for each dimension in the publicity evaluation space, wherein the decay parameter for the professional value dimension is less than that for the cultural value dimension; Step 502: Based on the differentiated time decay parameters, perform time-series decay correction on the dissemination coverage depth value generated by a single interaction.

6. The method for analyzing university publicity and operations based on AI big data according to claim 5, characterized in that, It also includes the following steps: Step 601: Perform time-series integral calculation on the dissemination coverage depth value within a preset time window to generate the cumulative effectiveness index of the corresponding publicity event; Step 602: Construct an evaluation threshold that dynamically adjusts with the content complexity weight, and compare the cumulative effectiveness index with the evaluation threshold.

7. The method for analyzing university publicity and operations based on AI big data according to claim 6, characterized in that, It also includes the following steps: Step 701: When the cumulative performance index exceeds the evaluation threshold and the character length of the corresponding interactive text is less than 5 characters, the corresponding audience is marked as a deep engagement target; Step 702: Calculate the conversion probability of the deep engagement target and generate a digital asset accumulation index to optimize resource allocation strategy.

8. The method for analyzing university publicity and operations based on AI big data according to claim 1, characterized in that, After generating the propagation coverage depth value in step 107, the following steps are also included: adjusting the weight distribution of subsequent push content according to the magnitude of the propagation coverage depth value; and increasing the push frequency for similar audience feature tags for display pages whose propagation coverage depth value exceeds the preset depth threshold.

Images