A marketing intelligent decision system based on multi-modal learning

Abstract

The application provides a marketing intelligent decision system based on multi-modal learning, and relates to the technical field of marketing intelligent decision.The application realizes the synchronous access and deep mining of three-source heterogeneous data of social media text flow, short video advertisement flow and transaction pulse flow through a multi-modal data analysis module, accurately extracts attribute vectors representing product function dimensions from the social media text flow, separates quality evaluation vectors and fluctuation value double indicators from the short video advertisement flow, and analyzes decision rhythm vectors and time data sets from the transaction pulse flow, thereby breaking through the limitations of traditional marketing systems in single-modal analysis;an inter-modal feature fusion module adopts three algebraic interaction mechanisms of element-by-element multiplication, dot product operation and vector module multiplication to generate three fusion features of attribute emergency index, advertisement timeliness coefficient and strategy matching degree, realizes the dynamic coupling of text sentiment, advertisement effect and transaction behavior, and solves the decision lag problem caused by the fragmentation of multi-modal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marketing intelligent decision-making, and particularly relates to a marketing intelligent decision-making system based on multi-modal learning. BACKGROUND

[0002] In the current digital marketing environment, enterprises face core challenges such as difficulty in integrating multi-source heterogeneous data, lag in decision-making response, and weakness in risk control. Traditional marketing decision-making systems usually rely on a single data modality, leading to one-sided cognition of product market performance. For example, social media text analysis can only capture user emotional tendencies, but cannot correlate with advertising effectiveness or real-time consumer behavior; while transaction data analysis can reflect sales results, it is difficult to trace the decision-making causes. Existing technologies attempt to improve decision-making efficiency through simple rule engines or multi-model parallel processing, but each modality data is still in a fragmented state, and text, video advertising performance and transaction behavior data lack deep coupling, making the system unable to capture key associations such as "how does emotional feedback after users see the advertisement affect the purchase decision rhythm", and in dynamic threshold control, traditional solutions use static empirical values or single-dimensional indicators, which are difficult to adapt to market sudden fluctuations, often resulting in misjudgment expansion or defensive delay.

[0003] Therefore, it is necessary to provide a marketing intelligent decision-making system based on multi-modal learning to solve the above technical problems. SUMMARY

[0004] To solve the above technical problems, the present application provides a marketing intelligent decision-making system based on multi-modal learning, which achieves the beneficial effect of integrating multi-modal data for intelligent decision-making of marketing.

[0005] The present application provides a marketing intelligent decision-making system based on multi-modal learning, comprising:

[0006] A multi-modal data analysis module: simultaneously accessing social media text stream, short video advertising stream and transaction pulse stream, extracting product attribute vector from social media text stream, quality evaluation vector and quality fluctuation value from short video advertising stream, and decision rhythm vector and transaction time data set from transaction pulse stream;

[0007] A cross-modal feature fusion module: based on the product attribute vector and the decision rhythm vector, calculating the attribute emergency index through element-by-element multiplication, based on the quality evaluation vector and the decision rhythm vector, generating the advertising timeliness coefficient through dot product operation, and based on the product attribute vector and the quality evaluation vector, generating the strategy matching degree through vector module multiplication;

[0008] A feature signal matrix generation module: combining the attribute emergency index, the advertising timeliness coefficient and the strategy matching degree into a current feature signal matrix, and extracting the principal eigenvalue and the secondary eigenvalue of the current feature signal matrix;

[0009] The double-threshold generation module calculates an expansion threshold based on the product attribute vector and the transaction time dataset, and calculates a defense threshold based on the quality fluctuation value, the current feature signal matrix, and the historical feature signal matrix;

[0010] The decision signal generation module generates an expansion signal and an expansion strength based on the main feature value and the expansion threshold, and generates a defense signal and a defense strength based on the secondary feature value and the defense threshold;

[0011] The decision instruction generation module inputs the expansion signal, the expansion strength, the defense signal, and the defense strength into the pulse decision engine, and maps them into platform executable decision instructions through a pre-set rule base.

[0012] Preferably, in the multi-modal data analysis module, the extraction step of the product attribute vector comprises:

[0013] The product function dimension is initialized by the product function aggregator of all dimensions based on the pre-constructed product function keyword tree;

[0014] Based on the pre-constructed product function keyword dictionary, the product function keywords of the current text in the social media text stream are matched, and the current text is assigned to the corresponding product function aggregator based on the matched product function keywords;

[0015] The sentiment words, degree adverbs, and negative words associated with the product function keywords in the current text in the product function aggregator are identified to construct a triple;

[0016] The function attribute strength is calculated based on the triple, and the average function attribute strength value is obtained by weighted average of all function attribute strengths in the same product function aggregator;

[0017] The average function attribute strength values in different product function aggregators are combined into a product attribute vector according to the arrangement order of the product function dimensions.

[0018] Preferably, in the multi-modal data analysis module, before extracting the product attribute vector, the social media text stream is forcibly converted from traditional Chinese to simplified Chinese.

[0019] Preferably, in the multi-modal data analysis module, the extraction step of the quality evaluation vector and the quality fluctuation value comprises:

[0020] The ad click-through rate, the video three-second skip rate, and the complete play rate of each short video advertisement in the short video advertisement stream are extracted;

[0021] The advertisement quality score is calculated based on the ad click-through rate, the video three-second skip rate, and the complete play rate;

[0022] The advertisement quality scores of all short video advertisements are weighted and fused to obtain an advertisement quality evaluation total score, and meanwhile, a quality fluctuation value is calculated based on the maximum value and the minimum value in the advertisement quality scores of all short video advertisements, and the advertisement quality evaluation total score is converted into a single-element quality evaluation vector.

[0023] Preferably, in the multi-modal data analysis module, the extraction steps of the decision rhythm vector and the transaction time data set comprise:

[0024] The transaction decision time intervals of each user in the transaction pulse stream are calculated, and the transaction decision type of each user is determined according to the preset decision time threshold of different industries, wherein the transaction decision type includes impulsive type, rational type and long cycle type;

[0025] The transaction decision time intervals of the users of the impulsive type are combined into an impulsive transaction time interval table, the transaction decision time intervals of the users of the rational type are combined into a rational transaction time interval table, and the transaction decision time intervals of the users of the long cycle type are combined into a long cycle transaction time interval table;

[0026] The proportion of the number of each transaction decision type is calculated, and the calculated proportion of the number of the impulsive type, the proportion of the number of the rational type and the proportion of the number of the long cycle type are sequentially combined into a decision rhythm vector output;

[0027] The impulsive transaction time interval table, the rational transaction time interval table and the long cycle transaction time interval table are combined into a transaction time data set output.

[0028] Preferably, in the feature signal matrix generation module, the calculation steps of the main feature value and the secondary feature value comprise:

[0029] The covariance matrix of the current feature signal matrix is calculated;

[0030] The covariance matrix is subjected to eigenvalue decomposition to obtain a plurality of eigenvalues, the eigenvalue with the largest absolute value is taken as the main feature value, and the eigenvalue with the smallest absolute value is taken as the secondary feature value.

[0031] Preferably, in the double-threshold generation module, the calculation step of the expansion threshold value comprises:

[0032] The impulsive transaction time interval table in the transaction time data set is extracted, and the arithmetic average time efficiency is calculated;

[0033] Based on the product attribute vector, the maximum functional attribute intensity value in the product attribute vector is obtained;

[0034] The arithmetic average time efficiency, the maximum functional attribute intensity value and the preset industry expansion coefficient are multiplied to obtain the expansion threshold value.

[0035] Preferably, in the dual-threshold generation module, the calculation step of the defense threshold comprises:

[0036] calculating the eigenvalue variance of the current feature signal matrix;

[0037] calculating the historical sub-eigenvalue based on the historical feature signal matrix, and counting the number of times that the historical sub-eigenvalue is less than the preset sub-eigenvalue threshold to obtain a risk number value, and taking the reciprocal of the risk number value as a historical risk index;

[0038] multiplying the reciprocal of the quality fluctuation value by the historical risk index, and then dividing by the eigenvalue variance to obtain the defense threshold.

[0039] Preferably, in the decision signal generation module, the generation conditions of the expansion signal and the defense signal comprise:

[0040] generating the expansion signal when the main eigenvalue is greater than the expansion threshold, and generating the defense signal when the sub-eigenvalue is less than the defense threshold.

[0041] Preferably, in the decision instruction generation module, it further comprises: when the defense intensity is greater than 1.5 times of the expansion intensity and less than 2 times of the expansion intensity, preferentially executing the defense instruction; and when the defense intensity is greater than or equal to 2 times of the expansion intensity, delaying the execution of the defense instruction.

[0042] Compared with the related art, the market marketing intelligent decision system based on multi-modal learning provided by the present application has the following beneficial effects:

[0043] The application realizes the synchronous access and deep mining of three-source heterogeneous data of social media text stream, short video advertisement stream and transaction pulse stream by constructing a multi-modal data analysis module, accurately extracts attribute vectors representing product function dimensions from social media text stream, separates quality evaluation vectors and fluctuation value double indicators from short video advertisement stream, and analyzes decision rhythm vectors and time data sets from transaction pulse stream, breaking through the limitations of traditional marketing system single-modal analysis; the cross-modal feature fusion module adopts three algebraic interaction mechanisms of element-by-element multiplication, dot product operation and vector module multiplication to generate three fusion features of attribute emergency index, advertisement timeliness coefficient and strategy matching degree, realizing the dynamic coupling of text sentiment, advertisement effect and transaction behavior; the feature signal matrix generation module combines multi-source features into a signal matrix and extracts main and secondary feature values through feature decomposition, accurately quantifying market expansion momentum and system vulnerability; the double threshold generation module dynamically calculates expansion threshold based on product attribute vectors and transaction time data sets, generates defense threshold combining quality fluctuation value, real-time signal matrix and historical matrix, forming a risk control boundary with adaptive industry characteristics; the decision signal generation module generates expansion and defense signals with intensity by comparing feature values and threshold values, and the pulse decision engine maps signal intensity to executable instructions through a rule base, finally realizing marketing decision response while solving the decision lag problem caused by multi-modal data fragmentation, improving market opportunity capture rate and providing decision support for cross-industry multi-category marketing scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 Figure 1 is a module structure diagram of a marketing intelligent decision system based on multi-modal learning according to the present application. DETAILED DESCRIPTION

[0045] The application will be described in further detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the application, and not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings, not all structures. Furthermore, the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0046] In addition, it should be noted that only the parts related to the application are shown in the drawings, not all contents. Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowchart describes each operation (or step) as a sequential process, many of the operations can be implemented in parallel, concurrently or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but can also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0047] Embodiment one

[0048] A multi-modal learning-based marketing intelligent decision system, in the specific implementation process, as shown in Figure 1 The multi-modal learning-based marketing intelligent decision system module structure diagram is shown, which includes:

[0049] Multi-modal data analysis module 100: simultaneously access social media text stream, short video advertising stream and transaction pulse stream, and extract product attribute vector from social media text stream, quality evaluation vector and quality fluctuation value from short video advertising stream, and decision rhythm vector and transaction time data set from transaction pulse stream.

[0050] Specifically, in the multi-modal data analysis module 100, the extraction step of the product attribute vector includes:

[0051] Determine the product function dimension initialization of all dimensions of the product function aggregator based on the pre-constructed product function keyword tree;

[0052] Based on the pre-constructed product function keyword dictionary, match the product function keywords of the current text in the social media text stream, and assign the current text to the corresponding product function aggregator based on the matched product function keywords;

[0053] Identify the sentiment words, degree adverbs and negative words associated with the product function keywords in the current text in the product function aggregator to construct a triple;

[0054] Calculate the function attribute intensity based on the triple, and obtain the average function attribute intensity value by weighted average of all function attribute intensities in the same product function aggregator;

[0055] Combine the average function attribute intensity values in different product function aggregators into a product attribute vector according to the arrangement order of the product function dimension.

[0056] Specifically, in the multi-modal data analysis module 100, it also includes forced conversion of traditional Chinese to simplified Chinese for the social media text stream before extracting the product attribute vector.

[0057] Specifically, in the multi-modal data analysis module 100, the extraction step of the quality evaluation vector and the quality fluctuation value includes:

[0058] Extract the advertising click-through rate, video three-second skip rate and complete play rate of each short video advertisement in the short video advertising stream;

[0059] Calculate the advertising quality score based on the advertising click-through rate, video three-second skip rate and complete play rate;

[0060] The advertisement quality scores of all short video advertisements are weighted and fused to obtain an advertisement quality evaluation total score, and meanwhile, a quality fluctuation value is calculated based on the maximum value and the minimum value in the advertisement quality scores of all short video advertisements, and the advertisement quality evaluation total score is converted into a single-element quality evaluation vector.

[0061] Specifically, in the multi-modal data analysis module 100, the extraction steps of the decision rhythm vector and the transaction time data set include:

[0062] The transaction decision time interval of each user in the transaction pulse stream is calculated, and the transaction decision type of each user is determined according to the preset decision time threshold of different industries, wherein the transaction decision type includes impulsive type, rational type and long cycle type;

[0063] The transaction decision time intervals of the users of the impulsive type are combined into an impulsive transaction time interval table, the transaction decision time intervals of the users of the rational type are combined into a rational transaction time interval table, and the transaction decision time intervals of the users of the long cycle type are combined into a long cycle transaction time interval table;

[0064] The number proportion of each transaction decision type is calculated, and the calculated impulsive number proportion, rational number proportion and long cycle number proportion are sequentially combined into a decision rhythm vector and output;

[0065] The impulsive transaction time interval table, the rational transaction time interval table and the long cycle transaction time interval table are combined into a transaction time data set and output.

[0066] In the implementation process, in the multi-modal data analysis module 100 of the marketing intelligent decision system, first, synchronize access to three types of heterogeneous data sources: social media text stream, short video advertisement stream, and transaction pulse stream. For the social media text stream, initialize the aggregator of the product function dimension based on the pre-built product function keyword tree. Match the keywords in the text with the product function keyword dictionary and assign them to the corresponding aggregator. The product function aggregator is a data structure allocated in memory for each predefined product function dimension. Examples of predefined product function dimensions include but are not limited to "battery life", "camera function", and "screen display". These dimensions are used to aggregate and calculate user feedback intensity in real time. The product function dimension is determined by the pre-built product function keyword tree, which defines the core function classification of the product and its hierarchical relationship. Based on the pre-built product function keyword dictionary, which is derived from the leaf nodes of the product function keyword tree, match the product function keywords contained in the current text in the social media text stream in real time. According to the matching result, assign the current text to its corresponding product function aggregator. In the product function aggregator, identify the sentiment words associated with the product function keywords for the current text, including but not limited to "good", "bad", degree adverbs, including but not limited to "very", "comparatively", and negative words, including but not limited to "not" and "no". Build a triple of sentiment words, degree adverbs, and negative words, such as "waterproof, very, no negative". Calculate the sentiment intensity value of the triple. The weighting coefficients of degree adverbs are 1.3 for non, 0.7 for comparative, and negative words are negative. Weighted average the sentiment intensity values of all texts in the same aggregator according to time decay, such as 1.5 for 24 hours, 1.2 for 6 hours, and 0.8 for more than 6 hours. Obtain the average intensity value of each function dimension and combine it into a product attribute vector in dimension order. At the same time, perform mandatory conversion from traditional Chinese to simplified Chinese to eliminate dialect interference. For the short video advertisement stream, extract the click-through rate, three-second skip rate, and complete playback rate of each advertisement in real time. Weighted fusion is performed through pre-set weight values to obtain the advertisement quality score.2 Obtain the advertisement quality score of a single advertisement, weight and integrate the advertisement quality scores of all advertisements to obtain an advertisement quality evaluation total score, encapsulate the scalar value of the advertisement quality evaluation total score as a 1-dimensional single-element quality evaluation vector, and calculate a quality fluctuation value according to the difference between the maximum and minimum values of the score; for the transaction pulse stream, calculate the difference between the product first contact timestamp and the transaction payment timestamp as the decision-making time interval, classify according to industry thresholds, for example, less than 2 hours for the beauty category is impulsive, 2 to 6 hours is rational, and more than 6 hours is long cycle, when the interval is equal to the threshold, it is classified as long cycle, and build an impulsive, rational, and long cycle transaction time interval table to form a transaction time dataset, and calculate the impulsive, rational, and long cycle person proportion, respectively, and combine the three scalar values into a decision rhythm vector in order to output, which directly reflects the decision type distribution of the current user group, and is used to represent the overall decision rhythm of the market, for example, a high proportion of impulsive type means a large corresponding value of the vector, indicating that the market decision rhythm is fast.

[0067] The cross-modal feature fusion module 200: based on the product attribute vector and the decision rhythm vector, the attribute emergency index is calculated by element-by-element multiplication, based on the quality evaluation vector and the decision rhythm vector, the advertisement timeliness coefficient is generated by dot product operation, and based on the product attribute vector and the quality evaluation vector, the strategy matching degree is generated by vector module multiplication.

[0068] In the specific implementation process, first, the product attribute vector, the quality evaluation vector, and the decision rhythm vector are subjected to format standardization processing, since the dimensions of the three are different, they need to be converted into a format that can be subjected to algebraic operation, specifically, the single-element quality evaluation vector is expanded into a vector with the same dimension as the product attribute vector, and each element value of the vector is the advertisement quality evaluation total score; the decision rhythm vector is regarded as a 3-dimensional vector.

[0069] Then a triple fusion operation is performed, including attribute emergency index calculation, multiplying the product attribute vector with the "impulsive proportion" component in the decision rhythm vector element by element, and the multiplication result is a new vector, and each element value of the new vector is the product of the function attribute heat and the impulsive user proportion. The multiplication result quantifies the degree to which the positive market feedback for the product function comes from the user group with the fastest decision rhythm and the most impulsive, and the higher the value, the more urgent the market response triggered by the function attribute. The advertisement timeliness coefficient calculation is to perform a dot product operation on the quality evaluation vector and the "rational proportion" component in the decision rhythm vector, and the dot product result is a scalar, and the value of the scalar is the product of the advertisement quality score and the rational proportion. The value of the scalar represents the degree to which high-quality advertisement content matches the user group that is cautious in decision-making and needs rational persuasion, and a high coefficient indicates that the advertisement is timely and effective in influencing rational users. The strategic matching degree calculation calculates the product of the module length of the product attribute vector and the module length of the quality evaluation vector, and the product is a scalar, which measures the coordination level between the "overall market appeal of the product" and the "overall quality of the advertisement". A high matching degree indicates that the product force and the advertisement effect are synchronized and good. Finally, the output results are subjected to dynamic constraints: the attribute emergency index value is truncated to the [-5, 5] interval to prevent overflow, the advertisement timeliness coefficient is compressed to the (-1, 1) range through the tanh function to ensure comparability, and the strategic matching degree retains the original value but marks the threshold value event, that is, when the strategic matching degree exceeds the preset matching degree threshold, an alarm is triggered. All operation results are written into the fusion log in real time for audit tracing and provide data basis for subsequent modules.

[0070] The feature signal matrix generation module 300 combines the attribute emergency index, the advertisement timeliness coefficient, and the strategic matching degree into the current feature signal matrix, and extracts the main eigenvalue and the secondary eigenvalue of the current feature signal matrix.

[0071] Specifically, in the feature signal matrix generation module 300, the calculation steps of the main eigenvalue and the secondary eigenvalue include:

[0072] Calculate the covariance matrix of the current feature signal matrix.

[0073] Perform eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues, take the eigenvalue with the largest absolute value as the main eigenvalue, and take the eigenvalue with the smallest absolute value as the secondary eigenvalue.

[0074] In the implementation process, firstly, three key indicator attributes emergency index output from the cross-modal feature fusion module 200 are received, reflecting the synergistic emergency degree of product attributes and user decision rhythm, the advertisement timeliness coefficient representing the dynamic correlation of advertisement quality and decision timeliness, and the strategic matching degree measuring the consistency of product attributes and advertisement quality strategy. The three scalar indicators are combined into the current feature signal matrix in a fixed order; then the covariance matrix of the matrix is calculated to capture the fluctuation correlation between indicators; then the covariance matrix is decomposed to obtain all eigenvalues, the core is to apply the principal component analysis, a well-known theory in the field of data analysis and machine learning. The largest eigenvalue of the covariance matrix, i.e. the principal eigenvalue, represents the direction of the largest variance in the multi-modal fusion feature signal, and its numerical size quantifies the trend and intensity of the consistency of the market in all dimensions. This conforms to the market trend consistency principle and opportunity window theory in marketing, i.e. when multiple key indicators simultaneously and dramatically point to a positive direction, it indicates that a high- confidence market expansion opportunity is emerging, so it is explained as the system expansion motive strength. Conversely, the smallest eigenvalue, i.e. the secondary eigenvalue, represents the direction of the smallest variance or noise in the signal, and its small numerical value indicates that the system has weak but critical instability or vulnerability in a certain dimension, which conforms to the system risk management theory and the bucket effect, i.e. the overall risk of the system is often indicated by the weakest link or the abnormal signal that has not been fully captured, so it is explained as the system vulnerability. Through this PCA-based interpretation framework, the principal and secondary eigenvalues successfully convert mathematical abstraction into decision indicators with clear marketing implications and are output in real time to the downstream module.

[0075] The double-threshold generation module 400 calculates the expansion threshold based on the product attribute vector and the transaction time data set, and calculates the defense threshold based on the quality fluctuation value, the current feature signal matrix and the historical feature signal matrix.

[0076] Specifically, in the double-threshold generation module 400, the calculation steps of the expansion threshold include:

[0077] Extract the impulsive transaction time interval table in the transaction time data set and calculate the arithmetic mean time efficiency;

[0078] Based on the product attribute vector, the maximum functional attribute intensity value in the product attribute vector is obtained;

[0079] The arithmetic mean time efficiency, the maximum functional attribute intensity value and the preset industry expansion coefficient are multiplied to obtain the expansion threshold.

[0080] Specifically, in the double-threshold generation module 400, the calculation steps of the defense threshold include:

[0081] Calculate the eigenvalue variance of the current feature signal matrix;

[0082] calculate a historical sub-eigenvalue based on the historical characteristic signal matrix, and count a number of times that the historical sub-eigenvalue is less than a preset sub-eigenvalue threshold to obtain a risk number value, and take a reciprocal of the risk number value as a historical risk index;

[0083] multiply the reciprocal of the quality fluctuation value by the historical risk index, divide the result by the eigenvalue variance to obtain a defense threshold.

[0084] In the implementation process, first extract the impulsive transaction time interval table in the transaction time data set, and calculate its reciprocal average value as the arithmetic average time efficiency, that is, the number of transactions per unit time, the higher the efficiency, the higher the value. This value reflects the conversion speed of the most active user group in the market, and the maximum functional attribute intensity value is obtained from the product attribute vector, which represents the current most popular product function point that can attract users. Multiply the arithmetic average time efficiency, the maximum functional attribute intensity value and the preset industry expansion coefficient to obtain the expansion threshold, wherein the industry expansion coefficient is an empirical constant based on historical data or industry benchmark to adjust the threshold sensitivity of different industries. For example, according to the product category, for example, take 0.85 for the beauty category and 0.78 for the 3C category. The expansion threshold aims to define a critical point of market opportunity. This calculation is based on a reasonable marketing assumption: when (market conversion efficiency) × (product maximum attraction) × (industry adjustment factor) reaches a certain level, it is considered that there is a clear expansion opportunity. The product form comprehensively considers market speed, product power and industry characteristics. The higher the value, the stronger the required expansion evidence. Here, the arithmetic average time efficiency, the maximum functional attribute intensity value and the preset industry expansion coefficient correspond to the market conversion efficiency, the product maximum attraction and the industry adjustment factor respectively. Then calculate the eigenvalue variance of the current feature signal matrix, and based on the historical feature signal matrix, count the number of times that the historical sub-eigenvalue is less than the preset sub-eigenvalue threshold, for example, the sub-eigenvalue threshold is 0.3. Take the reciprocal of this number as the historical risk index. Finally, multiply the reciprocal of the quality fluctuation value, that is, the higher the advertisement quality stability, the higher the value, by the historical risk index, and then divide it by the eigenvalue variance to obtain the defense threshold. When the advertisement quality fluctuates or the historical risk frequency increases, the defense threshold decreases, and when the system state fluctuates violently, the defense threshold increases. The core of the defense threshold is to build a risk perception system that can dynamically evolve with the market environment. The reciprocal of the quality fluctuation value reflects the brand asset theory and the signal theory. High and stable advertisement quality is a strong signal of trust that the brand transmits to the market. The reciprocal directly quantifies the accumulation degree of this trust capital. The higher the stability, the more "trust reserves" it has to resist risks. The historical risk index is rooted in the organizational learning theory and the crisis warning theory. The system learns the risk patterns of the market by digging into the frequency of system vulnerability outbreaks in historical data. The higher the frequency, the lower the historical risk index, indicating that the system is in a risk environment that must be continuously learned. A more sensitive defense threshold must be set. The eigenvalue variance reflects the market effectiveness theory. A high variance indicates that the market state is complex and variable, and the consensus is low. The signal is chaotic. At this time, the defense threshold should be increased to avoid overreaction to noise. Conversely, a low variance indicates that the market state is clear and stable. At this time, the defense threshold should be reduced to capture subtle risks.Finally, by multiplying the inverse of the quality fluctuation value representing the "trust reserve" with the historical risk index representing the "historical lesson", and dividing the product by the eigenvalue variance representing the "market chaos degree", the resulting defense threshold is a highly adaptive indicator that conforms to the concept of anti-fragility, enabling the system not only to resist risks, but also to learn and evolve from market fluctuations and historical experiences, thus achieving intelligent risk control.

[0085] The decision signal generation module 500 generates an expansion signal and an expansion strength based on the main eigenvalue and the expansion threshold, and generates a defense signal and a defense strength based on the secondary eigenvalue and the defense threshold.

[0086] Specifically, in the decision signal generation module 500, the generation conditions of the expansion signal and the defense signal include:

[0087] When the main eigenvalue is greater than the expansion threshold, the expansion signal is generated, and when the secondary eigenvalue is less than the defense threshold, the defense signal is generated.

[0088] In the specific implementation process, the main eigenvalue and the secondary eigenvalue from the feature signal matrix generation module 300 are received in real time, and the expansion threshold and the defense threshold calculated by the double threshold generation module 400 are obtained; when the main eigenvalue exceeds the expansion threshold, a binary expansion signal is immediately generated, and the expansion strength is calculated, the calculation formula is the main eigenvalue divided by the expansion threshold, the result is limited in the interval [0, 5] to reflect the degree of opportunity exceeding the safety boundary, when the secondary eigenvalue is lower than the defense threshold, a binary defense signal is generated and the defense strength is calculated, the calculation formula is the secondary eigenvalue divided by the defense threshold, after logarithmic compression, it is mapped to the interval [0, 10] to quantify the risk level; finally, the signal with the expansion signal and the expansion strength and the defense signal and the defense strength is output for use by the downstream decision instruction generation module 600.

[0089] The decision instruction generation module 600 inputs the expansion signal, the expansion strength, and the defense signal, the defense strength into the pulse decision engine, and maps them to platform executable decision instructions through a pre-set rule base.

[0090] Specifically, in the decision instruction generation module 600, when the defense strength is greater than 1.5 times the expansion strength and less than 2 times the expansion strength, the defense instruction is preferentially executed, and when the defense strength is greater than or equal to 2 times the expansion strength, the defense instruction is delayed.

[0091] In the implementation process, the decision instruction generation module 600 converts the input expansion signal, expansion strength, defense signal and defense strength into platform executable decision instructions through the pulse decision engine, the core pulse decision engine of which adopts a three-layer decision tree structure: the first layer is a signal priority arbitrator, according to the preset priority decision rule, that is, when the defense strength is greater than 1.5 times of the expansion strength and less than 2 times of the expansion strength, the defense instruction is triggered to execute the channel preferentially, when the defense strength is greater than or equal to 2 times of the expansion strength, the delayed execution queue is activated, for example, a 30-minute countdown is started and the defense strength is recalculated every 2 minutes, if the defense strength deteriorates by more than 10% during the period, the defense is executed immediately; the second layer is a rule matching matrix, which maps the signal combination to 132 preset scenes, for example, the expansion signal plus the expansion strength plus the defense signal plus the defense strength is mapped to scene 47, and dynamic compilation is realized through a decision tree with weighted labels, for example, when scene 47 is matched, the instruction of “reduce high-priced advertisement budget by 20% and start medium-risk area promotion” is generated; the third layer is an instruction executor, which converts the abstract instruction into a specific platform API call, and finally realizes the intelligent decision of market marketing based on multi-modal learning,

[0092] The working principle of the market marketing intelligent decision system based on multi-modal learning provided by the application is as follows:

[0093] Firstly, the multi-modal data analysis module 100 processes three source heterogeneous data in real time: product attribute vectors are extracted from social media text streams; single element quality evaluation vectors and quality fluctuation values reflecting the stability fluctuation of the advertisement are generated from short video advertisement streams; decision rhythm vectors and transaction time data sets are output from transaction pulse streams. Then, the cross-modal feature fusion module 200 realizes feature interaction through differential algebraic operation: the attribute emergency index is generated by element-by-element multiplication of the product attribute vector and the decision rhythm vector, the advertisement timeliness coefficient is output by the dot product operation of the quality evaluation vector and the decision rhythm vector, and the strategy matching degree is generated by the vector module multiplication of the product attribute vector and the quality evaluation vector. After the three are combined into a feature signal matrix by the feature signal matrix generation module 300, the main and secondary feature values are extracted by feature decomposition. The expansion threshold and the defense threshold are calculated in parallel by the double threshold generation module 400. Then, the decision signal generation module 500 generates expansion signals and defense signals with strength by comparing the relative relationship between the feature values and the thresholds, and finally the decision instruction generation module 600 maps the executable marketing action instructions based on the preset rule library through the pulse decision engine, and realizes instruction conflict arbitration according to the ratio of the defense strength to the expansion strength, forming a closed-loop intelligent decision system from multi-source perception, multi-modal learning, feature modulation, threshold decision to instruction generation threshold arbitration.

[0094] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the working examples illustrate the architecture, functionality, and operation of possible implementations of apparatuses (systems), methods and computer program products according to the present disclosure. In this regard, each flowchart and / or block diagram can represent a method, apparatus and / or computer program products according to the present disclosure. It will also be understood that each block of the flowchart and / or block diagrams and combinations of blocks in the flowchart and / or block diagrams can be implemented by computer program instructions such as program code. Figure 1 The flowchart and / or block diagram in the working examples illustrate the architecture, functionality, and operation of possible implementations of apparatuses (systems), methods and computer program products according to the present disclosure. In this regard, each flowchart and / or block diagram can represent a method, apparatus and / or computer program products according to the present disclosure. It will also be understood that each block of the flowchart and / or block diagrams and combinations of blocks in the flowchart and / or block diagrams can be implemented by computer program instructions such as program code.

[0095] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer readable storage medium, including Read-Only Memory (ROM), Random Access Memory (RAM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), One-time Programmable Read-Only Memory (OTPROM), Electrically-Erasable Programmable Read-Only Memory (EEPROM), Compact Disc Read-Only Memory (CD-ROM) or other optical disk memories, magnetic disk memories, magnetic tape memories, or any other medium readable by a computer that can be used to carry or store data.

[0096] It should also be noted that the terms "comprising" "containing" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without further limitation, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

Claims

1. A multi-modal learning based marketing intelligence decision system, characterized in that, The marketing intelligent decision system comprises the following modules: A multi-modal data analysis module: simultaneously access social media text stream, short video advertisement stream and transaction pulse stream, and extract product attribute vector from social media text stream, extract quality evaluation vector and quality fluctuation value from short video advertisement stream, and extract decision rhythm vector and transaction time data set from transaction pulse stream; A cross-modal feature fusion module: based on product attribute vector and decision rhythm vector, calculate attribute emergency index through element-by-element multiplication, based on quality evaluation vector and decision rhythm vector, generate advertisement timeliness coefficient through dot product operation, and based on product attribute vector and quality evaluation vector, generate strategy matching degree through vector module multiplication; A feature signal matrix generation module: combine attribute emergency index, advertisement timeliness coefficient and strategy matching degree into a current feature signal matrix, and extract principal eigenvalue and secondary eigenvalue of the current feature signal matrix; A double-threshold generation module: calculate expansion threshold based on product attribute vector and transaction time data set, and calculate defense threshold based on quality fluctuation value, current feature signal matrix and historical feature signal matrix; A decision signal generation module: generate expansion signal and expansion strength based on principal eigenvalue and expansion threshold, and generate defense signal and defense strength based on secondary eigenvalue and defense threshold; A decision instruction generation module: input expansion signal, expansion strength, defense signal and defense strength into pulse decision engine, and map them into platform executable decision instructions through a pre-set rule base.

2. The multi-modal learning based marketing intelligent decision system according to claim 1, wherein, In the multi-modal data analysis module, the extraction of the product attribute vector comprises: Initialize product function aggregators of all dimensions based on product function dimension determined by pre-constructed product function keyword tree; Match product function keywords of the current text in the social media text stream based on pre-constructed product function keyword dictionary, and assign the current text to the corresponding product function aggregator based on the matched product function keywords; Identify sentiment words, degree adverbs and negative words associated with product function keywords in the current text in the product function aggregator to construct triplets; Calculate function attribute intensity based on the triplets, and obtain average function attribute intensity value by weighted average of all function attribute intensity values in the same product function aggregator; Combine the average function attribute intensity values in different product function aggregators into a product attribute vector according to the arrangement order of product function dimensions.

3. The multi-modal learning based marketing intelligence decision system of claim 2, wherein, In the multi-modal data analysis module, it further comprises forcibly converting traditional Chinese in the social media text stream into simplified Chinese before extracting the product attribute vector.

4. The multi-modal learning based marketing intelligence decision system of claim 3, wherein, In the multi-modal data analysis module, the extraction of the quality evaluation vector and the quality fluctuation value comprises: Extract advertisement click-through rate, video first three seconds skip rate and complete play rate of each short video advertisement in the short video advertisement stream; Calculate advertisement quality score based on advertisement click-through rate, video first three seconds skip rate and complete play rate; Weightedly fuse advertisement quality scores of all short video advertisements to obtain advertisement quality evaluation total score, simultaneously calculate quality fluctuation value based on the maximum value and the minimum value in the advertisement quality scores of all short video advertisements, and convert the advertisement quality evaluation total score into a single-element quality evaluation vector.

5. The multi-modal learning based marketing intelligence decision system of claim 4, wherein, In the multi-modal data analysis module, the extraction step of the decision rhythm vector and the transaction time data set comprises: calculating the transaction decision time interval of each user in the transaction pulse stream, and determining the transaction decision type of each user according to the preset decision time threshold of different industries, wherein the transaction decision type includes impulsive, rational and long cycle; combining the transaction decision time interval of the impulsive user into an impulsive transaction time interval table, combining the transaction decision time interval of the rational user into a rational transaction time interval table, and combining the transaction decision time interval of the long cycle user into a long cycle transaction time interval table; calculating the proportion of the number of each transaction decision type, and combining the calculated impulsive proportion of the number, rational proportion of the number, and long cycle proportion of the number into a decision rhythm vector in order to output; combining the impulsive transaction time interval table, the rational transaction time interval table and the long cycle transaction time interval table into a transaction time data set output.

6. The multi-modal learning based marketing intelligence decision system of claim 5, wherein, In the feature signal matrix generation module, the calculation step of the main feature value and the secondary feature value comprises: calculating the covariance matrix of the current feature signal matrix; performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues, taking the eigenvalue with the largest absolute value as the main feature value, and taking the eigenvalue with the smallest absolute value as the secondary feature value.

7. The multi-modal learning based marketing intelligence decision system of claim 6, wherein, In the double threshold generation module, the calculation step of the expansion threshold comprises: extracting the impulsive transaction time interval table in the transaction time data set and calculating the arithmetic average time efficiency; based on the product attribute vector, obtaining the maximum functional attribute intensity value in the product attribute vector; multiplying the arithmetic average time efficiency, the maximum functional attribute intensity value and the preset industry expansion coefficient to obtain the expansion threshold.

8. The multi-modal learning based marketing intelligence decision system of claim 7, wherein, In the double threshold generation module, the calculation step of the defense threshold comprises: calculating the eigenvalue variance of the current feature signal matrix; based on the historical feature signal matrix, calculating the historical secondary eigenvalue, and counting the number of times that the historical secondary eigenvalue is less than the preset secondary eigenvalue threshold to obtain a risk number value, and taking the reciprocal of the risk number value as a historical risk index; multiplying the reciprocal of the quality fluctuation value by the historical risk index and dividing by the eigenvalue variance to obtain the defense threshold.

9. The multi-modal learning based marketing intelligence decision system of claim 8, wherein, In the decision signal generation module, the generation conditions of the expansion signal and the defense signal comprise: generating the expansion signal when the main feature value is greater than the expansion threshold, and generating the defense signal when the secondary feature value is less than the defense threshold.

10. The multi-modal learning based marketing intelligence decision system of claim 9, wherein, In the decision instruction generation module, it further comprises that when the defense intensity is greater than 1.5 times of the expansion intensity and less than 2 times of the expansion intensity, the defense instruction is preferentially executed, and when the defense intensity is greater than or equal to 2 times of the expansion intensity, the defense instruction is delayed.

Images