An advertisement pushing method based on big data analysis

Abstract

The application relates to the technical field of advertisement pushing, and specifically discloses an advertisement pushing method based on big data analysis, which comprises the following steps: calculating an attention concentration degree and an acoustic broadcast suitability level based on multidimensional data; obtaining the presentation attribute of each piece of advertisement content in a to-be-pushed advertisement material library, wherein the presentation attribute comprises a classification identifier and an associated score corresponding to different classification identifiers; filtering all the advertisement contents in the to-be-pushed advertisement material library according to the attention concentration degree and the acoustic broadcast suitability level, and generating a candidate advertisement set; calculating the priority score of the advertisement content based on a user interest preference model and the presentation attribute of each piece of advertisement content in the candidate advertisement set; and pushing the advertisement contents with high ranking to terminal equipment according to the order from high to low of the priority score. The application improves the overall effective reach rate and resource utilization efficiency of advertisement pushing.

Description

Technical Field

[0001] This invention relates to the field of advertising push technology, and more specifically to an advertising push method based on big data analysis. Background Technology

[0002] Existing ad delivery methods primarily rely on the analysis and modeling of users' historical behavioral data, such as click-through rates, browsing history, search keywords, and long-term interests. These methods match and rank ads based on users' past static interest profiles, but generally lack consideration for the real-time physical environment and the user's perceived state at the time of ad reception.

[0003] Current technologies cannot detect whether a user is in a noisy environment, is moving, or has their visual attention focused on the device screen. This results in the pushed advertising content being mismatched in form (such as relying heavily on audio or visuals) with the user's real-time reception capabilities. For example, pushing ads that are primarily based on voice information in noisy public places, or pushing ads that rely on complex visual presentations when the user's gaze has left the screen, wastes advertising exposure resources and reduces the actual dissemination effect. Summary of the Invention

[0004] The purpose of this invention is to provide an advertising push method based on big data analysis to solve the above-mentioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for pushing advertisements based on big data analytics includes the following steps: The system calculates the user's attention level on the terminal device screen and the appropriateness level of the acoustic broadcast based on multi-dimensional data from the user's terminal device. Obtain the presentation attributes of each advertisement in the advertising material library to be pushed, wherein the presentation attributes include the category identifier and the association score corresponding to different category identifiers; Based on the attention concentration level and the acoustic broadcast suitability level, all advertising content in the advertising material library to be pushed is filtered to generate a candidate advertising set; Based on the user interest preference model and the presentation attributes of each advertisement in the candidate advertisement set, the priority score of the advertisement content is calculated. Based on the priority scores from high to low, a number of advertisements are pushed to the terminal device.

[0006] As a further aspect of the present invention, the process of acquiring multi-dimensional data is as follows: The original audio sampling stream from the built-in microphone of the terminal device is continuously read, and the original audio sampling stream is divided into preset time windows. The sound pressure level of the audio signal in each time window is calculated, and the sound pressure level values ​​arranged in chronological order constitute the ambient sound pressure level waveform. The screen image frame is obtained by capturing the image data of the screen display area through the screen content capture interface. The output signals from the triaxial accelerometer and triaxial gyroscope of the inertial measurement unit are collected synchronously. The output signals are aligned and synthesized according to the same time base to generate continuous data describing the translation and rotation motion of the device in three-dimensional space, thus forming the device motion sequence. Ambient sound pressure level waveforms, screen image frames, and device motion sequences constitute multi-dimensional data.

[0007] As a further aspect of the present invention: the process of calculating attention concentration is as follows: Perform continuous multi-frame analysis on the screen image frames, calculate the overall pixel change amplitude of the screen display area frame by frame, and obtain the pixel change amplitude sequence; The motion sequence of the device is segmented and analyzed using a sliding time window. Within each sliding time window, periodic micro-motion patterns are identified and their frequency and amplitude are calculated. The rate of change of frequency and the rate of change of amplitude are calculated based on the frequency and amplitude of the current and previous windows. The sum of the rate of change of frequency and the rate of change of amplitude is obtained by weighted summation. The result of the power operation with the natural constant as the base and the negative sum of the rate of change as the exponent is calculated. The result of the power operation is used as the micro-motion stability score of the current sliding time window. Generate a micro-motion stability scoring sequence according to the window order, and perform linear interpolation on the micro-motion stability scoring sequence to generate a micro-motion stability scoring interpolation sequence X that corresponds one-to-one with the time points of the pixel change amplitude sequence. Extract the pixel change amplitude A1 and micro-motion stability score A2 corresponding to the same time point from the time-aligned pixel change amplitude sequence X and micro-motion stability score sequence. Input A1 and A2 into a preset fusion function to output the attention intensity. The fusion function stipulates that the lower the pixel change amplitude and the higher the micro-motion stability score, the higher the output attention intensity. The attention concentration is obtained by calculating the time-weighted average of the attention intensity within the current calculation cycle and then normalizing it.

[0008] As a further aspect of the present invention: the process of calculating the suitability level for acoustic broadcasting is as follows: The ambient sound pressure level waveform is subjected to a fast Fourier transform to obtain the spectrum. Within the preset low-frequency to mid-frequency sub-band range, the average energy of the sub-band is calculated for multiple consecutive time windows. When the average energy of the sub-band continuously exceeds the first preset energy threshold, it is determined that there is a continuous background noise band. The number of times the instantaneous sound pressure level of the ambient sound pressure level waveform exceeds the second preset energy threshold within a unit time is counted. When the number exceeds the preset frequency threshold, it is determined that there is an intermittent sudden high loudness event. The terminal device's camera captures scene images, which are then input into a pre-trained lightweight convolutional neural network model. The model outputs the probability that the image belongs to each of the three preset categories: public place, transportation, and private space. The category with the highest probability is taken as the visual semantic analysis result. When it is determined that there is a continuous background noise band and the visual semantic analysis result indicates a public place, or when it is determined that there is an intermittent sudden high-volume event, the acoustic broadcast suitability level is marked as unsuitable. When it is determined that there is no continuous background noise band and no intermittent sudden high-loudness events, and the visual semantic analysis result is a private space, the acoustic broadcasting suitability level is marked as suitable; Except for the two cases mentioned above, the appropriateness level for acoustic broadcasting is marked as neutral.

[0009] As a further aspect of the present invention: the process of obtaining presentation attributes is as follows: The ad content package is parsed to obtain file data, which includes visual media data, audio media data, and interactive script data. The classification identifier of the ad content is determined based on the file data. The visual media data volume is obtained by multiplying the total number of pixels in the image sequence by the resolution and duration of the video track; the audio media data volume is quantized by multiplying the sampling rate and duration of the audio track; the interactive script data volume is quantized by the number of lines of code or the file size of the script file. When the amount of visual media data exceeds the first preset threshold, and the amount of visual media data exceeds the sum of the amount of audio media data and the amount of interactive script data, the advertising content will be marked as a visually dominant advertising. When the amount of audio media data exceeds the second preset threshold, and the amount of audio media data exceeds the sum of the amount of visual media data and the amount of interactive script data, the category identifier of the advertising content will be marked as an audio-dominated advertisement. When the amount of interactive script data exceeds the third preset threshold, and the amount of interactive script data exceeds the sum of the amount of visual media data and audio media data, the advertising content will be marked as an interactive-driven advertising. When any two of the three data volumes—visual media data volume, audio media data volume, and interactive script data volume—are greater than the fourth preset threshold, and the data volume of any one of them does not exceed the sum of the data volumes of the other two, the advertising content will be marked as a composite advertisement. If the ad content package does not meet any of the above judgment conditions, the ad content will be marked as other types of ads. For content categorized as visually dominant ads, a key visual complexity score is generated by calculating the image entropy value of its keyframes and the distribution dispersion of color and edge features. For content categorized as audio-driven advertising, an average speech intelligibility score is generated by calculating the signal-to-noise ratio and harmonic distortion of its audio track within the standard speech frequency range.

[0010] For content categorized as interactive ads, composite ads, or other types of ads, the key visual complexity score and average speech clarity score are recorded as preset default values.

[0011] As a further aspect of the present invention, the process of generating a candidate advertisement set is as follows: When the attention concentration exceeds the preset high attention threshold, content categorized as visually dominant, interactive, or hybrid ads will be included in the candidate ad set. When the preset low attention threshold is less than or equal to the preset high attention threshold, the following steps are performed: If the acoustic broadcast suitability level is suitable, then content categorized as audio-driven or interactive-driven ads will be included in the candidate ad set. If the acoustic broadcast suitability level is unsuitable or neutral, then content categorized as interactive-driven advertising will be included in the candidate advertising set. When the level of attention is less than the preset low attention threshold, perform the following steps: If the acoustic broadcast suitability level is suitable, then content categorized as audio-dominant ads and other types of ads will be included in the candidate ad set; If the acoustic broadcast suitability level is unsuitable or neutral, then content categorized as other types of advertising will be included in the candidate advertising set.

[0012] As a further aspect of the present invention: the process of calculating the priority score is as follows: The user interest preference model is obtained. The long-term user interest preference model is a classification model trained using a logistic regression algorithm based on the user's historical ad click logs and the content tags subscribed by the user. It can output an interest prediction score representing the probability that the user is interested in the input ad content. For each ad content in the candidate ad set, obtain the interest prediction score of the ad content; For content categorized as visually dominant, the interest prediction score of the ad content is added to the key visual complexity score of the ad content to obtain the base score; For content categorized as audio-driven ads, the interest prediction score of the ad content is added to the average speech clarity score of the ad content to obtain the base score; For content categorized as a composite advertisement, the base score is obtained by adding the interest prediction score of the advertisement content to the arithmetic mean of the key visual complexity score and the average speech intelligibility score of the advertisement content. Using the level of attention concentration as the first adjustment coefficient, a mapping relationship between the appropriate level of acoustic broadcasting and the second adjustment coefficient is established to obtain the current second adjustment coefficient; The weighted sum of the first adjustment coefficient and the second adjustment coefficient yields the comprehensive coefficient. The base score is then multiplied by the comprehensive coefficient to obtain the priority score.

[0013] The beneficial effects of this invention compared to the prior art are as follows: This invention achieves a leap from static interest matching to dynamic context adaptation in advertising push strategies by introducing multi-dimensional perception and calculation of the user's real-time physical environment and perceptual state. It comprehensively utilizes the microphone, camera, and inertial measurement unit of the terminal device to simultaneously capture ambient sound pressure level waveforms, scene images, and device motion sequences, constructing a real-time perception data layer. Based on this, by analyzing pixel changes in screen image frames and the stability of device motion, the user's real-time visual attention concentration is accurately calculated; simultaneously, by analyzing the environmental acoustic spectrum characteristics and scene image semantics, the suitability level of the current environment for audio playback is objectively determined. These two core real-time state indicators, combined with the presentation attributes of the advertising content itself based on the objective classification of media data volume (such as visual-dominated, audio-dominated), drive a dynamic filtering logic to ensure that the generated candidate ad set is highly matched in form (visual, audio, interactive, composite, etc.) to the user's current attention level and acoustic environment. Furthermore, by integrating a long-term user interest model to finely prioritize candidate content, this significantly reduces invalid exposure caused by mismatch between ad format and user's instantaneous reception capacity, thereby improving the overall effective reach and resource utilization efficiency of advertising push. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating an advertising push method based on big data analysis according to the present invention. Figure 2 This is a schematic diagram of the process for calculating priority scores in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-2 As shown, this invention is an advertising push method based on big data analysis, comprising the following steps: The system calculates the user's attention level on the terminal device screen and the appropriateness level of the acoustic broadcast based on multi-dimensional data from the user's terminal device. In a preferred embodiment of the present invention, the process of acquiring multi-dimensional data is as follows: The device's built-in microphone continuously captures ambient sound, generating a continuous stream of raw audio samples. This stream is the result of analog sound waves being digitized by an analog-to-digital converter. A common sampling rate is 16,000 Hz, meaning 16,000 digital samples of sound amplitude are acquired per second. The system divides this continuous sample stream into a series of consecutive segments using a fixed time length, such as 500 milliseconds, as the basic analysis window. For each audio sample within a 500-millisecond window, its sound pressure level (SPL) value is calculated. The calculation first squares the amplitude value of each sample, then calculates the average of all squared values ​​within the window, and finally takes the logarithm to base 10 and multiplies it by 10, resulting in a value in decibels (dB). This value represents the average loudness of the ambient sound during that time period. Arranging the decibel values ​​calculated for each time window in chronological order creates an ambient SPL waveform, which serves as the foundational data for subsequent analysis of ambient noise levels.

[0018] While acquiring sound data, the system accesses the graphics buffer interface provided by the operating system's underlying layer. With user authorization, it periodically captures the currently displayed content on the screen. The capture frequency can be set to 10 times per second, or one frame every 100 milliseconds. Each capture yields a two-dimensional array representing the pixel color information of the entire screen, called a screen image frame. Continuous image frames constitute a visual record of the content actually viewed by the user.

[0019] To understand the user's grip and movements, the system also simultaneously reads data from the device's built-in inertial measurement unit, which includes a three-axis accelerometer and a three-axis gyroscope. The accelerometer measures the device's linear acceleration in the forward, backward, left, right, and up / down directions, typically in meters per second squared; the gyroscope measures the device's angular velocity around these three axes, typically in radians per second. The system aligns these two sets of sensor signals using the same timing reference, such as obtaining a unified timestamp from the system clock. After alignment, the three acceleration values ​​and three angular velocity values ​​acquired at the same moment are combined to form a six-dimensional motion state vector. Over time, these time-sequential vectors form the device's motion sequence, precisely describing the device's movement and rotation in space.

[0020] The dataset comprises three parts: ambient sound pressure level waveform, a sequence of screen image frames arranged by timestamps, and a device motion sequence. These together form the original multi-dimensional data set characterizing the user's environment and interaction state. The microphone converts sound, a mechanical vibration, into electrical signals; the voltage strength directly corresponds to the sound pressure level, providing the physical basis for assessing whether the environment is noisy and suitable for playing audio advertisements. Capturing screen frames involves digitally sampling the signals emitted by the display screen, directly reflecting the content and changes in the user's visual attention, serving as a direct input for determining whether the user is focused on the screen. The inertial sensor, based on its microelectromechanical structure, senses the device's acceleration and rotation speed. Its readings are closely related to the stability of the user's handheld device, whether they are walking or riding in a vehicle, and can indirectly infer whether the user is in a stable state for receiving information. These three types of data, from acoustic, visual optics, and kinematics perspectives, capture the user's external environment and behavioral characteristics in real time and synchronously, providing an indispensable and objective perceptual foundation for quantifying the user's attention concentration and the acoustic suitability of the environment in subsequent steps.

[0021] In another preferred embodiment of the present invention, the process of calculating attention concentration is as follows: Analyzing continuously acquired screen image frames, the current frame and its preceding frame are taken as a set of inputs. The absolute values ​​of the differences between the corresponding pixels in the red, green, and blue color channels of the entire screen display area are calculated. These absolute values ​​are all added together to obtain a sum, which is defined as the overall pixel change amplitude of the frame. It quantifies the degree of visual change in screen content between adjacent frames. Taking a typical 10Hz sampling frequency as an example, 10 such amplitude values ​​are generated per second. Arranging these values ​​in chronological order constitutes the pixel change amplitude sequence.

[0022] Simultaneously, the motion sequence from the inertial measurement unit is processed, consisting of six-dimensional motion vectors arranged in time. A sliding time window of 2 seconds is set, sliding across the sequence in 0.1-second steps. Within each 2-second window, the vertical acceleration signal component of the aligned and synthesized motion sequence is extracted first as the primary analysis object. During processing, vertical acceleration data points of 2 seconds within the current sliding time window are taken. Assuming a sampling rate of 50 Hz, the window contains 100 consecutive sampling points. To reduce edge effects during spectral analysis, a Hanning window function is applied to these 100 data points for weighting. Subsequently, a Fast Fourier Transform is performed on the windowed data sequence to transform it from the time domain to the frequency domain, obtaining a spectrum containing the energy magnitudes of each frequency component. Since the frequencies generated by physiological micro-movements of the human hand are usually in a low and specific range, the analysis focuses on the frequency band between 0.5 Hz and 3 Hz (other values ​​are acceptable and not limited here), searching for the peak point with the highest spectral energy within this range. The frequency corresponding to the peak point, for example, 1.2 Hz, is initially identified as the dominant frequency of the potential periodic micro-motion pattern. To ensure that the peak is not random noise, the energy amplitude at the peak frequency is checked to see if it exceeds a threshold dynamically calculated based on the total energy of the window signal; for example, the peak energy must be greater than 5% of the total energy of the window signal. When the condition is met, the periodic micro-motion pattern is confirmed, and the peak frequency is officially recorded as the frequency of the window micro-motion pattern. The amplitude of the micro-motion pattern is characterized by calculating the root mean square value of the original vertical acceleration signal within the same 2-second window. This calculation process involves first calculating the square of the acceleration value at each sampling point, then calculating the average of these squares, and finally taking the square root of the average. The final value represents the typical intensity of the acceleration oscillation within that time period. By using this method of detecting spectral peaks within a preset physiological frequency range and combining it with energy threshold determination, the periodic micro-motion component with physiological rhythm characteristics generated by the user's handheld operation can be stably separated from the device acceleration data, thus providing accurate frequency and amplitude characteristic parameters for subsequent assessment of motion stability. Understandably, when a user holds a device, the subtle adjustments and physiological tremors of their arm and wrist muscles impart a continuous, minute oscillation within a specific frequency range. This oscillation, acting as a carrier signal, is sensed by the accelerometer, and spectral analysis can distinguish it from other potentially large-amplitude, non-periodic motion interference. The identified frequency and amplitude parameters directly reflect the user's stability and physiological state while holding the device, providing crucial physical evidence for calculating their level of concentration.

[0023] Record the frequency and amplitude values ​​of the current window and the window immediately preceding it. Subtract the frequency value of the preceding window from the frequency value of the current window, take the absolute value, and divide it by the time interval between the center points of the two windows to obtain the frequency change rate. Calculate the amplitude change rate in the same way. Multiply the frequency change rate by a preset weighting coefficient, such as 0.6, and multiply the amplitude change rate by another preset weighting coefficient, such as 0.4. Add the two weighted results to obtain the total change rate. Then, calculate the result mathematically using the natural constant e as the base and the negative sum of the change rates as the exponent. The result of this mathematical operation will fall between 0 and 1 and is defined as the micro-motion stability score of the current sliding time window. The closer the value is to 1, the more stable the device's motion mode. As the window slides in 0.1-second increments, a series of micro-motion stability scores are generated, arranged according to the window's center time point.

[0024] Since the time points of this scoring sequence do not perfectly correspond to the time points of the pixel change amplitude sequence, linear interpolation is performed on the micro-motion stability scoring sequence. Specifically, for each sampling time point of the pixel change amplitude sequence, an estimated value is calculated using a linear equation based on the two nearest micro-motion stability scores before and after it, thus generating a micro-motion stability scoring interpolation sequence that is perfectly aligned with the pixel change amplitude sequence at the same time point. Next, for each identical time point, a value is extracted from the pixel change amplitude sequence and denoted as A1, and a value is extracted from the corresponding interpolation sequence and denoted as A2.

[0025] Inputting A1 and A2 into a preset fusion function, the function can be implemented by calculating the product of the normalized complements of A2 and A1. First, A1 is normalized by dividing it by a preset maximum pixel change amplitude. Then, this normalization ratio is subtracted from 1 to obtain the first factor. Multiplying the first factor by A2 yields the attention intensity value at that moment. This calculation rule ensures a higher output attention intensity when screen changes are small and device movement is stable. After obtaining the attention intensity values ​​at all time points within a calculation cycle, a time-weighted average of these values ​​is calculated, i.e., the sum of values ​​is calculated by weighting each value according to its proportion within the entire cycle for its corresponding time segment.

[0026] Finally, this time-weighted average is divided by a theoretically maximum possible attention intensity value for normalization, and the final output is a value between 0 and 1, which is the final attention concentration.

[0027] It is worth noting that when users are highly focused on the screen, the content they view typically does not flicker violently or switch rapidly, exhibiting low pixel-level variation between consecutive screen image frames. Simultaneously, users in a focused state tend to maintain a stable, possibly slightly physiologically trembling, posture when holding the device. This posture reflects a relatively stable periodic micro-motion pattern with varying frequency and amplitude within the device's motion sequence, and the frequency and amplitude of this pattern change at a low rate over time. By quantifying the drastic changes in screen content and the stability of the device's micro-motion pattern, and fusing these two physically observable indicators, the user's real-time visual attention level can be indirectly and reliably inferred. This attention level, as a key real-time state parameter, will be directly used by subsequent ad filtering logic. For example, it can be used to determine whether to push ads that rely on sophisticated visual recognition when the user's attention is scattered, thereby achieving dynamic matching between ad formats and the user's current actual reception capacity. This is one of the core mechanisms of this invention for improving the effectiveness of ad delivery.

[0028] In another preferred embodiment of the present invention, the process of calculating the appropriate level of acoustic broadcasting is as follows: An ambient sound pressure level waveform is a sequence of sound pressure level decibel values ​​arranged in chronological order. First, a Fast Fourier Transform (FFT) is performed on this waveform data to transform it from a time-domain representation to a frequency-domain representation, thus obtaining a spectrum describing the distribution of sound energy at different frequencies. The analysis focuses on the low- to mid-frequency sub-bands where the human ear is sensitive and where common environmental noise is concentrated, such as the frequency range from 200 Hz to 2000 Hz. Within this range, the spectrum is divided into several consecutive frequency sub-bands, and the average energy value of each sub-band is calculated. Then, using a basic time window of 0.5 seconds, the changes in the average energy of a specific sub-band are observed within multiple consecutive such windows, for example, observing five consecutive time windows. When it is found that the average energy value of a certain sub-band consistently exceeds a preset energy threshold, such as 65 dB, within five consecutive windows, it is determined that a persistent background noise band exists in the current acoustic environment. This stable and high-energy narrowband noise usually originates from machine operation, air conditioning, or continuous background conversation, which severely interferes with the clarity of speech information.

[0029] In parallel, the ambient sound pressure level waveform is directly analyzed in the time domain to detect instantaneous, high-loudness sound events. The instantaneous sound pressure level at each sampling point on the waveform is calculated, and a relatively high threshold, such as 75 dB, is set to identify sudden sound events that may constitute interference. The number of times the instantaneous sound pressure level exceeds this threshold within a fixed time unit, such as 10 seconds, is counted. When more than 3 such peak events are detected within 10 seconds, it is determined that there are intermittent sudden high-loudness events in the environment. These events may come from sudden honking, shouting, or collision sounds that can completely mask or interrupt normal audio broadcasting.

[0030] To combine visual information for comprehensive judgment, the rear camera of the terminal device captures a scene image of the current environment. This image is resized to 224 pixels by 224 pixels and then input into a pre-trained lightweight convolutional neural network model. This model has been trained on a large dataset of images labeled as public places, vehicle interiors, and private spaces. The model performs multiple convolutional and pooling operations on the input image to extract features, and finally outputs three values ​​through a classification layer, representing the probability that the input image belongs to one of the three categories. Each probability value is between 0 and 1, and the sum of the three is 1. The largest of these three probability values ​​is selected, and its corresponding category name is used as the visual semantic analysis result. For example, if the model outputs a probability of 0.8 for public place, 0.15 for vehicle interiors, and 0.05 for private space, then the visual semantic analysis result is "public place."

[0031] The final judgment is made based on the combined results of the three analyses above. When two conditions are met simultaneously—that is, a persistent background noise band is identified and the visual semantic analysis indicates a public place—or when only intermittent sudden high-volume events are identified, the current acoustic environment is considered extremely unfavorable for audio information broadcasting, and the acoustic broadcasting suitability level is marked as unsuitable. If and only if all three analysis results are favorable—that is, there is neither a persistent background noise band nor intermittent sudden high-volume events, and the visual semantic analysis indicates a private space—the environment is considered quiet and private, and the acoustic broadcasting suitability level is marked as suitable. All situations that do not meet the above "unsuitable" or "suitable" judgment conditions, such as a private space with sudden noise, or a public place with background noise not consistently exceeding the standard, are marked as neutral in terms of acoustic broadcasting suitability.

[0032] The physical principle behind this process is based on the characteristics of sound wave propagation and superposition in the air. Continuous background noise increases the overall ambient noise level and reduces the signal-to-noise ratio of the speech signal, while sudden high-volume sounds directly mask auditory perception. Visual scene analysis is incorporated because different physical spaces are inherently associated with typical acoustic environment patterns; public places typically represent complex and uncontrollable noise sources. By quantitatively assessing continuous noise and sudden interference in the environment and combining this with spatial semantics, the likelihood of users clearly receiving the core information of audio advertisements can be objectively determined. This acoustic suitability level, as a key environmental state parameter, will be directly used in subsequent ad filtering logic. For example, when an ad is marked as unsuitable, any advertisements with audio as their core information carrier will be avoided, ensuring the ad format's compatibility with the user's current physical environment. This is a crucial technical step in achieving accurate targeting and reducing invalid exposure.

[0033] Based on the attention concentration level and the acoustic broadcast suitability level, all advertising content in the advertising material library to be pushed is filtered to generate a candidate advertising set; In a preferred embodiment of the present invention, the process of obtaining presentation attributes is as follows: The ad content package is parsed to obtain file data, which includes visual media data, audio media data, and interactive script data. The classification identifier of the ad content is determined based on the file data. The visual media data volume is obtained by multiplying the total number of pixels in the image sequence by the resolution and duration of the video track; the audio media data volume is quantized by multiplying the sampling rate and duration of the audio track; the interactive script data volume is quantized by the number of lines of code or the file size of the script file. When the amount of visual media data exceeds the first preset threshold, and the amount of visual media data exceeds the sum of the amount of audio media data and the amount of interactive script data, the advertising content will be marked as a visually dominant advertising. When the amount of audio media data exceeds the second preset threshold, and the amount of audio media data exceeds the sum of the amount of visual media data and the amount of interactive script data, the category identifier of the advertising content will be marked as an audio-dominated advertisement. When the amount of interactive script data exceeds the third preset threshold, and the amount of interactive script data exceeds the sum of the amount of visual media data and audio media data, the advertising content will be marked as an interactive-driven advertising. When any two of the three data volumes—visual media data volume, audio media data volume, and interactive script data volume—are greater than the fourth preset threshold, and the data volume of any one of them does not exceed the sum of the data volumes of the other two, the advertising content will be marked as a composite advertisement. If the ad content package does not meet any of the above judgment conditions, the ad content will be marked as other types of ads. For content categorized as visually dominant ads, a key visual complexity score is generated by calculating the image entropy value of its keyframes and the distribution dispersion of color and edge features. For content categorized as audio-driven advertising, an average speech intelligibility score is generated by calculating the signal-to-noise ratio and harmonic distortion of its audio track within the standard speech frequency range.

[0034] For content categorized as interactive ads, composite ads, or other types of ads, the key visual complexity score and average speech clarity score are recorded as preset default values.

[0035] It should be noted that when obtaining the presentation attributes of advertising content, for content categorized as visually dominant, a key visual complexity score needs to be generated. The processing begins by parsing the advertising content package and extracting its keyframes. For example, three frames are uniformly extracted from a video advertisement as keyframe inputs representing the visual content of that advertisement. For each keyframe image, it is first converted from the original red, green, and blue color space to a grayscale image. Then, the brightness distribution of all pixels in the grayscale image is statistically analyzed to form a histogram containing 256 gray levels. The information entropy value corresponding to this histogram is calculated. This entropy value is called image entropy, which quantifies the randomness and information richness of the grayscale distribution in the image. An image full of detail and texture usually produces a high entropy value, such as 6.5. Next, the color feature distribution dispersion of the keyframes is analyzed. Specifically, the standard deviation of all pixel values ​​in the red, green, and blue color channels is calculated. For example, the standard deviation for the red channel is 45, for the green channel 38, and for the blue channel 50. The average of these three standard deviations is defined as the color distribution dispersion. A higher value indicates stronger color contrast and more uneven distribution in the advertisement. Further analysis of the edge feature distribution dispersion of the keyframes is performed. An edge detection operator is used to process the grayscale image to identify the contours of objects in the image, and the proportion of edge pixels to the total number of pixels is calculated. For example, an edge pixel ratio of 0.3 reflects the complexity of contours and structures in the image. Finally, the image entropy, color distribution dispersion, and edge pixel ratio are linearly combined, for example, with weights of 0.4, 0.3, and 0.3 respectively, and summed to obtain a key visual complexity score. This score is a comprehensive numerical representation of the visual information density and structural complexity of the advertisement.

[0036] For content categorized as audio-driven, an average speech intelligibility score needs to be generated. The process separates the audio track from the ad content package as input. First, the audio signal is analyzed within the standard speech frequency range of 300 Hz to 3400 Hz. The signal-to-noise ratio (SNR), typically expressed in decibels (dB), is calculated within this frequency range; for example, an SNR of 24 dB is obtained. This value measures the strength of the speech signal relative to background noise. Next, the harmonic distortion (HDC) of the audio signal is calculated by analyzing its spectrum using a Fast Fourier Transform (FFT) and measuring the ratio of total harmonic energy to fundamental energy; for example, a harmonic distortion of 5% is obtained. This value reflects the degree of nonlinear distortion generated during signal transmission or processing; lower distortion means more faithful speech reproduction. Finally, the signal-to-noise ratio (SNR) value and the harmonic distortion (HDC) value are fused. For example, the SNR is normalized and multiplied by a weight of 0.7, and the reciprocal of the HDC value is normalized and multiplied by a weight of 0.3. The two are then added together to obtain the average speech intelligibility score, which quantifies the intelligibility and fidelity of the speech information in the advertisement.

[0037] For content categorized as interactive, composite, or other types of advertising, since their presentation does not primarily rely on independent visual or audio quality for evaluation, the aforementioned calculation process is not performed. Instead, the key visual complexity score and average speech clarity score are directly recorded as a preset default value, such as a uniform 0.5. The physical principle behind this process is that different forms of advertising rely on different physical media for effective information delivery. Visually-driven advertising uses light signals to act on the human eye, and its information delivery efficiency is directly related to the information density, color contrast, and outline clarity of the image. These physical attributes can be quantified using image entropy and distribution dispersion. Audio-driven advertising uses sound waves to act on the human ear, and its information delivery efficiency is directly related to the ratio of speech signal intensity to noise and the waveform fidelity of the signal. These physical attributes can be quantified using signal-to-noise ratio and harmonic distortion. The purpose of generating these objective scores is to combine the objective factor of the information-carrying quality of the advertising content itself with the user's subjective interests and real-time status in the subsequent step of calculating the advertising priority score. For example, an ad with a high visual complexity score may be more impactful when pushed to users at times when their attention is focused; an ad with a high speech clarity score may be easier to hear when pushed to an environment with an appropriate acoustic level. For interactive or composite ads, since their experience quality is difficult to measure using a single visual or audio metric, default values ​​are used, and the ranking decision is weighted more heavily by user interest preference models and real-time state matching.

[0038] Based on the attention concentration level and the acoustic broadcast suitability level, all advertising content in the advertising material library to be pushed is filtered to generate a candidate advertising set; In a preferred embodiment of the present invention, the process of generating a candidate advertisement set is as follows: The process of generating a candidate ad set takes two real-time calculated state parameters as input: attention concentration and acoustic broadcast suitability level. Attention concentration is a value between 0 and 1, representing the degree to which the user's vision is focused on the screen. Acoustic broadcast suitability level is a categorical variable consisting of three labels: unsuitable, neutral, and suitable, representing how friendly the current environment is to listening to audio information.

[0039] The selection logic is determined by which predefined interval the user's attention level falls into. These intervals are divided by two preset thresholds, such as a high attention threshold of 0.7 and a low attention threshold of 0.3. When the input attention level is greater than 0.7, the user is considered to be in a high attention state, at which point the likelihood of deep visual interaction with the screen is highest. Therefore, all content categorized as visually dominant, interactive, and hybrid ads in the ad creative library is selected and included in the candidate ad set. In a high attention state, the user's visual information processing channel is in a highly efficient reception state, capable of effectively parsing complex images, animations, or interactive interfaces requiring visual feedback. Although audio-dominated ads can technically be played, it is more appropriate to use the more efficient visual channel to deliver information in this state. Therefore, visually dominant, interactive, and audiovisual hybrid ad types are prioritized to match the optimal perceptual bandwidth.

[0040] When the input attention level is between 0.3 and 0.7, the user is considered to be in a state of moderate attention. Further assessment based on the acoustic environment is needed. If the acoustic suitability level is marked as "suitable," indicating a quiet environment suitable for listening, then content categorized as audio-driven or interaction-driven ads will be included in the candidate ad set. The logic here is that under moderate attention, users may not be able to maintain sustained focus on complex visual content, but they can still receive audio information. Meanwhile, interactive ads, due to their degree of active participation, may still attract the user's limited attention. If the acoustic suitability level is marked as "unsuitable" or "neutral," indicating a noisy or undesirable environment, then only content categorized as interaction-driven ads will be included in the candidate ad set. This is because interactive experiences are relatively less dependent on sound, and even in unfavorable acoustic environments, touch operations may still attract users.

[0041] When the input attention level is less than 0.3, the user is considered to be in a low-attention state, meaning their visual focus on the screen is very limited. If the acoustic broadcast suitability level is marked as suitable, content categorized as audio-driven or other types of ads will be included in the candidate ad set. This is because, when the user's gaze is off the screen, delivering information through the sound channel is a viable alternative, while other types of ads serve as simpler, less disruptive alternatives. If the acoustic broadcast suitability level is marked as unsuitable or neutral, only content categorized as other types of ads will be included in the candidate ad set. This is based on the practical consideration that, under the dual disadvantages of scattered user attention and a noisy environment, pushing any ad that relies on visual focus or clear hearing may be ineffective; therefore, the simplest and least disruptive ad type will be selected as the default option.

[0042] The output of the entire screening process is a subset called the candidate ad set. This subset comes from the complete ad creative library but only includes ad content whose presentation format best matches the user's reception capabilities based on the current real-time perception state. The core principle of this step is information matching theory, which dynamically aligns the form of information with the audience's currently available perceptual channels and capabilities. By quantifying abstract attention states and physical acoustic environments into actionable thresholds and rules, this process achieves the first precise filtering from the full ad library to the contextualized candidate set. This provides a high-quality input set pre-adapted to the user's current state for subsequent fine-grained ranking based on content quality and user preferences, thereby improving the overall contextual relevance of ad delivery at the source. It should be noted that ad types not included in the above logic are not necessarily unsuitable; rather, the rules constructed by this invention based on real-time perception data suggest that, in the current specific state, the selected types have a higher expected matching degree and dissemination effectiveness.

[0043] Based on the user interest preference model and the presentation attributes of each advertisement in the candidate advertisement set, the priority score of the advertisement content is calculated. In a preferred embodiment of the present invention, the process of calculating the priority score is as follows: A pre-trained user interest preference model is loaded. Mathematically, this model is a logistic regression classifier with a clear hierarchical connection architecture to map raw ad features to interest probabilities. The model's input layer receives a carefully constructed numerical feature vector with a fixed dimension, such as 5000, where each dimension corresponds to a specific feature. These features mainly come from two parts: a numerical representation of the ad text content and the ad's category and metadata information. For text information such as ad titles and keywords, word segmentation is first performed, and then the TF-IDF method is used to convert it into a high-dimensional sparse vector. The TF-IDF method quantifies the importance of a word to the current ad by calculating the product of the frequency of each word in a single ad document and the logarithm of its frequency in the entire ad corpus, thus converting the text into a numerical value reflecting its key content. For discrete information such as the ad's category and brand, one-hot encoding is used to convert it into a binary vector, where each dimension represents a possible category or brand; if the ad belongs to that category, that dimension is 1, otherwise 0. All these numerical features are concatenated to form the feature vector of the input layer.

[0044] This high-dimensional feature vector is then fed into the decision core layer of the logistic regression model. This layer essentially performs a weighted summation plus a bias term. Specifically, the model maintains a corresponding weight parameter for each dimension of the input feature vector; these weights are learned and determined during training. The model multiplies the value of each dimension of the input vector with its corresponding weight, sums all the products, and adds a scalar bias to obtain a linear combination value. This linear combination value is then fed into a sigmoid activation function, a mathematically defined S-shaped curve that maps any real number to a value between 0 and 1. Through this mapping, the model's final output layer produces an interest prediction score between 0 and 1, for example, 0.72. This score can be interpreted probabilistically as the conditional probability that, given the current input ad features, the model predicts the user will click on the ad.

[0045] The model's ability to output meaningful predictions stems from its training process. Training data is derived from binary behavioral logs collected over a period, such as 90 days, documenting user clicks on numerous ads. Each log entry is associated with a feature vector of an ad, while also incorporating features derived from user-subscribed content tags. Supervised training using logistic regression aims to find an optimal set of weights and biases that minimizes the overall difference between the model's predicted probabilities for all historical training samples and the actual clicks made by the user—the logistic loss function. This optimization process is typically iteratively performed using gradient descent, for example, 100 iterations. Through this training, the model learns weights that encode the historical importance of different ad features in attracting a specific user. When a new ad's feature vector is input, the model performs forward computation through the aforementioned hierarchical connections. High-weight features contribute more to the final linear combination value, which is then transformed into a higher predicted probability via the sigmoid function, thus achieving a probabilistic prediction of the user's future interests based on historical behavioral patterns. This predicted value serves as a stable indicator for quantifying users' long-term interests and preferences, providing a personalized benchmark for subsequent calculations that integrate with real-time status and ad quality scores.

[0046] For each advertisement in the candidate ad set generated through contextual filtering, its content features are first extracted and input into the model to obtain the corresponding interest prediction score. Then, based on the classification identifier determined in the previous steps, its base score is calculated using different rules. If the advertisement's classification identifier is visually dominant, the pre-calculated key visual complexity score is retrieved from the ad attribute library. This score might be a value like 0.85, representing the complexity and sophistication of its visual design. The interest prediction score is directly added to this score to obtain the base score. If the advertisement's classification identifier is audio-dominant, its pre-calculated average speech clarity score is retrieved. This score might be a value like 0.90, representing the high fidelity of its speech quality. The interest prediction score is directly added to this score to obtain the base score. If the advertisement's classification identifier is a composite advertisement, both its key visual complexity score and average speech clarity score are retrieved simultaneously. The arithmetic mean of these two scores is calculated, and this mean is added to the interest prediction score to obtain the base score. For ads categorized as interactive or other types, since they lack directly related visual or audio quality scores, their base score is directly equivalent to the interest prediction score.

[0047] Next, a real-time state adjustment mechanism is introduced. The real-time calculated attention concentration is used as the first adjustment coefficient, for example, 0.8. Simultaneously, based on the real-time determined acoustic broadcast suitability level, a second adjustment coefficient is obtained through a preset mapping relationship. This mapping relationship is defined as follows: when the level is unsuitable, the coefficient is 0.5; when the level is neutral, the coefficient is 0.8; and when the level is suitable, the coefficient is 1.2. Assuming the current acoustic broadcast suitability level is determined to be "suitable," the second adjustment coefficient is 1.2. Then, these two adjustment coefficients are weighted and summed to obtain a comprehensive coefficient. Specifically, the first adjustment coefficient is multiplied by a preset first weight, for example, 0.6, and the second adjustment coefficient is multiplied by a preset second weight, for example, 0.4. The two products are then added together; for example, 0.8 multiplied by 0.6 plus 1.2 multiplied by 0.4 yields a comprehensive coefficient of 0.96.

[0048] Finally, the calculated base score is multiplied by this comprehensive coefficient, and the resulting product is the final priority score for the ad content. For example, a visually dominant ad with an interest prediction score of 0.72 and a key visual complexity score of 0.85 has a base score of 1.57. Multiplying this by the comprehensive coefficient of 0.96 yields a priority score of approximately 1.51. The physical and cognitive principle behind this process is that the final effect of ad delivery is determined by the user's long-term stable interest preferences, the objective production quality of the ad content itself, and the user's real-time physiological and physical reception state. The base score integrates subjective interest and objective quality, reflecting the potential attractiveness of the ad under ideal conditions. The comprehensive coefficient represents the "discount or gain factor of the ideal state" assessed from the two dimensions of user attention level and environmental noise level at the current moment. Essentially, it quantifies the degree to which current instantaneous conditions support or weaken the effective transmission of information. Multiplying the base score by this coefficient is equivalent to calibrating the static potential attractiveness according to the dynamic real-time reception capability, thereby obtaining a priority score that more accurately predicts the actual effect that pushing the ad may produce in the current specific context. This score will be used directly to rank all candidate ads in the final ranking, enabling a leap in decision-making from "what users might like" to "what is most suitable for users to receive here and now".

[0049] Based on the priority scores from high to low, a number of advertisements are pushed to the terminal device.

[0050] There is no limit to the number of push notifications, and the push order will be based on priority scores from high to low.

[0051] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art based on the actual situation. The weights and thresholds mentioned above can be set by those skilled in the art based on their experience, and are not restricted here.

[0052] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An advertising push method based on big data analysis, characterized in that, Includes the following steps: The system calculates the user's attention level on the terminal device screen and the appropriateness level of the acoustic broadcast based on multi-dimensional data from the user's terminal device. Obtain the presentation attributes of each advertisement in the advertising material library to be pushed, wherein the presentation attributes include the category identifier and the association score corresponding to different category identifiers; Based on the attention concentration level and the acoustic broadcast suitability level, all advertising content in the advertising material library to be pushed is filtered to generate a candidate advertising set; Based on the user interest preference model and the presentation attributes of each advertisement in the candidate advertisement set, the priority score of the advertisement content is calculated. Based on the priority scores from high to low, a number of advertisements are pushed to the terminal device.

2. The advertising push method based on big data analysis according to claim 1, characterized in that, The process of obtaining multi-dimensional data is as follows: The original audio sampling stream from the built-in microphone of the terminal device is continuously read, and the original audio sampling stream is divided into preset time windows. The sound pressure level of the audio signal in each time window is calculated, and the sound pressure level values ​​arranged in chronological order constitute the ambient sound pressure level waveform. The screen image frame is obtained by capturing the image data of the screen display area through the screen content capture interface. The output signals from the triaxial accelerometer and triaxial gyroscope of the inertial measurement unit are collected synchronously. The output signals are aligned and synthesized according to the same time base to generate continuous data describing the translation and rotation motion of the device in three-dimensional space, thus forming the device motion sequence. Ambient sound pressure level waveforms, screen image frames, and device motion sequences constitute multi-dimensional data.

3. The advertising push method based on big data analysis according to claim 2, characterized in that, The process of calculating attention concentration is as follows: Perform continuous multi-frame analysis on the screen image frames, calculate the overall pixel change amplitude of the screen display area frame by frame, and obtain the pixel change amplitude sequence; Set a sliding time window, identify the periodic micro-motion pattern within the sliding time window and calculate its frequency and amplitude, and calculate the rate of change of frequency and the rate of change of amplitude based on the frequency and amplitude of the current and previous windows. The micro-motion stability score is calculated based on the frequency change rate and amplitude change rate. The micro-motion stability score is then linearly interpolated and aligned with the pixel change amplitude sequence. After alignment, the attention concentration is calculated.

4. The advertising push method based on big data analysis according to claim 3, characterized in that, The process of calculating the acoustic broadcast suitability level is as follows: Based on the ambient sound pressure level waveform, determine whether there is a continuous background noise band and whether there are intermittent sudden high-loudness events; The camera device based on the terminal device captures the current scene image, and obtains the visual semantic analysis result based on the current scene image. The visual semantic analysis result includes public places, transportation vehicles, and private spaces. The appropriate level of acoustic broadcasting is determined based on the presence of a persistent background noise band, the presence of intermittent sudden high-volume events, and the results of visual semantic analysis.

5. The advertising push method based on big data analysis according to claim 4, characterized in that, The process of obtaining presentation attributes is as follows: Parse the ad content package to obtain file data, which includes visual media data, audio media data, and interactive script data. The classification identifiers for advertising content are determined based on file data. These identifiers include visually-driven ads, audio-driven ads, interactive ads, hybrid ads, and other types of ads. The association scores for advertisements of different categories are calculated separately. The association score for visually dominant advertisements includes the key visual complexity score, the association score for audio-dominated advertisements includes the average speech clarity score, and the association scores for interactive, composite, or other types of advertisements include both the key visual complexity score and the average speech clarity score.

6. The advertising push method based on big data analysis according to claim 5, characterized in that, The process of generating a candidate ad set is as follows: When the attention concentration exceeds the preset high attention threshold, content categorized as visually dominant, interactive, or hybrid ads will be included in the candidate ad set. When the preset low attention threshold is less than or equal to the preset high attention threshold, the following steps are performed: If the acoustic broadcast suitability level is suitable, then content categorized as audio-driven or interactive-driven ads will be included in the candidate ad set. If the acoustic broadcast suitability level is unsuitable or neutral, then content categorized as interactive-driven advertising will be included in the candidate advertising set. When the level of attention is less than the preset low attention threshold, perform the following steps: If the acoustic broadcast suitability level is suitable, then content categorized as audio-dominant ads and other types of ads will be included in the candidate ad set; If the acoustic broadcast suitability level is unsuitable or neutral, then content categorized as other types of advertising will be included in the candidate advertising set.

7. The advertising push method based on big data analysis according to claim 5, characterized in that, The calculation of priority scores includes: The interest prediction score of each advertisement in the candidate advertisement set is obtained based on the pre-trained interest preference model; The basic score for advertising content is obtained based on the interest prediction score and the association score. Using attention concentration as the first adjustment coefficient, a mapping relationship between the appropriate level of acoustic broadcasting and the second adjustment coefficient is established to obtain the current second adjustment coefficient; The weighted sum of the first adjustment coefficient and the second adjustment coefficient yields the comprehensive coefficient. The base score is then multiplied by the comprehensive coefficient to obtain the priority score.

Images