AI-based offline user behavior analysis and advertisement putting strategy optimization method

Abstract

The invention relates to the cross field of offline marketing and AI algorithms, and discloses an AI-based offline user behavior analysis and advertisement putting strategy optimization method. According to the method, recessive behavior data and environmental interference data are collected through an offline terminal device provided with a sensor assembly, preorder advertisement interaction data are obtained, the interest association degree is calculated, an AI optimization module is called for coupling calibration and time sequence correction, and an advertisement putting optimization decision is generated through a preset algorithm model. The method creatively integrates recessive behaviors, environmental interference and time sequence correlation coprocessing logic, solves the problem of insufficient optimization accuracy in the prior art, provides reliable basis for advertisement content, playing sequence and marketing strategy adjustment, and is suitable for offline advertisement terminal precision marketing scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of offline marketing and AI algorithm, in particular to an offline user behavior analysis and advertisement placement strategy optimization method based on AI. BACKGROUND

[0002] The existing offline advertisement placement optimization method mainly depends on the explicit data such as the number of off-hook triggers and the playing time of the advertisement, without considering the coupling influence of the implicit behavior of the user after off-hook and environmental interference, and lacking effective correction of the time sequence association between the previous advertisement and the current advertisement. This single-dimensional data processing method makes it difficult to accurately capture the real interest of the user, the advertisement placement decision is easily disturbed by environmental noise, the time sequence association information cannot be effectively converted into optimization basis, and finally the matching degree of the advertisement content, playing order and user demand is insufficient, which is difficult to meet the actual needs of precision marketing.

[0003] Based on the above problems, there is an urgent need for a technical solution that can integrate implicit behavior, environmental interference and time sequence association information, and generate optimization decisions through collaborative processing, to solve the core problem of insufficient optimization accuracy of existing methods. SUMMARY

[0004] The purpose of the present application is to solve the problems of evaluation lag, process control loss, insufficient incentive mechanism and data dimension fragmentation in the prior art, so as to provide an offline user behavior analysis and advertisement placement strategy optimization method based on AI, comprising the following steps:

[0005] S1: collecting implicit behavior data of the user after off-hook through an offline terminal device configured with a sensor component, the implicit behavior data being the holding stability of the user after off-hook; performing outlier rejection preprocessing on the implicit behavior data;

[0006] S2: collecting environmental interference data of the user during interaction through the offline terminal device, the environmental interference data including environmental light intensity and environmental sound interference degree;

[0007] S3: obtaining interaction data of the user in the previous advertisement, calculating the interest association degree of the previous advertisement and the current advertisement based on the interaction data, the interaction data including the actual stay time of the user in the previous advertisement and the average stay time of the previous advertisement;

[0008] S4: calling an AI optimization module, the AI optimization module receiving the preprocessed implicit behavior data, environmental interference data and interest association degree, coupling and calibrating the implicit behavior data and the environmental interference data first, then combining the interest association degree for time sequence correction, and generating an advertisement placement optimization decision through a preset algorithm model.

[0009] Preferably, in step S1, the grip stability is calculated in the following way: three-dimensional acceleration data under a standard stable grip state are collected in advance, and the variance of the three-dimensional acceleration data under the standard stable grip state is calculated and denoted as the standard stable acceleration variance; three-dimensional acceleration data after the user takes off the device is collected in real time, and the variance of the real-time collected three-dimensional acceleration data is calculated and denoted as the actual acceleration variance; the grip stability is the ratio of the standard stable acceleration variance to the actual acceleration variance.

[0010] More preferably, in step S2, the sensor components of the offline terminal device include a digital light sensor and a sound pressure sensor. The ambient light intensity is collected by the digital light sensor, and the ambient sound interference is obtained by collecting and converting the sound pressure sensor.

[0011] More preferably, in step S3, the interest relevance between the preceding advertisement and the current advertisement is calculated in the following way: an advertising keyword library is pre-built, and the keyword matching degree between the preceding advertisement and the current advertisement is calculated based on the advertising keyword library; the ratio of the user's actual dwell time on the preceding advertisement to the average dwell time of the preceding advertisement is calculated and recorded as the dwell time ratio; the interest relevance between the preceding advertisement and the current advertisement is the product of the keyword matching degree and the dwell time ratio.

[0012] More preferably, the preset algorithm model in step S4 includes a formula for the effective weighting of latent behavior-environmental interference collaboration, which is:

[0013] ;

[0014] Wherein, W is the effective weight of implicit behavior; S is the grip stability mentioned in step S1; L is the ambient light intensity mentioned in step S2; N is the ambient sound interference level mentioned in step S2; k1 is the light interference calibration coefficient, which is iteratively optimized by the AI ​​optimization module; k2 is the sound interference calibration coefficient, which is iteratively optimized by the AI ​​optimization module; and k3 is the stability enhancement calibration coefficient, which is iteratively optimized by the AI ​​optimization module.

[0015] More preferably, the preset algorithm model in step S4 further includes a hidden weight-temporal correlation coupling correction formula, which is:

[0016] ;

[0017] Wherein, M is a time sequence correlation correction coefficient; C is the interest correlation degree of the preceding advertisement and the current advertisement in step S3; W is the implicit behavior effective weight calculated by the implicit behavior-environment interference synergistic effective weight formula; k4 is a time sequence correlation calibration coefficient, which is iteratively optimized by the AI optimization module.

[0018] Further preferably, the preset algorithm model in step S4 further comprises an advertisement launching optimization decision coefficient formula, and the advertisement launching optimization decision coefficient formula is:

[0019] ;

[0020] Wherein, D is an advertisement launching optimization decision coefficient; M is a time sequence correlation correction coefficient calculated by the implicit weight-time sequence correlation coupling correction formula; T is a screen touch frequency before the user scans the code, which is collected by a capacitive touch sensor in the sensor assembly of the offline terminal device; a, b, and c are touch frequency fitting coefficients, which are fitted based on historical user behavior data; W is an implicit behavior effective weight calculated by the implicit behavior-environment interference synergistic effective weight formula.

[0021] Further preferably, in step S4, the AI optimization module iteratively optimizes the light interference calibration coefficient k1, the sound interference calibration coefficient k2, and the stability enhancement calibration coefficient k3 based on the corresponding relationship between historical user behavior data and advertisement launching results.

[0022] Further preferably, in step S4, the screen touch frequency T is collected in the following manner: the capacitive touch sensor counts the number of touches of the screen of the offline terminal device by the user within a timing period, and the number of touches is the screen touch frequency T.

[0023] Further preferably, in step S4, the AI optimization module calls the implicit behavior data, the environment interference data, and the interaction data collected in the last collection period as iterative training data each time the iteration is performed.

[0024] The present application has the following advantages:

[0025] The core creative technical point of the present application is to couple and calibrate the implicit behavior data and the environment interference data, and then perform synergistic processing logic of time sequence correction combined with the interest correlation degree. This technical point accurately solves the core problem of the single-dimensional processing data and the insufficient optimization accuracy of the existing method in the background technology, effectively filters the influence of environmental interference on user behavior data, fully excavates the user interest information in the time sequence correlation, makes the advertisement launching decision more in line with the real needs of users, and improves the adaptability and rationality of the advertisement launching. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 For the application, an AI-based offline user behavior analysis and advertising placement strategy optimization method flowchart. DETAILED DESCRIPTION

[0027] The application will be further described in detail below in conjunction with the drawings and specific embodiments. The embodiments take automobile repair enterprises as examples, but those skilled in the art should understand that the method of the application is also applicable to furniture manufacturing, packaging printing, industrial coating and other industrial enterprises involving volatile organic compounds (VOCs) emissions.

[0028] The traditional technical solution has the following technical problems: the existing offline advertising placement optimization method only relies on explicit data such as the number of off-hook times and the playing time, does not integrate implicit behavior, environmental interference and time sequence correlation information, and lacks a complete technical link from data collection to decision making, resulting in inaccurate capture of user real interest, insufficient precision of advertising placement decision, and difficulty for those skilled in the art to reproduce the complete optimization process.

[0029] Based on this, please refer to Figure 1The embodiment provides an offline user behavior analysis and advertisement putting strategy optimization method based on AI, and the implementation process is as follows: firstly, a terminal device configured with a sensor component is deployed, the sensor component at least comprises a MEMS three-dimensional accelerometer, a digital light sensor, a sound pressure sensor and a capacitive touch sensor, all the sensors are connected with a main control chip of the terminal device through an I2C bus, the main control chip adopts an STM32F4 series MCU, and is responsible for synchronous acquisition and preliminary transmission of sensor data. When step S1 is executed, the MEMS three-dimensional accelerometer collects three-dimensional acceleration data ax, ay, az of a user after taking off the handset in real time at a sampling frequency of 10 Hz, the collection duration is a complete interaction period from after the user taking off the handset to before releasing the handset, and the main control chip temporarily stores the collected acceleration data in a local SRAM, the SRAM capacity is not less than 16 KB, and the acceleration data is marked as original data of holding stability in implicit behavior data; then, the abnormal value elimination preprocessing of step S1.1 is executed, the main control chip calls a pre-stored Grubbs criterion algorithm, sets a significance level to 0.05, detects abnormal values for the original acceleration data, if the deviation of a sampling point is more than 3 times of a standard deviation, the sampling point is determined as an abnormal value, the abnormal value is replaced by the mean value of two adjacent effective sampling points through a linear interpolation method, and the effectiveness of the implicit behavior data is ensured. Step S2 is executed synchronously with step S1, the digital light sensor selects a BH1750 model, collects environmental light intensity at a sampling frequency of 1 Hz, and directly transmits a digital signal to the main control chip, and the unit is lux; the sound pressure sensor selects a MAX9814 model, collects environmental sound signals, converts the sound pressure signals into 0-3.3V analog voltage signals, converts the analog voltage signals into digital signals through a 12-bit ADC built in the main control chip, and converts the digital signals into environmental sound interference degrees through a preset voltage-dB conversion algorithm, the unit is dB, the voltage-dB conversion algorithm is based on a sensor datasheet calibration curve, and two types of environmental interference data and the implicit behavior data are synchronously stored. Step S3 is started after step S2 is executed, the main control chip obtains interaction data of a user in a previous advertisement from an advertisement background database through an Ethernet module, an actual stay duration in the interaction data is recorded by a timing module of the terminal device during playing of the previous advertisement, the timing module is based on a timer of the main control chip, an average stay duration is an arithmetic average value of stay durations of all users of the previous advertisement stored in the background database in the past 7 days, the main control chip calculates an interest correlation degree based on the two parameters and an advertisement keyword library, the advertisement keyword library is pre-stored in a flash memory of the terminal device, the flash memory capacity is not less than 128 KB, and the flash memory comprises core keywords and matching rules of various advertisements.In step S4, the master chip calls the AI optimization module through the UART interface, the AI optimization module is deployed in the edge computing unit of the terminal device, the edge computing unit adopts the NVIDIA Jetson Nano, the AI optimization module first acquires the preprocessed implicit behavior data, environmental interference data and interest correlation degree through the data receiving submodule, then the implicit behavior data and the environmental interference data are processed by the coupling calibration submodule, the data calibration is completed through the basic formula in the preset algorithm model, the time sequence dimension of the calibrated data is corrected by the time sequence correction submodule combined with the interest correlation degree, and finally the decision generation submodule generates an advertising optimization decision based on the corrected data. The decision result is transmitted in a structured data format, the structured data format adopts JSON, and is transmitted to the master chip. The master chip adjusts the advertising content loading order, the playing time length allocation and the marketing strategy triggering condition according to the decision result. The processing delay of the whole process is controlled within 100ms, which meets the real-time interaction demand in offline.

[0030] The prior art has the following technical problems: the prior art does not specify the specific calculation logic of the holding stability in the implicit behavior data, does not specify the calibration method of the standard stable state and the processing details of the actual data, so that the quantization results of the holding stability in different scenes lack uniformity, and cannot provide reliable data input for subsequent optimization.

[0031] Therefore, the calculation process of the holding stability in step S1 is as follows: first, the standard stable acceleration variance is pre-calibrated. In a laboratory environment, 100 testers with different hand sizes are selected, each tester holds the offline terminal device in a natural stable posture, and a MEMS three-dimensional accelerometer collects three-dimensional acceleration data of each tester for 30 seconds at a sampling frequency of 10Hz. The master chip calculates the variance var(ax ref ,ay ref ,az ref ) of each data segment, and then takes the arithmetic mean of the variances of the 100 testers as the standard stable acceleration variance. This value is pre-stored in the Flash memory of the terminal device and remains fixed during subsequent calculation. In the actual interaction process, the MEMS three-dimensional accelerometer collects real-time three-dimensional acceleration data of the user after picking up the phone, and the master chip calculates the actual acceleration variance using the sliding window method. The window size is set to 5 sampling points, corresponding to a time length of 0.5 seconds. Each new sampling point is removed from the earliest sampling point to ensure the real-time nature of the variance calculation. The holding stability is calculated by the master chip, which divides the standard stable acceleration variance by the current actual acceleration variance to obtain a dimensionless holding stability value. If the calculation result is greater than 2.0, it is taken as 2.0 to avoid the influence of extreme values. If the calculation result is less than 0.5, it is taken as 0.5 to represent an extremely unstable state. The calculation result is rounded to two decimal places and temporarily stored in the SRAM for subsequent step calling.

[0032] The prior art has the following technical problems: the prior art does not specify the hardware model, the collection frequency and the data conversion logic of the environmental interference data collection, so that the specific sensor selection and data processing method cannot be determined by the person skilled in the art, and the environmental data collection link is not fully disclosed.

[0033] Therefore, the collection process of the environmental interference data in step S2 is as follows: in the sensor assembly of the offline terminal device, a digital light sensor of model BH1750 is selected, the sensor is connected with the master control chip through an I2C interface, the default setting is a high resolution mode, the precision is ±20%, the sampling frequency is set to 1Hz, the sensor directly outputs a 16-bit digital light intensity value after each sampling, the unit is lux, and the master control chip reads the value through the I2C, and the value can be used as the environmental light intensity parameter without additional conversion; a sound pressure sensor of model MAX9814 is selected, the sensor integrates a preamplifier circuit and an automatic gain control function, is connected with an ADC pin of the master control chip through an analog signal output interface, the sensor converts the environmental sound signal into an analog voltage signal of 0-3.3V, the master control chip starts the built-in 12-bit ADC to sample the analog signal, the sampling frequency is set to 100Hz, and a digital quantity of 0-4095 is obtained after each sampling, and the digital quantity is converted into an environmental sound interference degree through a pre-stored calibration formula, the unit is dB, the calibration formula is obtained based on a voltage-dB curve fitting in the datasheet of the sensor, for example, when the ADC sampling value is 2048, the corresponding voltage is 1.65V, which is converted into an environmental sound interference degree of 60dB, and the voltage increases by 0.33V when the sampling value increases by 409.6, and the sound interference degree increases by 10dB, so that the accuracy of the conversion result is ensured; the two types of environmental interference data and the implicit behavior data are collected synchronously through the timer of the master control chip, the collection timestamp is accurate to the millisecond level, and the correlation of the data in the time dimension is ensured.

[0034] The prior art has the following technical problems: the prior art does not specify the calculation details of the interest correlation degree between the previous advertisement and the current advertisement, does not specify the construction rule of the keyword library and the calculation logic of the stay time ratio, so that the time sequence correlation information cannot be effectively quantified, and it is difficult to be converted into a reliable basis for optimization decision.

[0035] Based on this, the calculation process of interest correlation degree in step S3 is as follows: first, an advertisement keyword library is constructed, and in the advertisement background system, 3-5 core keywords are labeled for each type of advertisement, such as refrigerator, washing machine, energy saving for home appliance advertisements, and a weight is set for each keyword, the weight value is 0.2-1.0, the core keyword weight is set to 1.0, the keyword library is stored in a structured table form, including three fields of advertisement ID, keyword and keyword weight, and the terminal device synchronously updates the keyword library from the background through Ethernet at regular intervals; when calculating the keyword matching degree of the previous advertisement and the current advertisement, the master chip first obtains the IDs of the previous advertisement and the current advertisement, extracts the keywords and corresponding weights of the two types of advertisements from the locally stored keyword library, and calculates the matching degree by using the cosine similarity algorithm, for example, the previous advertisement keyword weight vector is [1.0, 0.8, 0.5], and the current advertisement keyword weight vector is [0.9, 0.8, 0.6], then the matching degree is the dot product of the two vectors divided by the product of the vector lengths, the calculation result is kept to 2 decimal places, and the value range is 0-1.0; when calculating the staying time ratio, the master chip obtains the actual staying time of the user in the previous advertisement from the background database, the unit is second, and the average staying time of the previous advertisement in the past 7 days is also obtained, the unit is second, if the average staying time is 0, it is set to 10 seconds by default to avoid the divisor being 0, and the staying time ratio is obtained by division operation, and kept to 2 decimal places; the interest correlation degree is the product of the keyword matching degree and the staying time ratio, if the product is greater than 1.5, take 1.5, to ensure that the value is within a reasonable range, and the calculation result is temporarily stored in SRAM for calling by the AI optimization module.

[0036] The traditional technical solution has the following technical problems: the existing method does not establish a collaborative processing model for implicit behavior data and environmental interference data, does not explain the theoretical basis, parameter dimension and calculation hardware support of the model, so that the two types of data are independent of each other, cannot effectively filter the influence of environmental interference on implicit behavior data, and it is difficult for those skilled in the art to realize the coupling calibration of data.

[0037] Based on this, the implementation process of the implicit behavior-environmental interference collaborative effective weight formula in step S4 is as follows: the formula is designed based on the space-time feature coupling theory and the Weber-Fechner law, and the core is to dynamically calibrate the reliability of the implicit behavior data through the environmental interference factor, and the formula expression is:

[0038] ;

[0039] Wherein the definition and dimension of each parameter are as follows: S is the holding stability calculated in step S1, which is the ratio of the standard stability acceleration variance to the actual acceleration variance, dimensionless, and the value range is 0.5-2.0; L is the ambient light intensity collected in step S2, the dimension is lux, and the value range is 0-10000 lux; N is the ambient sound interference degree converted in step S2, the dimension is dB and a dimensionless physical quantity, and the value range is 30-120 dB; k1 is the light interference calibration coefficient, the dimension is 1 / lux, the initial value is set to 0.001, and is iteratively optimized by the AI optimization module; k2 is the sound interference calibration coefficient, dimensionless, the initial value is set to 0.01, and is iteratively optimized by the AI optimization module; k3 is the stability enhancement calibration coefficient, dimensionless, the initial value is set to 0.5, and is iteratively optimized by the AI optimization module. The calculation process of the formula is executed by the coupling calibration sub-module of the AI optimization module, which is deployed on the GPU of the edge computing unit. The GPU of the edge computing unit uses the CUDA core of NVIDIA Jetson Nano, and uses a parallel computing architecture to ensure real-time performance: first, read the S, L, N values stored in the SRAM, and substitute them into the formula to calculate the exponential term and , wherein the exponential operation is approximated by Taylor expansion, expanded to 5 orders, and the precision is retained to 4 decimal places; then calculate and , these two terms represent the suppression degree of light and sound interference on implicit behavior data respectively, the stronger the interference, the smaller the value, and the more obvious the calibration effect; then calculate , this term amplifies the weight of high stability users through the square term of S, when S=2.0, the value of this term is 1+0.5x4=3.0, highlighting the implicit features of high interest users; finally, multiply the four terms to get the value of W, W is a dimensionless parameter, the value range is 0-2.5, the calculation result is transmitted to the time sequence correction sub-module, and stored in the local storage of the edge computing unit, the capacity of the local storage of the edge computing unit is not less than 8GB, for subsequent iterative optimization.

[0040] The traditional technical solution has the following technical problems: the existing method does not combine the credibility of implicit behavior data with time sequence association information, does not explain the specific algorithm and parameter optimization logic of the coupling correction of the two, resulting in lack of pertinence of time sequence association correction, unable to distinguish the time sequence association value corresponding to implicit behavior of different credibility, and insufficient accuracy of correction result.

[0041] Therefore, the implementation process of the implicit weight-time sequence association coupling correction formula in step S4 is as follows: the formula is designed based on information fusion theory, and the core is to dynamically adjust the correction strength of time sequence association information through the effective weight of implicit behavior, and the formula expression is:

[0042] ;

[0043] wherein the definitions and dimensions of each parameter are as follows: C is the interest correlation degree of the previous advertisement and the current advertisement calculated in step S3, dimensionless, with a value range of 0-1.5; W is the effective weight of the implicit behavior calculated by the formula of the effective weight of the implicit behavior-environment interference synergy, dimensionless, with a value range of 0-2.5; k4 is a time sequence correlation calibration coefficient, dimensionless, with an initial value of 0.2, which is iteratively optimized by the AI optimization module. The calculation process of the formula is performed by the time sequence correction submodule of the AI optimization module, which shares the GPU resources of the edge computing unit with the coupling calibration submodule and calculates in time sequence immediately after the coupling calibration: first, read the C value and the W value output by the coupling calibration submodule, calculate , where the logarithmic function takes the natural constant e as the base, and since W is dimensionless and has a value range of 0-2.5, , with a value range of 0-1.2, the logarithmic term reflects the diminishing marginal benefit of the effective weight of the implicit behavior, avoiding excessive amplification of the time sequence correlation correction strength when the W value is too large, and the calculation is quickly queried through the pre-stored logarithmic table, with a logarithmic table step of 0.01 and a precision of 4 decimal places, reducing the calculation time; then calculate , where the square term of C is used to enhance the time sequence correction basis of high correlation degree advertisements, and the exponential term is used to constrain the correction strength, when C = 1.5, , , the value of this term is 0.3624, avoiding excessive dependence on high correlation degree advertisements; finally, multiply C, and to get the M value, M is a dimensionless parameter with a value range of 0-1.8, and the calculation result is transmitted to the decision generation submodule and stored in association with the W value for subsequent decision coefficient calculation.

[0044] The traditional technical solution has the following technical problems: the existing method does not integrate implicit behavior, time sequence correlation and user active interaction data, does not explain the fusion logic of multi-dimensional data and the calculation details of the decision coefficient, resulting in a lack of comprehensive data support for decision generation, which cannot comprehensively reflect user interest, and the decision result is not specific enough.

[0045] Therefore, the implementation process of the advertisement optimization decision coefficient formula in step S4 is as follows: the formula is designed based on multi-factor decision theory, and the core is to integrate implicit behavior, time sequence correlation and active interaction three-dimensional information to generate the final decision basis, and the formula expression is:

[0046] ;

[0047] Wherein the definition and dimension of each parameter are as follows: M is the time correlation correction coefficient calculated by the implicit weight-time correlation coupling correction formula, dimensionless, the value range is 0-1.8; T is the screen touch frequency before the user scans the code, a dimensionless physical quantity and the unit is times / minute, the value range is 0-15 times / minute; a, b, c are touch frequency fitting coefficients, all are dimensionless, the initial value of a is-0.05, the initial value of b is 0.8, and the initial value of c is 0.2, which are obtained based on user historical behavior data fitting; W is the effective weight of implicit behavior calculated by the implicit behavior-environment interference synergistic effective weight formula, dimensionless, the value range is 0-2.5. The calculation implementation process of the formula is as follows: firstly, the value of T is collected by the capacitive touch sensor of the offline terminal device, the sensor is connected with the master control chip through the SPI interface, the master control chip sets a timing period of 1 minute, and the interruption is realized through the timer, and before the user triggers the scanning operation, the capacitive touch sensor detects an effective touch every time the interruption is triggered, the effective touch judgment standard is that the touch area is greater than or equal to 1 cm² and the touch time is greater than or equal to 50 ms, the master control chip counts the number of interruptions, and the counting result is the value of T; then the decision generation submodule of the AI optimization module reads the values of T, M and W, and reads the pre-fitted values of a, b and c from the local storage, the fitting process of a, b and c is based on 500 groups of user historical data, covering different age groups and interaction scenes, the least square method is used to perform quadratic curve fitting on T-actual conversion intention data, the fitting error is controlled within 5%, and the fitting error is controlled within 5%, to ensure that the touch frequency and the nonlinear relationship between the user interest can be accurately represented, for example, when T=6 times / minute, the value of this item is-0.05×36+0.8×6+0.2=3.8; then , which reflects the synergistic gain of the effective weight of implicit behavior and the time correlation correction coefficient, when W=2.0 and M=1.5, the value of this item is 1+2.0×1.5=4.0, which highlights the interest characteristics of high-value users; finally, M, and are multiplied to obtain the value of D, D is a dimensionless parameter, the value range is 0-5.0, the higher the value of D, the higher the priority of the corresponding advertisement, the decision generation submodule sorts the playing order of the advertisement content according to the value of D, adjusts the playing time of the advertisement, and the playing time is increased by 5 seconds every time the value of D increases by 1.0, and the corresponding marketing strategy is triggered, such as popping up a coupon pop-up window when D≥4.0, the decision result is transmitted to the master control chip through the UART interface, and the master control chip executes specific advertisement adjustment operation.

[0048] The prior art has the following technical problems: the prior art does not clearly define the iterative optimization logic of the calibration coefficient in the implicit behavior-environment interference synergistic effective weight formula, does not describe the source of the optimization data, the iteration frequency and the objective function, resulting in a fixed coefficient that cannot adapt to changes in data characteristics in different scenarios, and the accuracy of the formula calculation result decreases with changes in the scene.

[0049] Based on this, the iterative optimization process of k1, k2 and k3 by the AI optimization module in step S4 is as follows: first, build a historical database required for optimization, and at 24:00 every day, the terminal device uploads the implicit behavior data S, environmental interference data L, N and actual user interaction data collected that day to the cloud database, the actual user interaction data includes whether to scan the code, the interaction time, etc., and the cloud database stores historical data for nearly 30 days to ensure that the data volume is sufficient to support optimization; the parameter optimization submodule of the AI optimization module is deployed on the cloud server, the cloud server uses Intel Xeon E5 processor, and global iterative optimization is performed once a week, and the optimization first extracts historical data from the cloud database, and filters out valid samples, which must include complete S, L, N and actual interaction willingness labels, and the sample size is not less than 1000 groups; then set the optimization objective function as the minimum deviation of the implicit behavior effective weight W and the actual user interaction willingness, where the actual user interaction willingness is quantified by whether to scan the code, and scanning is 1 and not scanning is 0, and the deviation is calculated by the mean square error function, that is, , n is the sample size, Wi is the W calculation value of the i-th sample, and Yi is the actual interaction willingness label of the i-th sample; the gradient descent method is used to minimize the objective function, the iteration step is set to 0.001, the iteration number is set to 1000, the values of k1, k2 and k3 are updated after each iteration, and the error change is monitored in real time during the iteration process, and the iteration is stopped when the error does not decrease for 100 consecutive iterations; the optimized k1, k2 and k3 values are pushed to all offline terminal devices through the cloud, and the terminal device updates the local stored coefficient value when it starts next time, ensuring that the formula calculation result always adapts to the data characteristics of the current scene.

[0050] The prior art has the following technical problems: the prior art does not clearly define the collection timing, the setting method of the timing period and the counting rule of the touch frequency T of the screen, resulting in non-uniform T value collection standards for different terminal devices, poor data comparability, and affecting the accuracy of the advertising optimization decision coefficient.

[0051] Based on this, the collection process of the screen touch frequency T in step S4 is as follows: after the main control chip of the offline terminal device completes the loading of the advertisement playing interface, the collection function of the capacitive touch sensor is started, and a timer is initialized at the same time. The timer model is TIM2, the clock frequency is 72MHz, the timing period is set to 1 minute, and the period control is realized through the timer interrupt; the collection area of the capacitive touch sensor covers the interactive area of the terminal device screen, the interactive area is not less than 80% of the total area of the screen, the sensor has a built-in touch detection algorithm, the touch detection threshold is set to 100, the touch detection threshold is based on the factory calibration of the sensor and can be adjusted through software, when the strength of the detected touch signal exceeds the threshold and the touch duration is greater than or equal to 50ms, it is determined as an effective touch, and the sensor sends an interrupt signal to the main control chip through the interrupt pin; after the main control chip receives the interrupt signal, the interrupt service function is triggered, the number of effective touches is counted, and the counting variable is stored in the specified address of the SRAM. After each count, software anti-jitter is used to avoid false triggering, and the software anti-jitter method is to delay for 100ms; when the timer counts to 1 minute, the timer interrupt is triggered, the main control chip reads the current number of effective touches, and the number is the screen touch frequency T; if the user triggers the code scanning operation within the timing period, the main control chip immediately stops counting and reads the current number as the T value; if the user does not scan the code after the timing period ends, the T value is cleared, and the collection is restarted when the next advertisement is played, ensuring that the T value always corresponds to the user's active interaction behavior before scanning the code.

[0052] The traditional technical solution has the following technical problems: the existing method does not clearly specify the data source and timing of the AI optimization module iterative training, and does not explain the data selection rules and iteration trigger conditions, resulting in a lack of real-time data support for iterative training, and the model optimization cannot keep up with the changes in user behavior and environment, and the decision accuracy decreases over time.

[0053] Based on this, the AI optimization module in step S4 iterates the calling implementation process of the data as follows: the main control chip of the offline terminal device sets the data collection period to 1 hour, synchronously collects implicit behavior data, environmental interference data and interaction data in each collection period, filters the data after collection, the filtering rules include that the data collection time is not less than 10 seconds to ensure data integrity, the main control chip records fault codes when there is a sensor failure marker, the user interaction result includes explicit user interaction result, the user interaction result is code scanning or no code scanning, the effective data after filtering is stored in the local Flash memory of the terminal device, and the capacity of the local Flash memory is not less than 128 MB; the iteration training submodule of the AI optimization module sets the iteration period to 2 hours, reads the effective data in the latest collection period from the local Flash memory before each iteration, if the amount of effective data in the collection period is not less than 50 groups, the iteration training is started, if the data amount is insufficient, it is postponed to the next iteration period; during iteration training, the effective data is first divided into a training set and a validation set, the training set accounts for 80% and the validation set accounts for 20%, the training set is used to update model parameters, the model parameters include k1, k2, k3, k4 and a, b, c, the validation set is used to evaluate model performance, and the validation error needs to be less than 10% otherwise retraining; after training, if the model performance is improved compared with the last iteration, the model performance improvement standard is that the validation error is reduced by more than 5%, the parameter configuration of the AI optimization module is updated, if the performance is not improved, the original parameters are retained; at the same time, the terminal device uploads all iteration training data to the cloud every week, the cloud server performs global iteration optimization every 7 days, and the optimized model parameters are sent to all terminal devices, realizing the cooperation of local iteration and global iteration, and ensuring that the AI optimization module always maintains high decision accuracy based on the latest data.

[0054] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for optimizing offline user behavior and advertising placement strategies based on AI, characterized in that, Includes the following steps: S1: Collect implicit behavioral data of the user after taking off the device through an offline terminal device equipped with sensor components. The implicit behavioral data is the user's grip stability after taking off the device. Perform outlier removal preprocessing on the implicit behavioral data. S2: Collect environmental interference data during user interaction through the offline terminal device. The environmental interference data includes ambient light intensity and ambient sound interference. S3: Obtain user interaction data in the preceding advertisement, and calculate the interest correlation between the preceding advertisement and the current advertisement based on the interaction data. The interaction data includes the user's actual dwell time in the preceding advertisement and the average dwell time in the preceding advertisement. S4: Call the AI ​​optimization module. The AI ​​optimization module receives the preprocessed implicit behavior data, the environmental interference data, and the interest correlation. First, it couples and calibrates the implicit behavior data and the environmental interference data, and then performs time-series correction in combination with the interest correlation. Finally, it generates an advertising placement optimization decision through a preset algorithm model.

2. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 1, characterized in that, In step S1, the grip stability is calculated as follows: three-dimensional acceleration data under standard stable grip conditions are collected in advance, and the variance of the three-dimensional acceleration data under standard stable grip conditions is calculated and denoted as the standard stable acceleration variance; three-dimensional acceleration data after the user takes off the device is collected in real time, and the variance of the real-time collected three-dimensional acceleration data is calculated and denoted as the actual acceleration variance; the grip stability is the ratio of the standard stable acceleration variance to the actual acceleration variance.

3. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 1, characterized in that, In step S2, the sensor components of the offline terminal device include a digital light sensor and a sound pressure sensor. The ambient light intensity is collected by the digital light sensor, and the ambient sound interference level is obtained by collecting and converting the sound pressure sensor.

4. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 1, characterized in that, In step S3, the interest relevance between the preceding advertisement and the current advertisement is calculated in the following way: an advertising keyword library is pre-built, and the keyword matching degree between the preceding advertisement and the current advertisement is calculated based on the advertising keyword library; the ratio of the user's actual dwell time on the preceding advertisement to the average dwell time of the preceding advertisement is calculated and recorded as the dwell time ratio; the interest relevance between the preceding advertisement and the current advertisement is the product of the keyword matching degree and the dwell time ratio.

5. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 1, characterized in that, The preset algorithm model in step S4 includes a formula for the effective weight of the implicit behavior-environmental interference collaboration. The formula for the effective weight of the implicit behavior-environmental interference collaboration is as follows: ； Wherein, W is the effective weight of implicit behavior; S is the grip stability mentioned in step S1; L is the ambient light intensity mentioned in step S2; N is the ambient sound interference level mentioned in step S2; k1 is the light interference calibration coefficient, which is iteratively optimized by the AI ​​optimization module; k2 is the sound interference calibration coefficient, which is iteratively optimized by the AI ​​optimization module; and k3 is the stability enhancement calibration coefficient, which is iteratively optimized by the AI ​​optimization module.

6. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 5, characterized in that, The preset algorithm model in step S4 also includes a hidden weight-temporal correlation coupling correction formula, which is as follows: ； Where M is the temporal correlation correction coefficient; C is the interest correlation degree between the preceding advertisement and the current advertisement in step S3; W is the implicit behavior effective weight calculated by the implicit behavior-environment interference collaborative effective weight formula; k4 is the temporal correlation calibration coefficient, which is iteratively optimized by the AI ​​optimization module.

7. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 6, characterized in that, The preset algorithm model in step S4 also includes an advertising placement optimization decision coefficient formula, which is: ； Wherein, D is the advertising placement optimization decision coefficient; M is the time-series correlation correction coefficient calculated by the implicit weight-temporal correlation coupling correction formula; T is the screen touch frequency before the user scans the code, collected by the capacitive touch sensor in the sensor component of the offline terminal device; a, b, and c are touch frequency fitting coefficients, obtained by fitting based on the user's historical behavior data; and W is the implicit behavior effective weight calculated by the implicit behavior-environmental interference collaborative effective weight formula.

8. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 5, characterized in that, In step S4, the AI ​​optimization module iteratively optimizes the light interference calibration coefficient k1, the sound interference calibration coefficient k2, and the stability enhancement calibration coefficient k3 based on the correspondence between historical user behavior data and advertising results.

9. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 7, characterized in that, In step S4, the screen touch frequency T is collected by the capacitive touch sensor counting the number of times the user touches the screen of the offline terminal device within a timing period, and the number of touches is the screen touch frequency T.

10. The method for AI-based offline user behavior analysis and advertising strategy optimization according to claim 1, characterized in that, In step S4, the AI ​​optimization module calls the implicit behavior data, environmental interference data, and interaction data from the most recent collection period as iterative training data for each iteration.

Images