E-commerce operation analysis method based on artificial intelligence

Abstract

The invention discloses an e-commerce operation analysis method based on artificial intelligence, and the method comprises the following steps: S1, collecting multi-source data in an e-commerce platform, and carrying out the preprocessing of the multi-source data; s2, constructing an e-commerce operation analysis model; s3, performing global search on the hyper-parameter space of the e-commerce operation analysis model based on an artificial bee colony algorithm; s4, optimizing hyper-parameters by adopting a differential evolution algorithm, performing fine search, and dynamically adjusting the parameters; s5, designing a dynamic environment sensing module, and monitoring market dynamics and user behaviors in real time; s6, generating an e-commerce operation decision by using the fuzzy decision rule; and S7, applying an operation decision, monitoring an operation effect in real time, and feeding back the operation effect to the dynamic environment sensing module. According to the e-commerce operation analysis method based on artificial intelligence, efficient decision making and automatic optimization of the e-commerce platform in sales prediction, user behavior analysis and market dynamic response are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to an e-commerce operation analysis method based on artificial intelligence. Background Art

[0002] Currently, with the rapid development of Internet technology and big data analysis, the e-commerce field is undergoing unprecedented changes. Major e-commerce platforms are constantly exploring new methods to improve operation efficiency and user experience in the fierce market competition. Operation analysis based on artificial intelligence has gradually become a hot topic in research and application. Traditional e-commerce operation analysis methods mainly rely on statistical analysis, rule engines, and classic machine learning algorithms such as decision trees, support vector machines, and clustering analysis. These methods have high efficiency and stability in processing structured data, but they often seem powerless when facing large-scale, multi-source, and complex e-commerce data. Especially in the operation process of e-commerce platforms, sales data, user behavior data, inventory data, and market dynamics data have characteristics such as strong timeliness, diverse dimensions, and complex non-linear relationships. Traditional methods are difficult to capture the complex correlations between data in real time, thereby affecting the accuracy of decision-making and the real-time response ability.

[0003] In recent years, artificial intelligence technology has made remarkable progress in fields such as image recognition, natural language processing, and deep learning, and has gradually been applied to e-commerce operation analysis, attempting to achieve efficient prediction of sales trends, user preferences, and market dynamics through neural network models. However, the current existing technologies mainly focus on constructing prediction models using deep neural networks and optimizing parameters on this basis. Although the prediction accuracy can be improved to a certain extent, there are still many deficiencies. First, existing deep learning models often require a large amount of labeled data for training, and the data in e-commerce platforms has a high update frequency, complex data formats, and a lot of data noise, resulting in the model being easily interfered by noise during the training process, thereby affecting the stability of the prediction results. Second, most traditional model parameter optimization methods adopt fixed optimization strategies such as gradient descent or basic evolutionary algorithms, lacking an adaptive adjustment mechanism for the dynamic changes of real-time data, and unable to reflect the rapid changes in market fluctuations and user behavior in a timely manner, thus making it difficult for the prediction model's response speed and accuracy to meet the actual operation requirements.

[0004] In addition, there are obvious deficiencies in the current research on multi-objective optimization. E-commerce operation involves multiple objectives, such as maximizing sales, minimizing inventory costs, and enhancing user experience. There are complex relationships of mutual restriction and non-linear coupling among different objectives. Traditional optimization algorithms usually only focus on a single objective or use simple weighted methods for multi-objective optimization, making it difficult to balance various indicators, resulting in less-than-ideal decision-making effects in practical applications. More critically, the existing technologies lack research on the dynamic perception and real-time feedback mechanism of the e-commerce operation environment. During the e-commerce operation process, the market environment and user behavior are constantly changing. Traditional models often use static parameters for training and prediction, lacking means for real-time monitoring and adjustment of the dynamic environment. When facing emergencies or sharp market changes, the prediction ability and decision-making response of the models are both difficult to maintain stability.

[0005] Therefore, how to provide an e-commerce operation analysis method based on artificial intelligence is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose an e-commerce operation analysis method based on artificial intelligence. The present invention makes full use of deep learning technology, swarm intelligence optimization algorithms, dynamic environment perception modules, and self-supervised learning mechanisms, and details an algorithm for realizing efficient analysis and intelligent decision-making of e-commerce operation data through multi-source data preprocessing, deep model construction, global search and local optimization, dynamic adjustment of target weights, and self-supervised feedback. This method can capture market dynamics and user behavior changes in real time, comprehensively optimize multiple objectives such as sales trends, inventory management, and user experience within the e-commerce platform, thereby realizing refined and automated operation decisions, and having advantages such as fast response speed, high prediction accuracy, and strong adaptability.

[0007] The e-commerce operation analysis method based on artificial intelligence according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source data in the e-commerce platform and preprocess the multi-source data;

[0009] S2. Construct an e-commerce operation analysis model, which is used to predict sales trends, user behavior, and market dynamics. The e-commerce operation analysis model includes an input layer, a feature extraction layer, and a decision output layer;

[0010] S3. Based on the artificial bee colony algorithm, perform a global search on the hyperparameter space of the e-commerce operation analysis model. The employed bees randomly initialize the population individuals, calculate the fitness values, the onlooker bees select high-quality solutions according to probabilities, and the scout bees perform re-search operations in the area where the fitness value is lower than the threshold, and screen out a preliminary candidate solution set;

[0011] S4. Based on the obtained candidate solution set, use the differential evolution algorithm to optimize the hyperparameters of the e-commerce operation analysis model, conduct a fine search around the candidate solutions, and dynamically adjust the mutation factor and crossover probability;

[0012] S5. Design a dynamic environment perception module to monitor the market dynamics and user behavior changes in the e-commerce platform in real time. Combine with the fuzzy logic decision system to dynamically adjust the objective weights in the global search and local optimization processes during the operation of the e-commerce operation analysis model;

[0013] S6. Based on the optimized e-commerce operation analysis model, use the fuzzy decision rules to comprehensively evaluate multiple objectives, generate e-commerce operation decisions, and automatically correct and continuously optimize the parameters of the e-commerce operation analysis model by analyzing historical data and real-time feedback through the self-supervised learning module;

[0014] S7. Apply the generated e-commerce operation decisions to the actual operation process, monitor the execution effect of the e-commerce operation decisions in real time, and feedback the evaluation results to the dynamic environment perception module by collecting operation data and evaluating the actual effects.

[0015] Optionally, the S2 specifically includes:

[0016] S21. Classify and organize the multi-source data of the e-commerce platform after collection and preprocessing, clarify the data structure and dimensions of each type of data, and construct an input data set with a fixed format and multi-dimensional features;

[0017] S22. Based on the generated e-commerce operation data input set, construct the input layer of the e-commerce operation analysis model, and the input layer uniformly maps various types of original feature data to the internal data representation space;

[0018] S23. Construct a feature extraction layer on the basis of the input layer. The feature extraction layer conducts in-depth feature extraction for the sales trends, user behavior patterns, and market dynamics in e-commerce operations, and extracts information reflecting the core features of e-commerce operations from the input data through multi-layer non-linear mapping and data dimensionality reduction processing;

[0019] S24. On the feature representation output by the feature extraction layer, construct a decision output layer. The decision output layer further maps and integrates the extracted features, converts the feature information into specific e-commerce operation decision indicators, and forms a decision output with a clear structure and distinct levels;

[0020] S25. Integrate the input layer, feature extraction layer, and decision output layer constructed in steps S22 to S24 into a complete e-commerce operation analysis model, clarify the hyperparameters of each layer, and uniformly define the hyperparameters in the form of vectors or sets.

[0021] Optionally, the S3 specifically includes:

[0022] S31. Define the dynamic hyperparameter search space. The hyperparameter vector θ of the e-commerce operation analysis model is as follows:

[0023]

[0024] where W in and b in represent the weight and bias of the input layer respectively, W feat and b feat represent the weight and bias of the feature extraction layer respectively, W out and b out represent the weight and bias of the decision output layer respectively, ψ and χ represent the activation functions of each layer respectively. Define the dynamic hyperparameter search space Θ(t) as follows:

[0025] Θ(t) = {θ ∈ R k : θ min - αln(1 + t) ≤ θ ≤ θ max + βln(1 + t)};

[0026] where t represents the current iteration number, θ min and θ max are the initial lower and upper bounds of the hyperparameters respectively, α and β are positive real constants, k represents the dimension of the hyperparameters, ln is the logarithmic function, and R is the set of real numbers;

[0027] S32. The employed bees randomly initialize the population. According to the defined dynamic hyperparameter search space Θ(t), randomly generate a population P = {p1, p2,..., p N} under the e-commerce operation optimization requirements, where each candidate solution p i represents a set of model hyperparameters and is initially distributed by random number generation;

[0028] S33. For each candidate solution p i in the population, use the multi-objective fitness function to evaluate its performance in e-commerce operation prediction and decision-making:

[0029]

[0030] where F(p i ) represents the fitness of the candidate solution p i , E(p i ) represents the error of the candidate solution corresponding model in sales prediction, R(p i ) represents the error of the candidate solution corresponding model in operation metrics, M(p i ) represents the market response metric, ω1, ω2, and ω3 are weight coefficients, ∈ and δ ′To prevent small positive numbers from division by zero, κ is the steepness parameter of the response function, τ is the market response threshold, and exp() is the exponential function;

[0031] S34. According to the calculated fitness values, for each candidate solution p in the population i An improved probability selection mechanism is used to determine the probability of being selected, and the observing bees select high-quality solutions according to the probability:

[0032]

[0033] where K i represents the probability that the candidate solution p i is selected, λ is a positive real dynamic adjustment parameter, N represents the total number of candidate solutions in the population, λ is the dynamic adjustment parameter, N is the total number of candidate solutions in the population, δ is the exponential decay factor, and F(p j ) represents the fitness of the candidate solution p j ;

[0034] S35. The scouting bees perform a re-search operation in the area where the fitness value is lower than the threshold. For candidate solutions with fitness F(p i ) lower than the preset threshold F th , a re-search operation is performed in the area where it is located to generate an updated candidate solution p′ i . The updated candidate solution will replace the candidate solution in the original population that is lower than the preset threshold F th . Finally, a preliminary candidate solution set P final is formed:

[0035]

[0036] where p′ i represents the updated candidate solution, η is a random perturbation factor, θ max (t) and θ min (t) respectively represent the upper and lower limits of the hyperparameters of the dynamic search space at iteration t, T is the maximum number of iterations, p best is the candidate solution with the highest fitness in the current population, and ξ is a small positive number constant to prevent division by zero.

[0037] Optionally, the S4 specifically includes:

[0038] S41. For each candidate solution p in the screened preliminary candidate solution set P final , after randomly selecting three different candidate solutions p i ∈R k , p r1 , p r2 and p r3 , a mutation vector v i is generated:

[0039]

[0040] Among them, F(t) is the dynamic mutation factor, and η g is the gradient weight factor, represents the gradient vector of the candidate solution p i with respect to the fitness function, is the Euclidean norm of this gradient vector, and ξ g is a small positive number to prevent division by zero, and κ(t) is the chaos modulation function;

[0041] S42. For each candidate solution p i and the corresponding mutation vector v i perform a crossover operation to generate a trial vector u i , where each dimension j satisfies the following relationship:

[0042]

[0043] where u i,j represents the j-th component of the trial vector u i , v i,j and p i,j respectively represent the j-th components of the mutation vector v i and the candidate solution p i , r j is a random number, CR is the crossover probability, μ(t) is the chaos modulation factor, and j rand is a randomly selected index;

[0044] S43. Adopt a fusion geometric selection mechanism to fuse the trial vector u i and the original candidate solution p i through weighted geometric averaging to generate an updated candidate solution

[0045]

[0046] where F(u i ) and F(p i ) respectively represent the fitness values corresponding to the trial vector and the original candidate solution, exp() is the exponential function, ⊙ represents the element-wise product of corresponding elements, ln is the logarithmic function, and σ() represents the logistic function;

[0047] S44. Based on the complexity of e-commerce operation data and the dynamic changes in the environment, dynamically update the mutation factor F(t):

[0048]

[0049] Among them, F(t) represents the mutation factor at the current iteration t, F0 is the initial mutation factor, γ is a positive real number adjustment parameter, T is the maximum number of iterations, μ(t) is the chaotic modulation factor, and ΔE(p i ) represents the difference in fitness error between the candidate solution p i and the neighborhood, and E(p i ) is the fitness error of the candidate solution p i ; ∈ is a small positive number to prevent division by zero;

[0050] S45. Combine the updated candidate solutions to form the updated candidate solution set.

[0051] Optionally, the S5 specifically includes:

[0052] S51. Real-time collect the market dynamics and user behavior data of the e-commerce platform, construct the environmental index vector, and filter the original data to eliminate noise. The formula is

[0053]

[0054] where E(t) represents the environmental index vector at the current iteration t, and E i (t) represents the filtered value of the i-th environmental index at time t, represents the original environmental data collected at the past time t - kΔt, K is the filter window size, Δt is the time step, and n is the total number of environmental indices;

[0055] S52. Set the initial target weight vector according to the key indicators in e-commerce operation:

[0056]

[0057] where w 0 represents the initial target weight vector, represents the initial weight value corresponding to the i-th target, and m is the number of target indicators;

[0058] S53. Based on the constructed environmental index vector and the set initial target weight vector, update the target weight through the fuzzy logic decision mechanism:

[0059]

[0060] where w i (t) represents the updated weight of the i-th target at the current iteration t, represents the initial weight of the i-th target, α1 is a positive real number adjustment parameter, σ() is the logistic function, where λ is a positive real number parameter, and E i (t) represents the filtered value of the i-th environmental index at iteration t, τi represents the preset threshold of the i-th environmental indicator, n is the total number of environmental indicators, and E j (t) represents the filtered value of the j-th environmental indicator at iteration t, and τ j represents the preset threshold of the j-th environmental indicator;

[0061] S54. To further adapt to the complexity of indicator changes in the e-commerce operation environment, dynamically update the target weight adjustment factor:

[0062]

[0063] where ω(t) represents the dynamic adjustment factor at iteration t, ω0 is the initial adjustment factor, and λ ′ is the attenuation parameter, T is the maximum number of iterations, α2 is the amplification factor, and α3 is the steepness adjustment parameter of the exponential function, is the average value of environmental indicators, and τ E is the preset threshold of the environmental average value, and exp() is the exponential function;

[0064] S55. Integrate the updated target weight with the dynamic adjustment factor to generate the final target weight vector applicable to the e-commerce operation analysis model:

[0065] w(t) = α(t) ⊙ w(t - 1) + [1 - α(t)] ⊙ [ω(t) · w i (t)];

[0066] where w(t) represents the finally updated target weight vector at iteration t, w(t - 1) represents the target weight vector at the previous iteration, α(t) is the dynamic integration factor, and ⊙ represents the element-wise product operation.

[0067] Optionally, the S6 specifically includes:

[0068] S61. Use the optimized e-commerce operation analysis model to process the input data set X constructed in step S1 to generate a predicted output vector:

[0069]

[0070] where Y represents the predicted output, and respectively represent the weights and biases of the optimized input layer, and respectively represent the weights and biases of the optimized feature extraction layer, and respectively represent the weights and biases of the optimized decision output layer, ψ and respectively represent the activation functions of each layer;

[0071] S62. Calculate the evaluation function for the key objectives in e-commerce operations based on the predicted output Y, and construct the target index vector:

[0072] F(Y) = {f i (Y) | i = 1, 2, …, B};

[0073] Among them, F(Y) represents the target index vector, and f i (Y) represents the evaluation function value of the i-th target, and B is the number of target evaluation indicators;

[0074] S63. Adopt a fusion-based multi-dimensional non-linear fuzzy decision-making mechanism to fuse the target index and the environmental weight to generate the e-commerce operation decision vector:

[0075]

[0076] Among them, D represents the generated e-commerce operation decision vector, w i (t) represents the dynamic weight of the i-th target at the current iteration t, and f i (Y) represents the evaluation function value of the i-th target, is the i-th modulation coefficient, ψ i is the i-th activation adjustment parameter, τ i is the preset threshold for the i-th target evaluation, ln() and exp() are the natural logarithm and exponential function respectively, and K is the number of target evaluation indicators;

[0077] S64. Construct a self-supervised learning module to adaptively correct the optimized e-commerce operation analysis model through the fusion of fuzzy gradient feedback:

[0078]

[0079] Among them, Δθ represents the hyperparameter correction term, η sl represents the learning rate of self-supervised learning, D real represents the actual operation effect vector, tanh() represents the hyperbolic tangent function, A is the fuzzy gradient adjustment coefficient, Γ is the activation adjustment parameter, exp() represents the exponential function, ρ is the steepness adjustment parameter of the logistic function, G is the additional modulation factor, and ⊙ represents the element-wise multiplication operation;

[0080] S65. Apply the calculated hyperparameter correction term Δθ to the hyperparameters θ * of the currently optimized e-commerce operation analysis model to generate updated hyperparameters.

[0081] The beneficial effects of the present invention are:

[0082] By comprehensively utilizing advanced technologies such as deep learning, swarm intelligence optimization, dynamic environment perception, and self-supervised learning, the present invention constructs a closed-loop e-commerce operation analysis system to achieve efficient processing of massive and multi-source e-commerce data and intelligent decision-making. Compared with traditional methods, the deep model adopted by the present invention can fully exploit the complex non-linear relationships hidden in e-commerce data, effectively capture sales trends, inventory dynamics, and user behavior patterns, and achieve seamless connection between global search and local fine optimization through the artificial bee colony algorithm and differential evolution algorithm. This optimization strategy can quickly locate the optimal hyperparameter combination in massive data, avoiding the defect of traditional optimization methods being prone to falling into local optima, thus significantly improving the accuracy and robustness of model prediction.

[0083] Meanwhile, the present invention introduces a dynamic environment perception module and a fuzzy decision-making mechanism, enabling the system to monitor market changes and user behavior fluctuations in the e-commerce platform in real time, and dynamically adjust the weights of various target indicators (such as sales prediction accuracy, inventory management, and user experience, etc.). Through real-time collection and filtering of environmental indicators, the system constructs a standardized environmental indicator vector, and uses a fusion multi-dimensional fuzzy decision-making algorithm to non-linearly integrate the target weights. This method can balance multiple targets and improve the response speed and decision-making accuracy of the decision-making system in the face of sudden market changes. In addition, the introduction of the self-supervised multi-scale fuzzy gradient feedback mechanism enables the model to automatically correct and optimize hyperparameters according to the actual operation effect in each iteration process, forming a closed-loop adaptive optimization process, thereby ensuring the stability and continuous improvement ability of the system in long-term operation.

[0084] Through the close cooperation of the above-mentioned links, the present invention realizes the all-round intelligent analysis and decision support for e-commerce operation data. The system can not only achieve real-time data processing and prediction in a high-concurrency and changeable operation environment, but also has the remarkable advantages of adaptive adjustment, intelligent feedback, and multi-objective comprehensive optimization. This method shows excellent performance in improving sales prediction accuracy, optimizing inventory management, improving user experience, and enhancing overall operation efficiency, significantly overcoming the defects of slow response speed, fixed optimization strategy, and unreasonable multi-objective weight allocation in traditional methods. Through automatic correction and continuous optimization, the system can continuously improve its prediction and decision-making ability, providing strong technical support and data guarantee for e-commerce platforms in the fierce market competition, and ultimately achieving a significant improvement in operation efficiency and a steady enhancement of market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0086] Figure 1 Flow chart of the e-commerce operation analysis method based on artificial intelligence proposed by the present invention;

[0087] Figure 2 Structural schematic diagram of the e-commerce operation analysis model of the e-commerce operation analysis method based on artificial intelligence proposed by the present invention. Detailed implementation manners

[0088] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0089] Refer to Figure 1 and Figure 2 , the e-commerce operation analysis method based on artificial intelligence includes the following steps:

[0090] S1. Collect multi-source data in the e-commerce platform and preprocess the multi-source data;

[0091] S2. Construct an e-commerce operation analysis model, which is used to predict sales trends, user behaviors and market dynamics. The e-commerce operation analysis model includes an input layer, a feature extraction layer, and a decision output layer;

[0092] S3. Based on the artificial bee colony algorithm, globally search the hyperparameter space of the e-commerce operation analysis model. The employed bees randomly initialize the population individuals, calculate the fitness values, the onlooker bees select high-quality solutions according to probabilities, and the scout bees perform re-search operations in the area where the fitness value is lower than the threshold, and screen out the preliminary candidate solution sets;

[0093] S4. On the basis of the obtained candidate solution sets, use the differential evolution algorithm to optimize the hyperparameters of the e-commerce operation analysis model, perform fine search around the candidate solutions, and dynamically adjust the mutation factor and crossover probability;

[0094] S5. Design a dynamic environment perception module to monitor the market dynamics and user behavior changes in the e-commerce platform in real time. Combining with the fuzzy logic decision system, during the operation of the e-commerce operation analysis model, dynamically adjust the objective weights in the global search and local optimization processes;

[0095] S6. Based on the optimized e-commerce operation analysis model, use fuzzy decision rules to comprehensively evaluate multiple objectives, generate e-commerce operation decisions, and automatically correct and continuously optimize the parameters of the e-commerce operation analysis model by analyzing historical data and real-time feedback through the self-supervised learning module;

[0096] S7. Apply the generated e-commerce operation decisions to the actual operation process, monitor the execution effect of the e-commerce operation decisions in real time, and feedback the evaluation results to the dynamic environment perception module by collecting operation data and evaluating the actual effects.

[0097] In this embodiment, S2 specifically includes:

[0098] S21. Classify and organize the multi-source data of the e-commerce platform after collection and preprocessing, clarify the data structure and dimensions of each type of data, and construct an input data set with a fixed format and multi-dimensional features;

[0099] S22. Based on the generated input set of e-commerce operation data, construct the input layer of the e-commerce operation analysis model, and the input layer uniformly maps various types of original feature data into the internal data representation space;

[0100] S23. Construct a feature extraction layer on the basis of the input layer. The feature extraction layer performs in-depth feature extraction on the sales trend, user behavior pattern, and market dynamics in e-commerce operation. Through multi-layer non-linear mapping and data dimensionality reduction processing, extract information reflecting the core features of e-commerce operation from the input data;

[0101] S24. On the feature representation output by the feature extraction layer, construct a decision output layer. The decision output layer further maps and integrates the extracted features, converts the feature information into specific e-commerce operation decision indicators, and forms a decision output with a clear structure and distinct levels;

[0102] S25. Integrate the input layer, feature extraction layer, and decision output layer constructed in steps S22 to S24 into a complete e-commerce operation analysis model, clarify the hyperparameters of each layer, and uniformly define the hyperparameters in the form of a vector or a set.

[0103] In this embodiment, S3 specifically includes:

[0104] S31. Define the dynamic hyperparameter search space. The hyperparameter vector θ of the e-commerce operation analysis model is:

[0105]

[0106] where W in and b in respectively represent the weight and bias of the input layer, W feat and b feat respectively represent the weight and bias of the feature extraction layer, W out and b out respectively represent the weight and bias of the decision output layer, ψ and χ respectively represent the activation functions of each layer. Define the dynamic hyperparameter search space Θ(t) as:

[0107] Θ(t) = {θ ∈ R k : θ min - αln(1 + t) ≤ θ ≤ θ max + βln(1 + t)};

[0108] Among them, t represents the current iteration number, θ min and θ max are respectively the lower bound and upper bound of the initial hyperparameters, α and β are positive real constant, k represents the dimension of the hyperparameters, ln is the logarithmic function, and R is the set of real numbers;

[0109] S32. The employed bees randomly initialize the population. According to the defined dynamic hyperparameter search space Θ(t), randomly generate a population P = {p1, p2,..., p N} under the e-commerce operation optimization requirements, where each candidate solution p i represents a set of model hyperparameters and is initially distributed by random number generation;

[0110] S33. For each candidate solution p i in the population, use the multi-objective fitness function to evaluate its performance in e-commerce operation prediction and decision-making:

[0111]

[0112] Among them, F(p i ) represents the fitness of the candidate solution p i , E(p i ) represents the error of the candidate solution corresponding model in sales prediction, R(p i ) represents the error of the candidate solution corresponding model in operation indicators, M(p i ) represents the market response index, ω1, ω2, and ω3 are weight coefficients, ∈ and δ ′ are small positive numbers to prevent division by zero, κ is the steepness parameter of the response function, τ is the market response threshold, and exp() is the exponential function;

[0113] S34. According to the calculated fitness values, determine the selected probability for each candidate solution p i in the population using an improved probability selection mechanism, and the observing bees select high-quality solutions according to the probability:

[0114]

[0115] Among them, K i represents the selected probability of the candidate solution p i , λ is a positive real dynamic adjustment parameter, N represents the total number of candidate solutions in the population, λ is a dynamic adjustment parameter, N is the total number of candidate solutions in the population, δ is an exponential decay factor, and F(p j ) represents the fitness of the candidate solution p j ;

[0116] S35. The scout bees perform a re-search operation in the area where the fitness value is lower than the threshold, and for the fitness F(p iCandidate solutions lower than the preset threshold F th perform a resampling operation within the region where they are located to generate updated candidate solutions p′ i , and the updated candidate solutions will replace the candidate solutions lower than the preset threshold F th in the original population, and finally form a preliminary candidate solution set P final :

[0117]

[0118] where p′ i represents the updated candidate solution, η is a random perturbation factor, θ max (t) and θ min (t) represent the upper and lower bounds of the hyperparameters of the dynamic search space at iteration t respectively, T is the maximum number of iterations, p best is the candidate solution with the highest fitness in the current population, and ξ is a small positive constant to prevent division by zero.

[0119] In this embodiment, S4 specifically includes:

[0120] S41. For each candidate solution p final in the selected preliminary candidate solution set P i ∈R k , randomly select three different candidate solutions p r1 , p r2 and p r3 , and then generate a mutation vector v i :

[0121]

[0122] where F(t) is a dynamic mutation factor, η g is a gradient weight factor, represents the gradient vector of the candidate solution p i with respect to the fitness function, is the Euclidean norm of this gradient vector, ξ g is a small positive number to prevent division by zero, and κ(t) is a chaotic modulation function;

[0123] S42. Perform a crossover operation on each candidate solution p i and the corresponding mutation vector v i to generate a trial vector u i , where each dimension j satisfies the following relationship:

[0124]

[0125] where u i,j represents the j-th component of the trial vector u i , v i,j and pi,j respectively represent the mutant vector v i and the j-th component of the candidate solution p i , r j is a random number, CR is the crossover probability, μ(t) is the chaotic modulation factor, j rand is a randomly selected index;

[0126] S43. Adopt a fusion geometric selection mechanism to fuse the trial vector u i and the original candidate solution p i to generate an updated candidate solution through weighted geometric averaging

[0127]

[0128] where F(u i ) and F(p i ) respectively represent the fitness values corresponding to the trial vector and the original candidate solution, exp() is the exponential function, ⊙ is the element-wise product indicating multiplication of corresponding elements, ln is the logarithmic function, and σ() represents the logistic function;

[0129] S44. Based on the complexity of e-commerce operation data and the dynamic changes of the environment, dynamically update the mutation factor F(t):

[0130]

[0131] where F(t) represents the mutation factor at the current iteration t, F0 is the initial mutation factor, γ is a positive real adjustment parameter, T is the maximum number of iterations, μ(t) is the chaotic modulation factor, ΔE(p i ) represents the difference between the candidate solution p i and the fitness error within the neighborhood, E(p i ) is the fitness error of the candidate solution p i , and ∈ is a small positive number to prevent division by zero;

[0132] S45. Combine the updated candidate solutions to form an updated candidate solution set uniformly.

[0133] In this embodiment, the specific steps of S5 include:

[0134] S51. Real-time collect the market dynamics and user behavior data of the e-commerce platform, construct an environmental index vector, and filter the original data to eliminate noise. The formula is

[0135]

[0136] where E(t) represents the environmental index vector at the current iteration t, E iThe filtered value of the i-th environmental indicator at time t is denoted as (t), which represents the original environmental data collected at the past time t - kΔt. K is the size of the filtering window, Δt is the time step, and n is the total number of environmental indicators;

[0137] S52. Set the initial target weight vector according to the key indicators in e-commerce operation:

[0138]

[0139] where, w 0 represents the initial target weight vector, represents the initial weight value corresponding to the i-th target, and m is the number of target indicators;

[0140] S53. Based on the constructed environmental indicator vector and the set initial target weight vector, update the target weight through the fuzzy logic decision-making mechanism:

[0141]

[0142] where, w i (t) represents the updated weight of the i-th target at the current iteration t, represents the initial weight of the i-th target, α1 is a positive real number adjustment parameter, σ() is the logistic function, where λ is a positive real number parameter, E i (t) represents the filtered value of the i-th environmental indicator at iteration t, τ i represents the preset threshold of the i-th environmental indicator, n is the total number of environmental indicators, E j (t) represents the filtered value of the j-th environmental indicator at iteration t, τ j represents the preset threshold of the j-th environmental indicator;

[0143] S54. To further adapt to the complexity of indicator changes in the e-commerce operation environment, dynamically update the target weight adjustment factor:

[0144]

[0145] where, ω(t) represents the dynamic adjustment factor at iteration t, ω0 is the initial adjustment factor, λ ′ is the decay parameter, T is the maximum number of iterations, α2 is the amplification coefficient, α3 is the steepness adjustment parameter of the exponential function, is the average value of environmental indicators, τ E is the preset threshold of the environmental average value, and exp() is the exponential function;

[0146] S55. Integrate the updated target weight and the dynamic adjustment factor to generate the final target weight vector applicable to the e-commerce operation analysis model:

[0147] w(t) = α(t) ⊙ w(t - 1) + [1 - α(t)] ⊙ [ω(t) · w i (t)];

[0148] Among them, w(t) represents the finally updated target weight vector at iteration t, w(t - 1) represents the target weight vector at the previous iteration, α(t) is the dynamic fusion factor, and ⊙ represents the element-wise product operation.

[0149] In this embodiment, S6 specifically includes:

[0150] S61. Use the optimized e-commerce operation analysis model to process the input data set X constructed in step S1 to generate a predicted output vector:

[0151]

[0152] Among them, Y represents the predicted output, and respectively represent the weights and biases of the optimized input layer, and respectively represent the weights and biases of the optimized feature extraction layer, and respectively represent the weights and biases of the optimized decision output layer, ψ and respectively represent the activation functions of each layer;

[0153] S62. Based on the predicted output Y, calculate the evaluation function for the key objectives in e-commerce operations and construct the target index vector:

[0154] F(Y) = {f i (Y) | i = 1, 2,..., B};

[0155] Among them, F(Y) represents the target index vector, f i (Y) represents the value of the i-th target evaluation function, and B is the number of target evaluation indicators;

[0156] S63. Adopt a fusion-based multi-dimensional non-linear fuzzy decision-making mechanism to fuse the target index and the environmental weight to generate an e-commerce operation decision vector:

[0157]

[0158] Among them, D represents the generated e-commerce operation decision vector, w i (t) represents the dynamic weight of the i-th target at the current iteration t, f i (Y) represents the value of the i-th target evaluation function, is the i-th modulation coefficient, ψi is the i-th activation adjustment parameter, τ i is the preset threshold for the i-th target evaluation. ln() and exp() are the natural logarithm and exponential function respectively, and K is the number of target evaluation indicators;

[0159] S64. Construct a self-supervised learning module to adaptively correct the optimized e-commerce operation analysis model by fusing fuzzy gradient feedback:

[0160]

[0161] where Δθ represents the hyperparameter correction term, η sl represents the learning rate of self-supervised learning, D real represents the actual operation effect vector, tanh() represents the hyperbolic tangent function, A is the fuzzy gradient adjustment coefficient, Γ is the activation adjustment parameter, exp() represents the exponential function, ρ is the steepness adjustment parameter of the logistic function, G is the additional modulation factor, and ⊙ represents the element-wise multiplication operation;

[0162] S65. Apply the calculated hyperparameter correction term Δθ to the hyperparameters θ of the currently optimized e-commerce operation analysis model * to generate the updated hyperparameters.

[0163] Example 1:

[0164] To verify the feasibility of the present invention in implementation, the present invention is applied to a large e-commerce platform. Due to the huge daily transaction volume, the sales data, user behavior data, inventory data, and market dynamic data generated by the platform reach tens of millions of records every day. When traditional statistical analysis or simple machine learning methods are used to process these data, there are often problems such as slow data processing speed, low prediction accuracy, and inability to respond to market changes in a timely manner. The platform used a traditional prediction model in the past, mainly relying on historical data to establish a prediction model with fixed parameters. As a result, when facing promotional activities, holidays, or emergencies, the model cannot quickly adapt to the new environment, and the prediction results often deviate. The inventory management and promotional strategy adjustments lag, directly affecting the overall operation efficiency and user experience of the platform.

[0165] In this embodiment, in response to the above problems of the platform, we implemented an e-commerce operation analysis method based on artificial intelligence. This method first deploys a data collection system in the main data center of the platform to collect multi-source data such as sales, user clicks, browsing, placing orders, and logistics distribution in real time, and processes the data through a data preprocessing module for cleaning, denoising, and normalization to generate a standardized input data set. Subsequently, a deep learning model is used to construct an e-commerce operation analysis model, which includes an input layer, a feature extraction layer, and a decision output layer. The model parameters are automatically optimized through an initial global search and local optimization algorithm to ensure that the model can accurately capture key information such as sales trends, changes in user behavior, and inventory status.

[0166] In practical applications, the platform comprehensively monitors data on daily operations by deploying the method of the present invention. For example, during the period from November 2024 to January 2025, when the platform carried out Double Eleven and Spring Festival promotion activities, the method of the present invention was used to monitor sales data and inventory status in real time, and the promotion strategy and inventory control plan were automatically adjusted according to the prediction results. Compared with the traditional method, after adopting the method of the present invention, the sales prediction accuracy of the platform has increased from 78% to 92%, the inventory turnover speed has increased by about 25%, and the user page response time has been shortened by nearly 40%. For example, on November 15, 2024, during the peak promotion period, the data volume processed by this system reached 120 million records per day. Through a closed-loop self-supervised feedback mechanism, the model completed an adaptive parameter update within 15 minutes, making subsequent predictions more accurate, thus greatly improving the response speed of the promotion activities and promoting the sales volume to increase by more than 35% year-on-year.

[0167] In addition, the dynamic environment perception module in the method of the present invention can obtain data on environmental changes inside and outside the platform in real time, such as competitor promotion information, user evaluation feedback, and logistics distribution status, so as to dynamically adjust the weights of each operation target. The actual operation data shows that during December 2024, this module updated the target weights once per hour, enabling the decision output to reflect market changes in a timely manner and effectively reducing the risk of inventory backlog. For example, through comparative analysis, before adopting this method, the platform had a high lag in inventory management, with an average inventory backlog rate of 18%. After adopting the method of the present invention, this ratio dropped to 12%, the inventory cost was significantly reduced, and the user satisfaction was improved.

[0168] Meanwhile, this embodiment utilizes a self-supervised multi-scale fuzzy gradient feedback mechanism to achieve automatic calibration of model parameters. The system compares the current prediction decision with the actual operation results every certain period (such as every 2 hours), and automatically adjusts the model parameters through error feedback to form a closed-loop adaptive optimization process. Experimental data shows that during the continuous operation in December 2024, after self-supervised update, the average prediction error of the model decreased by 15%, enabling the platform to predict abnormal sales fluctuations 10 - 15 minutes in advance during peak periods, thereby deploying emergency measures in advance and effectively avoiding operation risks caused by prediction lag.

[0169] Table 1 Comparison Table of Operational Data Analysis Effects of Platform A

[0170]

[0171] According to the data in Table 1, it can be seen that the e-commerce operation analysis method based on artificial intelligence of the present invention has shown remarkable effects in practical applications. In terms of sales prediction, the prediction accuracy of the method of the present invention has been increased from 78% to 92%, an increase of 14 percentage points, indicating more accurate prediction of sales trends during peak periods. In inventory management, the inventory turnover speed has been increased from 1.8 days per time to 1.35 days per time, an increase of 25%, effectively reducing inventory costs and improving the capital turnover efficiency.

[0172] In terms of user experience, this method has shortened the page response time from 3.2 seconds to 1.9 seconds, with a speed increase of 40%, significantly enhancing the browsing experience of users and reducing the user churn rate caused by slow loading. In promotional activities, this method has increased the sales growth rate from 20% of the traditional method to 35%, an increase of 15 percentage points, effectively improving the promotion conversion rate through dynamic decision-making. In addition, this method realizes automatic update of model parameters every 2 hours, can respond to market changes in real time, and can predict abnormal sales fluctuations 10 - 15 minutes in advance, effectively reducing operation risks.

[0173] Generally speaking, through these data analyses, it can be seen that the e-commerce operation analysis method of the present invention not only performs excellently in specific operation indicators such as sales prediction, inventory management, and user experience, but also has obvious technical advantages in real-time response and risk prevention and control. Through the combination of multiple innovative technologies, the method of the present invention has achieved beneficial effects of accurate prediction, rapid response, and adaptive adjustment in the operation of e-commerce platforms, greatly improving the operation efficiency and market competitiveness of the platforms, and providing an intelligent, automated, and efficient operation analysis solution for the e-commerce industry.

[0174] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. An e-commerce operation analysis method based on artificial intelligence, characterized in that, It includes the following steps: S1. Collect multi-source data in the e-commerce platform and preprocess the multi-source data; S2. Construct an e-commerce operation analysis model, which is used to predict sales trends, user behaviors and market dynamics. The e-commerce operation analysis model includes an input layer, a feature extraction layer, and a decision output layer; S3. Conduct a global search on the hyperparameter space of the e-commerce operation analysis model based on the artificial bee colony algorithm. The employed bees randomly initialize the population individuals, calculate the fitness values, the onlooker bees select high-quality solutions according to probabilities, and the scout bees perform re-search operations in the area where the fitness value is lower than the threshold, and filter out a preliminary candidate solution set; S4. On the basis of the obtained candidate solution set, use the differential evolution algorithm to optimize the hyperparameters of the e-commerce operation analysis model, conduct a fine search around the candidate solutions, and dynamically adjust the mutation factor and crossover probability; S5. Design a dynamic environment perception module to monitor the market dynamics and changes in user behaviors in the e-commerce platform in real time. Combining with the fuzzy logic decision system, during the operation of the e-commerce operation analysis model, dynamically adjust the objective weights in the global search and local optimization processes; S6. Based on the optimized e-commerce operation analysis model, use fuzzy decision rules to comprehensively evaluate multiple objectives, generate e-commerce operation decisions, and automatically correct and continuously optimize the parameters of the e-commerce operation analysis model by analyzing historical data and real-time feedback through the self-supervised learning module; S7. Apply the generated e-commerce operation decisions to the actual operation process, monitor the execution effect of the e-commerce operation decisions in real time, and feedback the evaluation results to the dynamic environment perception module by collecting operation data and evaluating the actual effects.

2. The e-commerce operation analysis method based on artificial intelligence according to claim 1, wherein The specific content of S2 includes: S21. Classify and sort out the multi-source data of the e-commerce platform after collection and preprocessing, clarify the data structures and dimensions of various types of data, and construct an input data set with a fixed format and multi-dimensional features; S22. Based on the generated e-commerce operation data input set, construct the input layer of the e-commerce operation analysis model, and the input layer uniformly maps various types of original feature data into the internal data representation space; S23. On the basis of the input layer, construct a feature extraction layer. The feature extraction layer conducts in-depth feature extraction for sales trends, user behavior patterns and market dynamics in e-commerce operations. Through multi-layer non-linear mapping and data dimensionality reduction processing, extract information reflecting the core features of e-commerce operations from the input data; S24. On the feature representation output by the feature extraction layer, construct a decision output layer. The decision output layer further maps and integrates the extracted features, converts the feature information into specific e-commerce operation decision indicators, and forms a decision output with clear structure and distinct levels; S25. Integrate the input layer, feature extraction layer and decision output layer constructed in steps S22 to S24 into a complete e-commerce operation analysis model, clarify the hyperparameters of each layer, and uniformly define the hyperparameters in the form of vectors or sets.

3. The e-commerce operation analysis method based on artificial intelligence according to claim 1, wherein The specific content of S3 includes: S31. Define the dynamic hyperparameter search space. The hyperparameter vector θ of the e-commerce operation analysis model is: Among them, W in and b in represent the weights and biases of the input layer respectively. W feat and b feat represent the weights and biases of the feature extraction layer respectively. W out and b out represent the weights and biases of the decision output layer respectively. ψ and χ represent the activation functions of each layer respectively. Define the dynamic hyperparameter search space Θ(t) as: Θ(t) = {θ ∈ R k : θ min -α ln(1 + t) ≤ θ ≤ θ max +β ln(1 + t)}; where \(t\) represents the current iteration number, \(\theta\) min and \(\theta\) max are respectively the initial lower and upper bounds of the hyperparameters, \(\alpha\) and \(\beta\) are positive real constants, \(k\) represents the dimension of the hyperparameters, \(\ln\) is the logarithmic function, and \(\mathbb{R}\) is the set of real numbers; S32. The employed bees randomly initialize the population and randomly generate a population \(P = \{p_1, p_2, \ldots, p\}\) according to the defined dynamic hyperparameter search space \(\Theta(t)\) under the e-commerce operation optimization requirements. Each candidate solution \(p\) represents a set of model hyperparameters and is initially distributed by means of random number generation; N} where each candidate solution \(p\) i represents a set of model hyperparameters and is initially distributed by means of random number generation; S33. For each candidate solution p in the population i , evaluate its performance in e-commerce operation prediction and decision-making using the multi-objective fitness function: Among them, F(p i ) represents the fitness of the candidate solution p i , E(p i ) represents the error of the model corresponding to the candidate solution in sales prediction, R(p i ) represents the error of the model corresponding to the candidate solution in operation metrics, M(p i ) represents the market response metric, ω1, ω2, and ω3 are weight coefficients, ∈ and δ ′ are small positive numbers to prevent division by zero, κ is the steepness parameter of the response function, τ is the market response threshold, and exp() is the exponential function; S34. According to the calculated fitness value, for each candidate solution p in the population i An improved probability selection mechanism is used to determine the probability of being selected, and the observing bee selects high-quality solutions according to the probability: Among them, K i represents the probability that the candidate solution p i is selected. λ is a positive real dynamic adjustment parameter, N represents the total number of candidate solutions in the population, λ is the dynamic adjustment parameter, N is the total number of candidate solutions in the population, δ is the exponential decay factor, and F(p j ) represents the fitness of the candidate solution p j ; S35. The scouting bees perform a re-search operation in the area where the fitness value is lower than the threshold, and for the candidate solution with fitness F(p i ) lower than the preset threshold F th , a re-search operation is performed within the area to generate an updated candidate solution p′ i . The updated candidate solution will replace the candidate solution in the original population that is lower than the preset threshold F th , and finally a preliminary candidate solution set P final is formed: Among them, p' i represents the updated candidate solution, η is the random perturbation factor, and θ max (t) and θ min (t) represent the upper and lower bounds of the hyperparameters of the dynamic search space at iteration t respectively, T is the maximum number of iterations, p best is the candidate solution with the highest fitness in the current population, and ξ is a small positive constant to prevent division by zero.

4. The e-commerce operation analysis method based on artificial intelligence according to claim 1, characterized in that The specific content of S4 includes: S41. For each candidate solution p in the initially selected candidate solution set P final where p i ∈R k , randomly select three different candidate solutions p r1 , p r2 and p r3 , and then generate a mutation vector v i : Among them, F(t) is the dynamic mutation factor, η g is the gradient weight factor, represents the gradient vector of the candidate solution p i with respect to the fitness function, is the Euclidean norm of the gradient vector, ξ g is a small positive number to prevent division by zero, and κ(t) is the chaotic modulation function; S42. For each candidate solution p i and the corresponding mutation vector v i perform a crossover operation to generate a trial vector u i , where each dimension j satisfies the following relationship: where, u i,j represents the j-th component of the test vector u i , v i,j and p i,j represent the j-th components of the mutant vector v i and the candidate solution p i respectively, r j is a random number, CR is the crossover probability, μ(t) is the chaos modulation factor, and j rand is a randomly selected index; S43. Adopt a fusion-based geometric selection mechanism to fuse the test vector u i with the original candidate solution p i to generate an updated candidate solution through weighted geometric averaging where F(u i ) and F(p i ) represent the fitness values corresponding to the test vector and the original candidate solution respectively, exp() is the exponential function, ⊙ is the element-wise product representing the multiplication of corresponding elements, ln is the logarithmic function, and σ() represents the logistic function; Based on the complexity of e-commerce operation data and the dynamic changes in the environment, the mutation factor F(t) is dynamically updated: Among them, F(t) represents the mutation factor at the current iteration t, F0 is the initial mutation factor, γ is a positive real number adjustment parameter, T is the maximum number of iterations, μ(t) is the chaotic modulation factor, and ΔE(p i ) represents the difference in fitness error between the candidate solution p i and the neighborhood, E(p i ) is the fitness error of the candidate solution p i , and ∈ is a small positive number to prevent division by zero; S45. Combine the updated candidate solution p i m to form the updated candidate solution set uniformly.

5. The e-commerce operation analysis method based on artificial intelligence according to claim 1, wherein The specific steps of S5 are as follows: S51: Real-time collect the market dynamics and user behavior data of the e-commerce platform, construct an environmental index vector, and filter the original data to eliminate noise. The formula is Among them, \(E(t)\) represents the environmental index vector at the current iteration \(t\), and \(E\) i (t) represents the filtered value of the \(i\)-th environmental index at time \(t\), represents the original environmental data collected at the past time \(t - k\Delta t\), \(K\) is the size of the filtering window, \(\Delta t\) is the time step, and \(n\) is the total number of environmental indices; S52: Set the initial target weight vector according to the key indicators in e-commerce operation: Among them, w 0 represents the initial target weight vector, represents the initial weight value corresponding to the i-th target, and m is the number of target indicators; S53: Based on the constructed environmental index vector and the set initial target weight vector, update the target weight through a fuzzy logic decision-making mechanism: where, w i (t) represents the updated weight of the i-th target at the current iteration t, represents the initial weight of the i-th target, α1 is a positive real-valued adjustment parameter, σ() is the logistic function, where λ is a positive real-valued parameter, E i (t) represents the filtered value of the i-th environmental indicator at iteration t, τ i represents the preset threshold of the i-th environmental indicator, n is the total number of environmental indicators, E j (t) represents the filtered value of the j-th environmental indicator at iteration t, τ j represents the preset threshold of the j-th environmental indicator; S54: To further adapt to the complexity of index changes in the e-commerce operation environment, dynamically update the target weight adjustment factor: Among them, ω(t) represents the dynamic adjustment factor at iteration t, ω0 is the initial adjustment factor, λ ′ is the attenuation parameter, T is the maximum number of iterations, α2 is the amplification coefficient, α3 is the steepness adjustment parameter of the exponential function, is the average environmental index, τ E is the preset threshold of the environmental average, and exp() is the exponential function; S55: Integrate the updated target weight with the dynamic adjustment factor to generate a target weight vector finally applicable to the e-commerce operation analysis model: w(t) = α(t) ⊙ w(t - 1) + [1 - α(t)] ⊙ [ω(t) · w i (t)]; Among them, w(t) represents the finally updated target weight vector at iteration t, w(t - 1) represents the target weight vector at the previous iteration, α(t) is the dynamic integration factor, and ⊙ represents the element-wise product operation.

6. The e-commerce operation analysis method based on artificial intelligence according to claim 1, wherein The specific steps of S6 are as follows: S61: Use the optimized e-commerce operation analysis model to process the input data set X constructed in step S1 to generate a predicted output vector: where Y represents the predicted output, and represent the weight and bias of the optimized input layer respectively, and represent the weight and bias of the optimized feature extraction layer respectively, and represent the weight and bias of the optimized decision output layer respectively, ψ and θ represent the activation functions of each layer respectively; S62: Based on the predicted output Y, calculate the evaluation function for the key objectives in e-commerce operation and construct a target index vector: F(Y) = {f i (Y) | i = 1, 2, …, B}; Among them, F(Y) represents the target index vector, and f i (Y) represents the value of the i-th target evaluation function, and B is the number of target evaluation indicators; S63: Adopt a fusion multi-dimensional non-linear fuzzy decision-making mechanism to integrate the target index with the environmental weight to generate an e-commerce operation decision vector: Among them, D represents the generated e-commerce operation decision vector, w i (t) represents the dynamic weight of the i-th objective at the current iteration t, f i (Y) represents the evaluation function value of the i-th objective, is the i-th modulation coefficient, ψ i is the i-th activation adjustment parameter, τ i is the preset threshold for the evaluation of the i-th objective, ln() and exp() are the natural logarithm and exponential functions respectively, and K is the number of objective evaluation indicators; S64: Construct a self-supervised learning module to adaptively correct the optimized e-commerce operation analysis model through the integration of fuzzy gradient feedback: where, Δθ represents the hyperparameter correction term, η sl represents the learning rate of self-supervised learning, D real represents the actual operation effect vector, tanh() represents the hyperbolic tangent function, A is the fuzzy gradient adjustment coefficient, Γ is the activation adjustment parameter, exp() represents the exponential function, ρ is the steepness adjustment parameter of the logistic function, G is the additional modulation factor, and ⊙ represents the element-wise multiplication operation; S65. Apply the calculated hyperparameter correction term Δθ to the hyperparameter θ of the currently optimized e-commerce operation analysis model * to generate the updated hyperparameter.

Images