Marketing customer portrait analysis and intelligent recommendation method based on artificial intelligence

Abstract

The invention discloses a marketing customer portrait analysis and intelligent recommendation method based on artificial intelligence, and the method comprises the following steps: S1, collecting customer data, and constructing an interest feature vector set; s2, constructing a customer portrait through a neural dynamic capsule network model, and performing hierarchical decomposition on the interest feature vector set by adopting an interest capsule unit; s3, adopting an eagle strategy optimization algorithm to realize personalized recommendation optimization, and executing high-altitude circling search in the commodity recommendation pool to obtain a preliminary candidate commodity set; s4, executing dive attack optimization, performing sorting optimization on the preliminary candidate commodity set, and determining a recommendation sorting rule through an adaptive attack strategy; s5, dynamically adjusting interest capsule unit weight parameters of the neural dynamic capsule network model and recommendation sorting rule parameters in an eagle strategy optimization algorithm; and S6, generating a final customer personalized recommendation list. According to the method, the neural dynamic capsule network and the eagle strategy optimization algorithm are combined, so that customer portrait construction and intelligent recommendation optimization are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a marketing customer portrait analysis and intelligent recommendation method based on artificial intelligence. Background Art

[0002] In today's digital age, with the rapid development of e-commerce, social media, and various online services, intelligent recommendation systems have become an important technical means in the marketing field. Recommendation systems analyze customers' historical behaviors, interest preferences, and social relationships to achieve precise push of products or content, thereby enhancing the customer experience and improving the commercial conversion rate. Currently, common intelligent recommendation methods in the market mainly include collaborative filtering-based, content matching-based, and deep learning-based recommendation methods.

[0003] Traditional collaborative filtering recommendation methods achieve personalized recommendations by mining the similarity between customers or the interaction relationship between customers and products. This method can be divided into customer-based collaborative filtering and item-based collaborative filtering. The customer-based collaborative filtering method uses the historical behaviors of similar customers to predict the preferences of target customers, while the item-based collaborative filtering method recommends by calculating the similarity between products. However, such methods have problems of data sparsity and cold start, that is, when new customers or new products are added to the system, due to the lack of sufficient interaction data, it is difficult for the recommendation system to provide accurate personalized recommendations. In addition, collaborative filtering methods usually rely on customer rating data, but in actual application scenarios, customer rating behaviors are less, which is difficult to support high-quality recommendation effects.

[0004] Content matching-based recommendation methods calculate the matching degree by analyzing the characteristic information of customers and products. For example, in e-commerce recommendations, the system analyzes the categories, brands, keywords of products, as well as customers' search, click, and purchase records to generate a personalized recommendation list. Although the content matching method can alleviate the data sparsity problem of the collaborative filtering method to a certain extent, it depends on the accurate description of product or customer characteristics and is difficult to effectively capture the dynamic changes of customers' interests. In addition, this method cannot fully utilize the potential relevance between customers, resulting in lower recommendation diversity.

[0005] In recent years, with the progress of artificial intelligence technology, deep learning has been widely applied to personalized recommendation systems. Recommendation methods based on deep learning usually adopt neural networks to extract deep features of customer behavior and achieve accurate recommendations through complex modeling methods. Among them, convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) are used to process image and sequence data, while the self-attention mechanism (Transformer) is used to capture long-range dependencies. In addition, reinforcement learning methods have also been introduced into recommendation systems, enabling the system to continuously optimize recommendation strategies through customer feedback. However, existing deep learning recommendation systems still have multiple technical bottlenecks. First, traditional neural network models are difficult to construct hierarchical customer interest representations, resulting in poor performance of the system when dealing with multi-level customer preferences. Second, most recommendation systems rely on fixed recommendation strategies and are difficult to dynamically adjust recommendation results in real-time interactions. In addition, existing methods usually only focus on short-term interests and ignore the medium-term and long-term interest changes of customers, thus affecting the long-term accuracy of recommendations.

[0006] Therefore, how to provide an artificial intelligence-based marketing customer portrait analysis and intelligent recommendation method is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose an artificial intelligence-based marketing customer portrait analysis and intelligent recommendation method. The present invention combines a neural dynamic capsule network with an eagle strategy optimization algorithm to achieve customer portrait construction and intelligent recommendation optimization. By modeling the short-term, medium-term, and long-term interests of users through interest capsule units, the interest feature weights are dynamically adjusted to improve the accuracy of interest representation. At the same time, an altitude circling search and dive attack optimization strategy is adopted to optimize the recommendation ranking based on the global search combined with user behavior data, ensuring a high matching degree of recommended content. Combined with a real-time feedback mechanism, the recommendation model and sorting rules are adaptively adjusted to achieve dynamic optimization. This method has the advantages of high accuracy, strong real-time adaptive optimization ability, and excellent personalized recommendation effect.

[0008] An artificial intelligence-based marketing customer portrait analysis and intelligent recommendation method according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect customer data, perform deduplication, missing value filling, and outlier processing on the customer data, and construct an interest feature vector set;

[0010] S2. Construct a customer portrait through a neural dynamic capsule network model, and hierarchically decompose the interest feature vector set by using interest capsule units, including short-term interests, medium-term interests, or long-term interests, to obtain hierarchical customer interest features;

[0011] S3. Use the eagle strategy optimization algorithm to achieve personalized recommendation optimization, perform high-altitude hovering search in the product recommendation pool, calculate the matching degree between hierarchical customer interest features and product features, and obtain a preliminary set of candidate products;

[0012] S4, perform dive attack optimization, optimize the ranking of the preliminary candidate product set according to the customer's historical click and purchase behavior, and distinguish customers as stable customers or customers with fluctuating interests through adaptive attack strategies, and determine recommendation ranking rules dominated by short-term interests, medium-term interests or long-term interests;

[0013] S5, collecting customer click-through rate, purchase rate, bounce rate and page dwell time in real time, and dynamically adjusting the interest capsule unit weight parameters of the neural dynamic capsule network model and the recommendation sorting rule parameters in the eagle strategy optimization algorithm;

[0014] S6. Generate a final customer personalized recommendation list based on the dynamically adjusted interest capsule unit weight parameters and recommendation sorting rule parameters.

[0015] Optionally, the S2 specifically includes:

[0016] S21, define the interest feature vector set X = {x1, x2, ..., x n}, where x i represents the i-th interest feature, and n represents the interest feature dimension;

[0017] S22. Build customer portraits based on the neural dynamic capsule network model and define the interest capsule unit matrix as U = {u1,u2,...,u m}, where u j represents the jth interest capsule unit, m represents the total number of interest capsule units, and each interest capsule unit corresponds to an interest category;

[0018] S23, dynamically calculate the routing of interest features and interest capsule units, optimize the weight distribution between capsule units through an adaptive iterative mechanism, and use a dynamic nonlinear mapping function to aggregate capsule features:

[0019]

[0020] Among them, s j represents the input aggregate feature of the jth interest capsule unit, c ij represents the routing weight from the i-th interest feature to the j-th interest capsule unit, W ij represents the mapping weight matrix from the i-th interest feature to the j-th interest capsule unit, λ represents the dynamic adjustment parameter, R jk represents the hierarchical dependency matrix between the jth interest capsule unit and the kth interest capsule unit, u kDenote the k-th interest capsule unit, δ denote the interest propagation adjustment factor, and tanh denote the hyperbolic tangent function. Denote the correlation degree between the j-th interest capsule unit and the l-th interest capsule unit, u l Denote the l-th interest capsule unit;

[0021] S24. Calculate the output vector of the interest capsule unit by using normalization transformation and non-linear activation function:

[0022]

[0023] Among them, v j Denote the output vector of the j-th interest capsule unit, ||·|| denote the norm, ρ denote the adaptive weight adjustment factor, ReLU denote the non-linear activation function, and α denote the non-linear adjustment coefficient;

[0024] S25. Calculate the short-term interest, medium-term interest and long-term interest features, and define the dynamic piecewise decay function of the interest feature intensity changing with time as:

[0025]

[0026] Among them, I t Denote the interest intensity at the current moment, I t-1 Denote the interest intensity at the previous moment, Δt denote the time interval, A t Denote the new interest feature intensity at the current moment, β S Denote the decay coefficient of the short-term interest, β M Denote the decay coefficient of the medium-term interest, β L Denote the decay coefficient of the long-term interest, satisfying β S >β M >β L ; γ S Denote the update coefficient of the short-term interest, γ M Denote the update coefficient of the medium-term interest, γ L Denote the update coefficient of the long-term interest; ξ S Denote the time adjustment coefficient of the short-term interest, ξ M Denote the time adjustment coefficient of the medium-term interest, ξ L Denote the time adjustment coefficient of the long-term interest, T S Denote the time boundary between the short-term interest and the medium-term interest, T M Denote the time boundary between the medium-term interest and the long-term interest;

[0027] S26. Set the adaptive weight coefficient w S of the short-term interest, the adaptive weight coefficient w M of the medium-term interest, and the adaptive weight coefficient wL , satisfying \(w\) S + \(w\) M + \(w\) L = 1 and \(w\) S > \(w\) M > \(w\) L , weighted fusion of short - term interest, medium - term interest and long - term interest is performed to generate hierarchical customer interest characteristics.

[0028] Optionally, the S3 specifically includes:

[0029] S31. According to the hierarchical customer interest characteristics and product feature data obtained from the customer portrait, construct a multi - dimensional matching space, and use the high - altitude hovering search mechanism in the eagle strategy optimization algorithm to perform high - altitude hovering search in the product recommendation pool, and calculate the matching degree between the hierarchical customer interest characteristics and product features:

[0030]

[0031] where \(M\) pq represents the matching degree between the \(p\) - th hierarchical customer interest characteristic and the \(q\) - th product feature, \(N\) represents the total dimension number of the hierarchical customer interest characteristics, \(\alpha\) k , \(\beta\) k and \(\zeta\) represent feature weight adjustment parameters, exp represents the exponential function, \(I\) p represents the \(p\) - th hierarchical customer interest characteristic, \(P\) q represents the matching degree of the \(q\) - th product feature, \(\sigma\) represents the matching smoothing parameter, \(\Delta\) k represents the feature scale parameter, \(I\) pk represents the \(k\) - th component of the \(p\) - th hierarchical customer interest characteristic, \(P\) qk represents the \(k\) - th component of the \(q\) - th product feature;

[0032] S32. Based on the calculation result of the matching degree between the hierarchical customer interest characteristics and product features, set a pre - screening threshold \(T\) θ ;

[0033] S33. Classify the products whose matching degree between the hierarchical customer interest characteristics and product features is greater than the pre - screening threshold \(T\) θ into the preliminary candidate product set.

[0034] Optionally, the S4 specifically includes:

[0035] S41. Organize each candidate product in the preliminary candidate product set and extract customer historical click and purchase behavior data;

[0036] S42. Determine the customer status using statistical methods based on the volatility of the customer's historical click and purchase behavior data, and generate a customer status indicator variable σ. For stable customers, σ = 1; for customers with fluctuating interests, σ = 0.

[0037] S43. For each candidate product in the preliminary candidate product set, calculate the dive attack score using the dynamic behavior weight data:

[0038]

[0039] where A r represents the dive attack score of the r-th candidate product, T represents the preset time window, ω t represents the weight coefficient, represents the normalized behavior value of the r-th candidate product at time t, ρ represents the cumulative index, Δ rs represents the difference value of the r-th candidate product in the s-th dynamic feature dimension, and λ s represent the adjustment parameters, and N1 represents the number of dynamic feature dimensions;

[0040] S44. Combine the customer status indicator variable with the short-term interest, medium-term interest, and long-term interest indicators corresponding to the candidate products, and calculate the final ranking score for the candidate products:

[0041]

[0042] where F r represents the final ranking score of the r-th candidate product, I rS represents the short-term interest indicator of the r-th candidate product, I rM represents the medium-term interest indicator of the r-th candidate product, I rL represents the long-term interest indicator of the r-th candidate product, and η1 and η2 represent the non-linear correction parameters;

[0043] S45. Sort the candidate products in descending order according to the final ranking score, and conduct statistical analysis on the proportions of the short-term interest, medium-term interest, and long-term interest indicators of each candidate product in the preliminary candidate product set to determine the recommended ranking rules: when the short-term interest indicator in the preliminary candidate product set is the largest, adopt the recommended ranking rule dominated by short-term interest; when the medium-term interest indicator in the preliminary candidate product set is the largest, adopt the recommended ranking rule dominated by medium-term interest; when the long-term interest indicator in the preliminary candidate product set is the largest, adopt the recommended ranking rule dominated by long-term interest.

[0044] Optionally, the specific content of S5 includes:

[0045] S51. Collect customer click-through rate, purchase rate, bounce rate, and page dwell time data in real time, and construct a feedback data set through customer behavior logs;

[0046] S52. Perform data cleaning, outlier removal, missing value filling, and normalization on the feedback data set to form a standardized feedback data set;

[0047] S53. Construct a dynamic feedback weight index based on the standardized feedback data set:

[0048]

[0049] where Ψ represents the dynamic feedback weight index, CTR represents the click-through rate, CVR represents the purchase rate, BR represents the bounce rate, TST represents the page dwell time, τ1, τ2, τ3, and τ4 represent feedback index weight coefficients, ε1 and ε2 represent smoothing factors, ζ1 represents the feedback exponent, and ζ2 represents the decay adjustment parameter;

[0050] S54. Adjust the weight parameters of the interest capsule units in the neural dynamic capsule network model and the recommendation ranking rule parameters in the eagle strategy optimization algorithm according to the dynamic feedback weight Ψ to achieve closed-loop regulation:

[0051]

[0052] where Θ' represents the adjusted parameter, Θ represents the parameter before adjustment, tanh represents the hyperbolic tangent function, Ψ0 represents the preset feedback weight reference value, μ represents the ranking adjustment coefficient, and ν represents the non-linear adjustment exponent.

[0053] The beneficial effects of the present invention are:

[0054] First of all, the present invention uses a neural dynamic capsule network to construct a hierarchical customer portrait, realizing multi-level modeling of short-term interest, medium-term interest, and long-term interest. Through the dynamic routing mechanism of interest capsule units, the interest features at different time scales can adaptively adjust their weights, ensuring that important interest signals are effectively captured while reducing noise interference. This mechanism can overcome the limitations of traditional neural networks in interest representation, making the recommendation system more accurate in dealing with user interest changes. In addition, during the interest capsule modeling process of the present invention, a time decay mechanism is introduced, enabling interest features to be dynamically updated over time, ensuring that the recommendation system still maintains a high personalized recommendation accuracy during long-term operation.

[0055] Secondly, in terms of the recommended optimization strategy, the present invention adopts the Eagle Strategy Optimization Algorithm. Through a two-stage optimization strategy of high-altitude hovering search and precise dive attack, the screening accuracy of recommended products is improved. In the high-altitude hovering search stage, the system conducts a global search in the product recommendation pool, calculates the matching degree between the hierarchical customer interest characteristics and product characteristics, and screens out a preliminary candidate product set. In the dive attack optimization stage, the system further combines data such as the customer's historical click-through rate and purchase rate, dynamically adjusts the product sorting rules, and differentiates stable customers from customers with fluctuating interests through an adaptive attack strategy, ensuring that the recommended strategy can be adaptively adjusted according to changes in customer behavior. This optimization process greatly improves the flexibility of the recommendation system, enabling it to automatically adjust the recommended strategy under different customer behavior patterns, thereby increasing the user's click-through rate and conversion rate.

[0056] In addition, the present invention introduces a real-time feedback mechanism in the recommendation system. By continuously monitoring interaction data such as the customer's click-through rate, purchase rate, bounce rate, and page stay duration, the weights of the interest capsule units in the neural dynamic capsule network and the recommended sorting rules of the Eagle Strategy Optimization Algorithm are dynamically optimized. The closed-loop optimization mechanism based on the feedback data enables the system to automatically adjust the recommendation model according to the user's real-time interaction behavior, avoiding the limitations brought by fixed recommendation strategies and improving the adaptability and personalization degree of the recommendation results. Compared with traditional recommendation systems that can only perform offline optimization based on historical data, the recommended strategy of the present invention can continuously learn and optimize during the user interaction process to achieve more efficient personalized recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0058] Figure 1 is the overall flowchart of a marketing customer portrait analysis and intelligent recommendation method based on artificial intelligence proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0060] Refer to Figure 1 , a marketing customer portrait analysis and intelligent recommendation method based on artificial intelligence, including the following steps:

[0061] S1. Collect customer data, perform deduplication, missing value filling, and outlier processing on the customer data, and construct an interest feature vector set;

[0062] S2. Construct a customer profile through a neural dynamic capsule network model, and hierarchically decompose the interest feature vector set using interest capsule units, including short-term interest, medium-term interest, or long-term interest, to obtain hierarchical customer interest features;

[0063] S3. Implement personalized recommendation optimization using the eagle strategy optimization algorithm, perform high-altitude hovering search in the commodity recommendation pool, calculate the matching degree between the hierarchical customer interest features and commodity features, and obtain a preliminary candidate commodity set;

[0064] S4. Execute dive attack optimization, sort and optimize the preliminary candidate commodity set according to the customer's historical click and purchase behaviors, and distinguish customers as stable customers or interest-fluctuating customers through an adaptive attack strategy to determine the recommendation sorting rules dominated by short-term interest, medium-term interest, or long-term interest;

[0065] S5. Real-time collect the customer click-through rate, purchase rate, bounce rate, and page stay duration, and dynamically adjust the weight parameters of the interest capsule units of the neural dynamic capsule network model and the recommendation sorting rule parameters in the eagle strategy optimization algorithm;

[0066] S6. Generate a final customer personalized recommendation list according to the dynamically adjusted weight parameters of the interest capsule units and the recommendation sorting rule parameters.

[0067] The marketing customer profile analysis and intelligent recommendation method based on artificial intelligence provided by the present invention can effectively improve the accuracy and adaptability of the recommendation system. By constructing an interest feature vector set and hierarchically modeling customer interests using a neural dynamic capsule network, the system can more accurately capture the short-term interest, medium-term interest, and long-term interest of users. The eagle strategy optimization algorithm is used for personalized recommendation optimization, the preliminary candidate commodity set is screened through high-altitude hovering search, and the recommendation effect is improved by combining the dive attack optimization strategy to ensure that the recommendation results are more in line with the user's current interest state. At the same time, by real-time collecting feedback data such as the customer click-through rate, purchase rate, bounce rate, and page stay duration, the weight of the interest capsule units of the neural dynamic capsule network and the recommendation sorting rules are dynamically adjusted to achieve closed-loop optimization of personalized recommendations. The present invention can effectively adapt to the dynamic changes of user interests, improve the real-time performance and personalization degree of the recommendation system, and thus enhance the user experience and improve the commercial conversion rate.

[0068] In this embodiment, the S2 specifically includes:

[0069] S21. Define an interest feature vector set X = {x1, x2,..., x n}, where x i represents the i-th interest feature, and n represents the interest feature dimension;

[0070] S22. Construct a customer portrait based on the neural dynamic capsule network model, and define the interest capsule unit matrix as U = {u1, u2,..., u m}, where u j represents the j-th interest capsule unit, m represents the total number of interest capsule units, and each interest capsule unit corresponds to an interest category;

[0071] S23. Perform dynamic routing calculations on the interest features and interest capsule units, optimize the weight allocation between capsule units through an adaptive iteration mechanism, and use a dynamic non-linear mapping function for capsule feature aggregation:

[0072]

[0073] Among them, s j represents the input aggregation feature of the j-th interest capsule unit, c ij represents the routing weight from the i-th interest feature to the j-th interest capsule unit, W ij represents the mapping weight matrix from the i-th interest feature to the j-th interest capsule unit, λ represents the dynamic adjustment parameter, R jk represents the hierarchical dependence matrix between the j-th interest capsule unit and the k-th interest capsule unit, u k represents the k-th interest capsule unit, δ represents the interest propagation adjustment factor, tanh represents the hyperbolic tangent function, represents the mutual correlation degree between the j-th interest capsule unit and the l-th interest capsule unit, u l represents the l-th interest capsule unit;

[0074] S24. Calculate the output vector of the interest capsule unit using normalization transformation and non-linear activation function:

[0075]

[0076] Among them, v j represents the output vector of the j-th interest capsule unit, ||·|| represents the norm, ρ represents the adaptive weight adjustment factor, ReLU represents the non-linear activation function, and α represents the non-linear adjustment coefficient;

[0077] S25. Calculate the short-term interest, medium-term interest, and long-term interest features, and define the dynamic piecewise decay function of the interest feature intensity changing with time as:

[0078]

[0079] Among them, I t represents the interest intensity at the current moment, I t-1 represents the interest intensity at the previous moment, Δt represents the time interval, A tRepresents the intensity of the new interest feature at the current moment, β S Represents the decay coefficient of short-term interest, β M Represents the decay coefficient of medium-term interest, β L Represents the decay coefficient of long-term interest, satisfying β S >β M >β L ; γ S Represents the update coefficient of short-term interest, γ M Represents the update coefficient of medium-term interest, γ L Represents the update coefficient of long-term interest; ξ S Represents the time adjustment coefficient of short-term interest, ξ M Represents the time adjustment coefficient of medium-term interest, ξ L Represents the time adjustment coefficient of long-term interest, T S Represents the time boundary between short-term interest and medium-term interest, T M Represents the time boundary between medium-term interest and long-term interest;

[0080] S26. Set the adaptive weight coefficient w of short-term interest S , the adaptive weight coefficient w of medium-term interest M and the adaptive weight coefficient w of long-term interest L , satisfying w S +w M +w L =1 and w S >w M >w L , and perform weighted fusion on short-term interest, medium-term interest and long-term interest to generate hierarchical customer interest features.

[0081] The present invention uses a neural dynamic capsule network to construct a customer portrait, and through the hierarchical decomposition of interest capsule units, realizes the fine modeling of short-term, medium-term and long-term interests, enabling the system to capture user interest features at different time scales. The dynamic routing mechanism ensures the adaptive adjustment of the weights of interest features between different levels, reduces the influence of invalid features, and improves the accuracy of interest representation. The output of the interest capsule unit is optimized using normalization transformation and non-linear activation functions to enhance the learning ability of the model. At the same time, a dynamic piecewise decay function is introduced to enable the interest features to be dynamically updated over time, ensuring that the recommendation system can adapt to changes in user interests in the long term. In addition, through the adaptive weighted fusion of short-term, medium-term and long-term interests, the hierarchical expression ability of the customer portrait is further improved, providing more accurate input data for subsequent personalized recommendation optimization.

[0082] In this embodiment, the S3 specifically includes:

[0083] S31. Construct a multi-dimensional matching space based on the hierarchical customer interest characteristics and product feature data obtained from the customer portrait, and use the high-altitude hovering search mechanism in the Eagle Strategy Optimization Algorithm to perform high-altitude hovering search in the product recommendation pool, and calculate the matching degree between the hierarchical customer interest characteristics and the product features:

[0084]

[0085] Among them, M pq represents the matching degree between the p-th hierarchical customer interest characteristic and the q-th product feature, N represents the total dimension number of the hierarchical customer interest characteristics, α k , β k and ζ represent feature weight adjustment parameters, exp represents the exponential function, I p represents the p-th hierarchical customer interest characteristic, P q represents the matching degree of the q-th product feature, σ represents the matching smoothing parameter, Δ k represents the feature scale parameter, I pk represents the k-th component of the p-th hierarchical customer interest characteristic, P qk represents the k-th component of the q-th product feature;

[0086] S32. Set a pre-screening threshold T θ ;

[0087] S33. Classify the products with the matching degree between the hierarchical customer interest characteristics and the product features greater than the pre-screening threshold T θ into the preliminary candidate product set.

[0088] The present invention adopts the high-altitude hovering search mechanism in the Eagle Strategy Optimization Algorithm to perform global search in the product recommendation pool, calculates the matching degree between the hierarchical customer interest characteristics and the product features, and realizes accurate product screening. By constructing a multi-dimensional matching space and using a complex matching degree calculation formula, it is ensured that the selection of recommended products not only depends on a single interest dimension, but combines multiple levels of interest characteristics, improving the comprehensiveness and accuracy of matching. Set a pre-screening threshold, and screen out the preliminary candidate product set according to the calculated interest matching degree, ensuring that the products entering the next stage of optimization have a high interest relevance. At the same time, this method can dynamically adjust the screening threshold, making the system flexible in different recommendation scenarios, improving the diversity and personalization of recommended products, and thus enhancing the user's acceptance of the recommended content.

[0089] In this embodiment, the specific steps of S4 include:

[0090] S41. Organize each candidate product in the preliminary candidate product set, and extract the customer's historical click and purchase behavior data;

[0091] S42. Determine the customer status using statistical methods based on the volatility of the customer's historical click and purchase behavior data, and generate a customer status indicator variable σ. For stable customers, σ = 1; for customers with fluctuating interests, σ = 0.

[0092] S43. For each candidate product in the preliminary candidate product set, calculate the dive attack score using the dynamic behavior weight data:

[0093]

[0094] where A r represents the dive attack score of the r-th candidate product, T represents the preset time window, ω t represents the weight coefficient, represents the normalized behavior value of the r-th candidate product at time t, ρ represents the cumulative index, Δ rs represents the difference value of the r-th candidate product in the s-th dynamic feature dimension, and λ s represent adjustment parameters, and N1 represents the number of dynamic feature dimensions;

[0095] S44. Combine the customer status indicator variable with the short-term interest, medium-term interest, and long-term interest indicators corresponding to the candidate products, and calculate the final ranking score for the candidate products:

[0096]

[0097] where F r represents the final ranking score of the r-th candidate product, I rS represents the short-term interest indicator of the r-th candidate product, I rM represents the medium-term interest indicator of the r-th candidate product, I rL represents the long-term interest indicator of the r-th candidate product, and η1 and η2 represent non-linear correction parameters;

[0098] S45. Sort the candidate products in descending order according to the final ranking score, and conduct statistical analysis on the proportions of the short-term interest, medium-term interest, and long-term interest indicators of each candidate product in the preliminary candidate product set to determine the recommended ranking rules: when the short-term interest indicator in the preliminary candidate product set is the largest, adopt a recommended ranking rule dominated by short-term interest; when the medium-term interest indicator in the preliminary candidate product set is the largest, adopt a recommended ranking rule dominated by medium-term interest; when the long-term interest indicator in the preliminary candidate product set is the largest, adopt a recommended ranking rule dominated by long-term interest.

[0099] The dive attack optimization mechanism proposed by the present invention can, on the basis of the preliminary candidate product set, combine the customer's historical click and purchase behaviors to dynamically optimize the sorting of products. By calculating the dive attack score, products with high relevance are further screened out, and combined with customer status analysis, stable customers and customers with fluctuating interests are distinguished, making the recommendation strategy more personalized. The final sorting score is optimized using short-term, medium-term, and long-term interest indicators to ensure that the system can adjust the recommendation strategy according to the user's interest status, so that different types of users can obtain the most suitable recommendation results. In addition, this method automatically determines the sorting rules dominated by short-term interest, medium-term interest, or long-term interest by statistically analyzing the interest ratio of the preliminary candidate product set, improving the adaptability and accuracy of the recommendation strategy, thereby effectively enhancing the stability of the recommendation system and user satisfaction.

[0100] In this embodiment, step S5 specifically includes:

[0101] S51. Real-time collect data on customer click-through rate, purchase rate, bounce rate, and page stay duration, and construct a feedback data set through the customer behavior log;

[0102] S52. Perform data cleaning, outlier removal, missing value filling, and normalization processing on the feedback data set to form a standardized feedback data set;

[0103] S53. Construct a dynamic feedback weight index based on the standardized feedback data set:

[0104]

[0105] Among them, Ψ represents the dynamic feedback weight index, CTR represents the click-through rate, CVR represents the purchase rate, BR represents the bounce rate, TST represents the page stay duration, τ1, τ2, τ3, and τ4 represent the feedback index weight coefficients, ε1 and ε2 represent the smoothing factors, ζ1 represents the feedback exponent, and ζ2 represents the decay adjustment parameter;

[0106] S54. Adjust the weight parameters of the interest capsule units of the neural dynamic capsule network model and the recommendation sorting rule parameters in the eagle strategy optimization algorithm according to the dynamic feedback weight Ψ to achieve closed-loop regulation:

[0107]

[0108] Among them, Θ' represents the adjusted parameter, Θ represents the parameter before adjustment, tanh represents the hyperbolic tangent function, Ψ0 represents the preset feedback weight reference value, μ represents the sorting adjustment coefficient, and ν represents the non-linear adjustment exponent.

[0109] The present invention establishes a dynamic feedback mechanism by collecting feedback data such as customer click-through rate, purchase rate, bounce rate, and page dwell time in real time, ensuring that the recommendation system can adjust the recommendation strategy at any time to improve the accuracy and real-time performance of recommendations. The dynamic feedback weight index is used to calculate the influence degree of user behavior, and based on this weight, the weights of the interest capsule units of the neural dynamic capsule network and the recommendation sorting rules of the eagle strategy optimization algorithm are dynamically adjusted to achieve continuous optimization of personalized recommendations. Through the adaptive weight adjustment formula, the system can optimize the recommendation parameters in real time according to the actual interaction situation of users, avoiding the problem of insufficient adaptability caused by fixed parameters in traditional recommendation systems. In addition, the closed-loop optimization mechanism of the present invention can ensure that the recommendation system remains efficient and stable during long-term operation, improve user satisfaction with the recommended content, and enhance the commercial value of the platform.

[0110] Embodiment 1:

[0111] To verify the feasibility of the present invention in implementation, the present invention is deployed to the intelligent recommendation system of a large e-commerce platform to optimize the personalized recommendation effect, improve the user click-through rate and purchase conversion rate, reduce the bounce rate, and increase the user dwell time. This e-commerce platform has a large user group, generating hundreds of millions of user behavior data every day, with a wide variety of product categories. The recommendation system needs to accurately match user needs in a complex and changing market environment. However, the existing recommendation systems rely on traditional collaborative filtering or content-based matching methods, which are difficult to distinguish short-term interests, medium-term interests, and long-term interests, resulting in a lack of dynamic adjustment ability in the recommendation results, poor user experience, and limited commercial conversion rate.

[0112] In this embodiment, the intelligent recommendation method of the present invention is integrated into the recommendation engine to achieve more accurate and personalized recommendations. First, the system collects users' behavioral data such as browsing, searching, clicking, purchasing, and favoriting, and constructs an interest feature vector set. Through the Neural Dynamic Capsule Network (NDCN), hierarchical modeling of users' interests is performed to extract short-term interests (based on recent behaviors), medium-term interests (based on continuous interaction records), and long-term interests (based on long-term historical preferences) respectively, generating hierarchical interest features. Subsequently, the system adopts the Hawk Strategy Optimization algorithm (HSO) to perform high-altitude hovering search in the product recommendation pool, calculate the matching degree between interest features and product features, and screen out a preliminary candidate product set. Then, the system performs dive attack optimization on the candidate product set, sorts and optimizes the products according to data such as users' historical click-through rates, purchase rates, and bounce rates, and differentiates stable users from users with fluctuating interests through an adaptive attack strategy to determine the recommendation sorting rules, ensuring that the recommended content can conform to users' dynamic interest preferences. In addition, the system continuously monitors users' interaction data during the recommendation process, including click-through rates, purchase conversion rates, bounce rates, and user stay durations, and adjusts the weights of interest capsule units in the Neural Dynamic Capsule Network model and the recommendation sorting rules of the Hawk Strategy Optimization algorithm in real time to form a closed-loop optimized recommendation system.

[0113] To evaluate the recommendation effect of the present invention, a comparative test was conducted on the performance of the recommendation system of the present invention and the traditional recommendation system at multiple stages, and core indicators such as click-through rate, purchase conversion rate, bounce rate, and average user stay time were statistically analyzed. The results show that in the case of using the traditional recommendation system, the click-through rate is relatively low, the user purchase conversion rate is limited, the bounce rate is relatively high, and the average user stay time is relatively short. However, after using the recommendation system of the present invention, the click-through rate has increased significantly, the purchase conversion rate has increased significantly, the bounce rate has decreased, and the average user stay time has been extended. Especially during promotional activities, the recommendation system of the present invention can more accurately capture changes in users' needs and timely adjust the recommendation strategy, comprehensively improving the user experience and business conversion effect.

[0114] Table 1 Comparison of the effects of the traditional recommendation system and the recommendation system of the present invention on the e-commerce platform

[0115]

[0116]

[0117] As can be seen from the data in Table 1 above, compared with traditional recommendation methods, the method of the present invention has significant improvements in multiple key indicators. The increase in click-through rate indicates that the recommended content can attract users' attention more effectively. The growth in purchase conversion rate shows that the recommended products are more in line with users' interests. The decrease in bounce rate and the increase in user stay time indicate that the present invention effectively enhances user engagement and improves users' satisfaction with the recommendation results. In addition, when faced with large-scale user data and diverse product categories, the method of the present invention still maintains high computational efficiency and real-time performance, and can quickly adjust the recommendation strategy when user behavior changes, thereby improving the accuracy of personalized recommendations and the dynamic adaptation ability of the system.

[0118] In summary, the recommendation method of the present invention combines a neural dynamic capsule network with an eagle strategy optimization algorithm, which not only optimizes user interest modeling, enabling the recommendation system to accurately distinguish short-term interests, medium-term interests, and long-term interests, but also enhances the matching accuracy of the recommended content through the eagle strategy optimization algorithm, improving the adaptability of the recommendation system. Combined with a real-time feedback mechanism, the present invention realizes the closed-loop optimization of the recommendation strategy, ensuring that the system can dynamically adjust the recommended content according to users' real-time interactions, thereby enhancing the user experience, optimizing the commercial conversion effect of the platform, and providing an efficient and accurate intelligent recommendation solution for scenarios such as e-commerce, intelligent marketing, and personalized content recommendation.

[0119] The above is only a preferred specific embodiment 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, should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing marketing customer portraits and intelligent recommendation based on artificial intelligence, characterized in that, It includes the following steps: S1. Collect customer data, perform deduplication, missing value filling, and outlier processing on the customer data, and construct an interest feature vector set; S2. Construct a customer portrait through a neural dynamic capsule network model, and hierarchically decompose the interest feature vector set using interest capsule units, including short-term interest, medium-term interest, or long-term interest, to obtain hierarchical customer interest features; S3. Implement personalized recommendation optimization using the eagle strategy optimization algorithm, perform high-altitude hovering search in the commodity recommendation pool, calculate the matching degree between the hierarchical customer interest features and commodity features, and obtain a preliminary candidate commodity set; S4. Perform dive attack optimization, sort and optimize the preliminary candidate commodity set according to the customer's historical click and purchase behaviors, and distinguish customers as stable customers or interest-fluctuating customers through an adaptive attack strategy, and determine the recommendation sorting rules dominated by short-term interest, medium-term interest, or long-term interest; S5. Collect the customer click-through rate, purchase rate, bounce rate, and page stay duration in real time, and dynamically adjust the weight parameters of the interest capsule units in the neural dynamic capsule network model and the recommendation sorting rule parameters in the eagle strategy optimization algorithm; S6. Generate a final customer personalized recommendation list according to the dynamically adjusted weight parameters of the interest capsule units and the recommendation sorting rule parameters.

2. The method for analyzing marketing customer portraits and intelligent recommendation based on artificial intelligence according to claim 1, wherein The specific content of S2 includes: S21. Define the set of interest feature vectors \(X = \{x_1, x_2, \ldots, x\) n \}, where \(x\) i represents the \(i\)-th interest feature, and \(n\) represents the dimension of interest features; S22. Build a customer portrait based on the neural dynamic capsule network model, and define the interest capsule unit matrix as U = {u1, u2,..., u m}, where u j represents the j-th interest capsule unit, m represents the total number of interest capsule units, and each interest capsule unit corresponds to an interest category; S23. Perform dynamic routing calculation on the interest features and interest capsule units, optimize the weight allocation between capsule units through an adaptive iteration mechanism, and perform capsule feature aggregation using a dynamic non-linear mapping function: Among them, s j represents the input aggregation feature of the j-th interest capsule unit, c ij represents the routing weight from the i-th interest feature to the j-th interest capsule unit, W ij represents the mapping weight matrix from the i-th interest feature to the j-th interest capsule unit, λ represents the dynamic adjustment parameter, R jk represents the hierarchical dependence matrix between the j-th interest capsule unit and the k-th interest capsule unit, u k represents the k-th interest capsule unit, δ represents the interest propagation adjustment factor, tanh represents the hyperbolic tangent function, represents the correlation degree between the j-th interest capsule unit and the l-th interest capsule unit, u l represents the l-th interest capsule unit; S24. Calculate the output vector of the interest capsule unit using normalization transformation and non-linear activation function: Among them, v j represents the output vector of the j-th interest capsule unit, ||·|| represents the norm, ρ represents the adaptive weight adjustment factor, ReLU represents the non-linear activation function, and α represents the non-linear adjustment coefficient; S25. Calculate short-term interest, medium-term interest, and long-term interest features, and define the dynamic piecewise decay function of the interest feature intensity changing with time as: Among them, I t represents the interest intensity at the current moment, I t-1 represents the interest intensity at the previous moment, Δt represents the time interval, A t represents the new interest feature intensity at the current moment, β S represents the decay coefficient of short-term interest, β M represents the decay coefficient of medium-term interest, β L represents the decay coefficient of long-term interest, satisfying β S >β M >β L ; γ S represents the update coefficient of short-term interest, γ M represents the update coefficient of medium-term interest, γ L represents the update coefficient of long-term interest; ξ S represents the time adjustment coefficient of short-term interest, ξ M represents the time adjustment coefficient of medium-term interest, ξ L represents the time adjustment coefficient of long-term interest, T S represents the time boundary between short-term interest and medium-term interest, T M represents the time boundary between medium-term interest and long-term interest; S26. Set the adaptive weight coefficient w of short-term interest S , the adaptive weight coefficient w of medium-term interest M and the adaptive weight coefficient w of long-term interest L , satisfying w S +w M +w L =1 and w S >w M >w L , and perform weighted fusion on short-term interest, medium-term interest and long-term interest to generate hierarchical customer interest characteristics.

3. A method for analyzing marketing customer portraits and intelligent recommendation based on artificial intelligence according to claim 1, characterized in that, The specific content of S3 includes: S31. According to the hierarchical customer interest features and commodity feature data obtained from the customer portrait, construct a multi-dimensional matching space, and perform high-altitude hovering search in the commodity recommendation pool using the high-altitude hovering search mechanism in the eagle strategy optimization algorithm, and calculate the matching degree between the hierarchical customer interest features and commodity features; Among them, M pq represents the matching degree between the p-th hierarchical customer interest feature and the q-th commodity feature, N represents the total dimension number of the hierarchical customer interest features, α k , β k and ζ represent feature weight adjustment parameters, exp represents the exponential function, I p represents the p-th hierarchical customer interest feature, P q represents the matching degree of the q-th commodity feature, σ represents the matching smoothing parameter, Δ k represents the feature scale parameter, I pk represents the k-th component of the p-th hierarchical customer interest feature, P qk represents the k-th component of the q-th commodity feature; S32. Set a pre-screening threshold T according to the calculation result of the matching degree between the hierarchical customer interest characteristics and the product characteristics θ ; S33. Classify the products whose matching degree between the hierarchical customer interest features and the product features is greater than the pre-screening threshold T θ into the preliminary candidate product set.

4. A method for analyzing marketing customer portraits and intelligent recommendation based on artificial intelligence according to claim 1, characterized in that, The specific content of S4 includes: S41. Sort each candidate commodity in the preliminary candidate commodity set, and extract the customer's historical click and purchase behavior data; S42. According to the volatility of the customer's historical click and purchase behavior data, use statistical methods to determine the customer status, generate a customer status indicator variable σ, where σ = 1 for stable customers and σ = 0 for interest-fluctuating customers; S43. For each candidate commodity in the preliminary candidate commodity set, calculate the dive attack score using dynamic behavior weight data; Among them, A r represents the dive attack score of the r-th candidate product, T represents the preset time window, and ω t represents the weight coefficient, represents the normalized behavior value of the r-th candidate product at time t, ρ represents the cumulative index, and Δ rs represents the difference value of the r-th candidate product in the s-th dynamic feature dimension, and λ s represent the adjustment parameters, and N1 represents the number of dynamic feature dimensions; S44. Combine the customer status indicator variable with the short-term interest, medium-term interest, and long-term interest indicators corresponding to the candidate commodity, and calculate the final sorting score for the candidate commodity; Among them, F r represents the final ranking score of the r-th candidate product, I rS represents the short-term interest index of the r-th candidate product, I rM represents the medium-term interest index of the r-th candidate product, I rL represents the long-term interest index of the r-th candidate product, and η1 and η2 represent non-linear correction parameters; S45. Sort the candidate products in descending order according to the final sorting scores, and conduct statistical analysis on the proportions of short-term interest, medium-term interest, and long-term interest indicators of each candidate product in the preliminary candidate product set to determine the recommendation sorting rules: when the short-term interest indicator in the preliminary candidate product set is the largest, adopt a recommendation sorting rule dominated by short-term interest; when the medium-term interest indicator in the preliminary candidate product set is the largest, adopt a recommendation sorting rule dominated by medium-term interest; when the long-term interest indicator in the preliminary candidate product set is the largest, adopt a recommendation sorting rule dominated by long-term interest.

5. A method for analyzing marketing customer portraits and intelligent recommendation based on artificial intelligence according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. Real-time collect data on customer click-through rate, purchase rate, bounce rate, and page stay duration, and construct a feedback data set through customer behavior logs. S52. Perform data cleaning, outlier removal, missing value filling, and normalization processing on the feedback data set to form a standardized feedback data set. S53. Construct a dynamic feedback weight index based on the standardized feedback data set: Among them, Ψ represents the dynamic feedback weight index, CTR represents the click-through rate, CVR represents the purchase rate, BR represents the bounce rate, TST represents the page stay duration, τ1, τ2, τ3, and τ4 represent feedback index weight coefficients, ε1 and ε2 represent smoothing factors, ζ1 represents the feedback exponent, and ζ2 represents the attenuation adjustment parameter. S54. Adjust the weight parameters of the interest capsule unit in the neural dynamic capsule network model and the recommendation sorting rule parameters in the eagle strategy optimization algorithm according to the dynamic feedback weight Ψ to achieve closed-loop regulation: Among them, Θ' represents the adjusted parameter, Θ represents the parameter before adjustment, tanh represents the hyperbolic tangent function, Ψ0 represents the preset feedback weight reference value, μ represents the sorting adjustment coefficient, and ν represents the non-linear adjustment exponent.

Images