Product recommendation method for electric power marketing management information platform

Abstract

The invention relates to the technical field of electric power marketing management, in particular to a product recommendation method for an electric power marketing management information platform, which comprises the following steps: S1, collecting user data of the electric power marketing management information platform; s2, performing standardization processing on the user data collected in the S1; s3, constructing a dynamic user portrait; s4, matching is carried out, and a candidate recommended product set is generated; s5, performing two-stage sorting on the candidate recommendation product set to generate a personalized recommendation list; and S6, pushing the personalized recommendation list generated in the step S5 to a user terminal, and dynamically adjusting recommendation logic parameters in two-stage sorting. According to the method, the dynamic portrait is constructed by fusing the multi-dimensional data of the user, and the recommendation logic is dynamically optimized in combination with the real-time operation strategy, so that the unification of the accuracy and the self-adaptability of personalized recommendation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power marketing management, and in particular to a product recommendation method for a power marketing management information platform. Background Art

[0002] In the process of informatization in the power industry, the power marketing management information platform has gradually become a key support carrier for power enterprises to carry out marketing, serve users, and make operation decisions; by analyzing users' purchase records, power consumption characteristics, and various feedback information, power enterprises can more accurately grasp users' needs and consumption trends, and then launch differentiated power products and value-added services; with the continuous growth of data scale and the continuous improvement of the diversification degree of users' needs, the traditional recommendation methods relying on static information and single-dimensional user information are difficult to meet the high standards that require both real-time performance and accuracy.

[0003] However, the existing technologies lack comprehensive consideration of dynamic monitoring of users' behaviors and real-time operation strategies when constructing user portraits and performing product recommendations, making it difficult to capture changes in users' needs in a timely manner and adaptively adjust the recommendation algorithms; in addition, some solutions lack multi-dimensional feature fusion in the process of matching the power product database with user portraits, resulting in deviations between the recommended results and users' actual needs; this limitation makes it difficult for the power marketing management information platform to fully exploit the potential of big data analysis and unable to ensure the simultaneous improvement of marketing efficiency and user satisfaction. Summary of the Invention

[0004] Based on the above purpose, the present invention provides a product recommendation method for a power marketing management information platform.

[0005] A product recommendation method for a power marketing management information platform includes the following steps:

[0006] S1: Collect user data of the power marketing management information platform, including users' historical purchase records, power consumption data, and user feedback data;

[0007] S2: Perform standardization processing on the user data collected in S1 to generate standardized user data;

[0008] S3: Construct a dynamic user portrait based on the standardized user data, where the dynamic user portrait includes a static attribute layer and a dynamic behavior layer;

[0009] S4: Match the dynamic user portrait with the power product database to generate a set of candidate recommended products;

[0010] S5: Combine the real-time operation strategy data of the power marketing platform to perform two-level sorting on the set of candidate recommended products to generate a personalized recommendation list;

[0011] S6: Push the personalized recommendation list generated in S5 to the user terminal, and dynamically adjust the recommendation logic parameters in the two-level sorting based on the click-through rate and purchase conversion rate of the recommended products by the user.

[0012] Optionally, the S1 specifically includes:

[0013] S11: Extract the user's historical purchase records from the marketing business system of the power marketing management information platform. The user's historical purchase records include user identification, types of purchased power products, purchase time, and corresponding power consumption equipment codes.

[0014] S12: Obtain the user's power consumption data from the power metering automation system. The power consumption data includes the time-series power consumption data divided by the billing cycle, the peak-to-valley power consumption ratio, and is associated and matched with the user identification.

[0015] S13: Collect the user's feedback data from the power customer service system. The user's feedback data includes the text feedback content actively submitted by the user, the text of the customer service call recording, and the corresponding user satisfaction rating. The text feedback content is associated and stored with the user identification.

[0016] Optionally, the S2 specifically includes:

[0017] S21: Perform number mapping conversion on the types of power products and power consumption equipment codes involved in the user's historical purchase records collected in S1, and uniformly convert them into fixed-length numerical feature vectors.

[0018] S22: Perform periodic normalization processing on the time-series power consumption data in the user's power consumption data. Use the maximum-minimum normalization method to normalize the power consumption in each period to the [0, 1] interval.

[0019] S23: Use the Z-score standardization method for the peak-to-valley power consumption ratio to calculate its standard deviation offset value in the overall distribution.

[0020] S24: Convert the non-standard power terms in the user's feedback data. By extracting the power-related words in the user's feedback text and matching the words with the preset power standard term library, replace them with the corresponding standard terms.

[0021] S25: Concatenate and integrate all the feature vectors processed by S21 to S24 according to the user identification to generate standardized user data with a unified dimension and consistent format.

[0022] Optionally, the S3 specifically includes:

[0023] S31: Select the information reflecting the basic characteristics of users from the standardized user data to construct a static attribute layer, where the static attribute layer includes the geographical area to which the user belongs, the user industry classification, and the types of common electrical equipment used by the user;

[0024] S32: Select the information that can reflect the user's consumption preferences and electricity consumption behavior patterns from the standardized user data to construct a dynamic behavior layer, where the dynamic behavior layer includes the purchase frequency within a continuous time period, the electricity peak-valley fluctuation pattern, and the feedback evaluation trend;

[0025] S33: Perform an association mapping on the static attribute layer and the dynamic behavior layer, and establish a cross-layer connection relationship through the user identifier, so that the data from different feature dimensions form a multi-dimensional attribute representation with a unified structure in the same user portrait, thereby generating a dynamic user portrait.

[0026] Optionally, the S33 specifically includes:

[0027] S331: Correspondingly match the data of the static attribute layer and the dynamic behavior layer according to the user identifier to construct an initial mapping pair set, specifically represented as a triple <U, A, B>, where U represents the user identifier; A represents the static attribute vector; B represents the dynamic behavior vector;

[0028] S332: Perform a splicing operation on the vectors A and B to construct a fused feature vector F, and its expression is: F = α1·a1, α2·a2,..., α i ·a i , β1·b1, β2·b2,..., β j ·b j , where: α i represents the weight factor of the static attribute a i ; β j represents the weight factor of the dynamic behavior b j ;

[0029] S333: Bind the fused feature vector F to the user identifier U to form the final dynamic user portrait D, and the expression is: D = {U, F}.

[0030] Optionally, the S4 specifically includes:

[0031] S41: Extract the fused feature vector of the user based on the dynamic user portrait, and extract the attribute feature vectors of various power products from the power product database;

[0032] S42: Use the feature similarity measurement method to calculate the matching similarity between the user fused feature vector and the power product attribute feature vector to obtain the product similarity score for each user;

[0033] S43: Perform threshold screening on the product similarity scores, and preliminarily screen the power products that meet the similarity threshold conditions to form an initial product set;

[0034] S44: Remove duplicates from the power products in the initial product set to form a candidate recommended product set for each user.

[0035] Optionally, the specific steps of S42 include:

[0036] S421: Normalize the user's fused feature vector and the attribute feature vector of the power product respectively to generate a vector form with a unified numerical range;

[0037] S422: Use the cosine similarity method to calculate the matching similarity between the normalized user fused feature vector and the power product attribute feature vector. The specific calculation formula is: where: Sim is the similarity score between the user fused feature vector and the power product attribute feature vector; u k is the k-th eigenvalue in the user fused feature vector; v k is the k-th eigenvalue in the power product attribute feature vector; p is the dimension of the feature vector.

[0038] Optionally, the specific steps of S5 include:

[0039] S51: Based on the real-time operation strategy data, obtain the promotion weights of the power products, and perform primary sorting on the products in the candidate recommended product set according to the promotion weights to generate a primary sorted list; where the calculation formula for the promotion weight is: W op (j) = γ × Q(j) + (1 - γ) × R(j), where, W op (j) represents the primary sorting promotion weight of candidate product j; Q(j) is the current promotion coefficient of candidate product j defined in the real-time operation strategy; R(j) is the historical sales popularity coefficient of candidate product j in the operation strategy; γ is the trade-off coefficient between the promotion coefficient and the historical sales popularity coefficient;

[0040] S52: Calculate the user preference matching degree according to the user fused feature vector and the product feature vector in the primary sorted list, and generate a personalized matching score P(u, j);

[0041] S53: Perform secondary re-sorting on the primary sorted list according to the calculated personalized matching score P(u, j) to generate a personalized recommendation list. The expression for the secondary re-sorting is:

[0042] L(u) = Sort{P(u, j)}, where j ∈ J, L(u) is the final personalized recommendation list for user u; J is the set of candidate recommended products after primary sorting; Sort{·} is a function that sorts in descending order according to the matching score.

[0043] Optionally, S6 specifically includes:

[0044] S61: Push the personalized recommendation list generated in S5 to the user terminal interface in real time in the form of interactive messages through the data communication interface of the power marketing management information platform, and record the browsing behavior of the user terminal for the recommended products;

[0045] S62: Calculate the click-through rate and purchase conversion rate of each recommended product by the user according to the browsing records of the recommended products by the user terminal;

[0046] S63: Dynamically adjust the trade-off coefficient γ between the promotion coefficient and the historical sales popularity coefficient in 55 by using the gradient optimization method according to the changes in the click-through rate CTR(j) and purchase conversion rate CR(j) of the product.

[0047] Optionally, S63 specifically includes:

[0048] S631: Based on the click-through rate and purchase conversion rate data within two consecutive statistical periods, calculate the change in click-through rate ΔCTR(j) and the change in purchase conversion rate ΔCR(j) of the recommended product respectively;

[0049] S632: Construct a loss function L(γ) for parameter adjustment with the change in click-through rate and the change in purchase conversion rate;

[0050] S633: Perform gradient calculation on the trade-off coefficient γ based on the loss function L(γ) to obtain the gradient value of the loss function with respect to the trade-off coefficient γ. The gradient calculation formula is:

[0051] where, is the gradient value of the loss function L(γ) with respect to the trade-off coefficient γ; represents the partial derivative of the click-through rate with respect to the trade-off coefficient γ; represents the partial derivative of the purchase conversion rate with respect to the trade-off coefficient γ;

[0052] S634: Update the trade-off coefficient according to the gradient descent optimization method. The calculation formula is: where γ new is the updated trade-off coefficient; γ old is the trade-off coefficient of the previous cycle; η is the learning rate.

[0053] Advantages of the present invention:

[0054] In the present invention, through multi-source data fusion and dynamic user portrait construction, the efficient integration of user static attributes and dynamic behaviors is achieved, thereby making the product recommendation process more accurate and comprehensive; a systematic closed-loop of data processing, feature extraction, and matching calculation is realized between each step through strictly defined calculation formulas and weighing coefficients, providing stable and discriminative user feature inputs for the power marketing management information platform, and significantly improving the relevance and personalized matching degree of the candidate recommended product set.

[0055] In the present invention, by adopting real-time operation strategy data and gradient optimization methods, the recommendation logic parameters are dynamically and adaptively adjusted, achieving an instant response to changes in user click-through rate and purchase conversion rate; this optimization mechanism effectively balances the relationship between the platform promotion needs and user personalized preferences, enabling the recommendation list to continuously maintain an efficient and accurate recommendation effect, providing strong data support and decision-making basis for power enterprises in the actual marketing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 Schematic diagram of the product recommendation method according to an embodiment of the present invention;

[0058] Figure 2 Schematic diagram of the process for generating a candidate recommended product set according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The present invention will be described in detail below in conjunction with the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0060] It should be pointed out that in the specification, it is mentioned that "an embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not necessarily every embodiment includes such specific features, structures, or characteristics. In addition, when combining embodiments to describe specific features, structures, or characteristics, implementing such features, structures, or characteristics in combination with other embodiments (whether explicitly described or not) should be within the knowledge scope of those skilled in the relevant art.

[0061] Generally, terms can be understood at least in part from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or property in the singular sense, or can be used to describe a combination of features, structures, or properties in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but rather, depending at least in part on the context, can allow for other factors that may not be explicitly described.

[0062] As Figure 1 - Figure 2 shown, a product recommendation method for an electric power marketing management information platform includes the following steps:

[0063] S1: Collect user data of the electric power marketing management information platform, including user historical purchase records, power consumption data, and user feedback data;

[0064] S2: Standardize the user data collected in S1 to generate standardized user data;

[0065] S3: Build a dynamic user profile based on the standardized user data, where the dynamic user profile includes a static attribute layer and a dynamic behavior layer;

[0066] S4: Match the dynamic user profile with the electric power product database to generate a set of candidate recommended products;

[0067] S5: Combine the real-time operation strategy data of the electric power marketing platform to perform two-level sorting on the set of candidate recommended products to generate a personalized recommendation list;

[0068] S6: Push the personalized recommendation list generated in S5 to the user terminal, and dynamically adjust the recommendation logic parameters in the two-level sorting based on the click-through rate and purchase conversion rate of the recommended products by the user.

[0069] S1 specifically includes:

[0070] S11: Extract user historical purchase records from the marketing business system of the electric power marketing management information platform, where the user historical purchase records include user identification, types of purchased electric power products, purchase time, and corresponding power consumption equipment codes;

[0071] S12: Obtain user power consumption data from the electric power metering automation system, where the power consumption data includes power consumption time series data divided by billing cycles, peak-to-valley power consumption ratios, and perform associated matching with user identification;

[0072] S13: Collect user feedback data from the power customer service system. The user feedback data includes the text feedback content actively submitted by users, the text of customer service call recordings, and the corresponding user satisfaction ratings. The text feedback content is stored in association with user identifiers. The above steps achieve a comprehensive characterization of users' electricity consumption behaviors and product preferences by uniformly collecting users' historical purchase behaviors, electricity load characteristics, and subjective feedback content and binding them to user identifiers, providing high-quality raw data support for subsequent standardized processing and dynamic portrait construction.

[0073] S2 specifically includes:

[0074] S21: Perform number mapping conversion on the types of power products and electricity equipment codes involved in the user historical purchase records collected in S1, and unify them into fixed-length numerical feature vectors.

[0075] S22: Perform periodic normalization processing on the electricity consumption time series data in the user electricity consumption data. Using the maximum-minimum normalization method, normalize the electricity consumption in each period to the interval [0, 1].

[0076] S23: Use the Z-score standardization method for the peak-to-valley electricity consumption ratio to calculate its standard deviation offset value in the overall distribution.

[0077] S24: Convert non-standard power terms in the user feedback data. By extracting power-related words from the user feedback text and matching the words with a preset power standard term library, replace them with the corresponding standard terms.

[0078] S25: Concatenate and integrate all the feature vectors processed in S21 to S24 according to user identifiers to generate standardized user data with unified dimensions and consistent formats. The above steps achieve the consistency of data structure and the unity of numerical distribution by performing specific standardized processing on various types of information in the user raw data, providing an accurate, efficient, and comparable input data basis for subsequent user portrait construction and model calculation, thereby enhancing the generalization ability and adaptability of the recommendation method among different user groups.

[0079] S3 specifically includes:

[0080] S31: Select information reflecting users' basic characteristics from the standardized user data to construct a static attribute layer. The static attribute layer includes the geographical region where the user belongs, the user industry classification, and the types of users' commonly used electricity equipment.

[0081] S32: Select information that can reflect users' consumption preferences and electricity consumption behavior patterns from the standardized user data to construct a dynamic behavior layer. The dynamic behavior layer includes the purchase frequency within a continuous time period, the electricity peak-valley fluctuation pattern, and the feedback evaluation trend.

[0082] S33: Perform correlation mapping on the static attribute layer and the dynamic behavior layer, establish a cross-layer connection relationship through the user identifier, so that data from different feature dimensions form a multi-dimensional attribute representation with a unified structure in the same user portrait, thereby generating a dynamic user portrait; the above steps realize the comprehensive characterization of the long-term characteristics and short-term behaviors of users through the hierarchical construction of static attributes and dynamic behaviors of standardized user data, facilitating the subsequent accurate positioning of users' multi-dimensional characteristics in the matching and recommendation processes, thereby improving the flexibility and accuracy of product recommendations.

[0083] S33 specifically includes:

[0084] S331: Correspondingly match the data of the static attribute layer and the dynamic behavior layer according to the user identifier, and construct an initial mapping pair set, which is specifically represented as a triple <U, A, B>, where U represents the user identifier; A represents the static attribute vector; B represents the dynamic behavior vector;

[0085] S332: Perform a concatenation operation on vectors A and B to construct a fused feature vector F, and its expression is: F = α1·a1, α2·a2,..., α i ·a i , β1·b1, β2·b2,..., β j ·b j , where: α i represents the weight factor of the static attribute a i ; β j represents the weight factor of the dynamic behavior b j ;

[0086] S333: Bind the fused feature vector F to the user identifier U to form the final dynamic user portrait D, and the expression is: D = {U, F}; the above steps effectively improve the discrimination ability and modeling accuracy of the dynamic user portrait through the structured fusion of the static attribute vector and the dynamic behavior vector, providing a stable and discriminative data basis for the subsequent product recommendation matching strategy.

[0087] S4 specifically includes:

[0088] S41: Extract the fused feature vector of the user based on the dynamic user portrait, and extract the attribute feature vectors of various power products from the power product database;

[0089] S42: Use the feature similarity measurement method to calculate the matching similarity between the user fused feature vector and the power product attribute feature vector, and obtain the product similarity score for each user;

[0090] S43: Perform threshold screening on the product similarity scores, and initially screen the power products that meet the similarity threshold conditions to form an initial product set;

[0091] S44: Deduplicate the power products in the initial product set to form a candidate recommended product set for each user. The above steps achieve the precise correspondence between user needs and product attributes through the matching and similarity calculation of the dynamic user profile and the characteristics of the power product database, effectively improving the accuracy and relevance of the candidate recommended product set and providing a high-quality data input basis for the subsequent recommendation process.

[0092] S42 specifically includes:

[0093] S421: Normalize the fused feature vector of the user and the attribute feature vector of the power product respectively to generate a vector form with a unified numerical range.

[0094] S422: Use the cosine similarity method to calculate the matching similarity between the normalized fused feature vector of the user and the attribute feature vector of the power product. The specific calculation formula is: where: Sim is the similarity score between the fused feature vector of the user and the attribute feature vector of the power product; u k is the k-th eigenvalue in the fused feature vector of the user; v k is the k-th eigenvalue in the attribute feature vector of the power product; p is the dimension of the feature vector. The above steps achieve the precise quantitative matching between user needs and power product attributes through a clear cosine similarity calculation method, effectively improving the accuracy in the candidate product recommendation process.

[0095] The threshold conditions in step S43 specifically include:

[0096] Threshold condition 1: Set a similarity score threshold T sim , and screen and retain the power products whose product similarity score Sim calculated in step S42 is greater than or equal to the threshold T sim .

[0097] Threshold condition 2: Set a product popularity threshold T hot , and select and retain the products whose purchase frequency is greater than or equal to the threshold T hot according to the purchase frequency of the power products by users in the historical period.

[0098] Threshold condition 3: Set a user feedback score threshold T fb , obtain the historical user satisfaction scores of the power products from the power customer service system, and retain the products whose user satisfaction scores are greater than or equal to the threshold T fb .

[0099] Threshold condition 4: Set a product inventory threshold T inv , detect the inventory quantity of the power products, and retain the power products whose inventory quantity is greater than or equal to the threshold T inv .

[0100] S5 specifically includes:

[0101] S51: Based on real-time operation strategy data, obtain the promotion weight of power products, and perform a primary sorting on the products in the candidate recommended product set according to the promotion weight to generate a primary sorted list; the calculation formula for the promotion weight is: W op (j) = γ × Q(j) + (1 - γ) × R(j), where, W op (j) represents the primary sorting promotion weight of candidate product j; Q(j) is the current promotion coefficient of candidate product j defined in the real-time operation strategy; R(j) is the historical sales heat coefficient of candidate product j in the operation strategy; γ is the trade-off coefficient between the promotion coefficient and the historical sales heat coefficient, and its value range is [0, 1];

[0102] S52: Calculate the user preference matching degree according to the user fusion feature vector and the product feature vector in the primary sorted list, and generate a personalized matching score P(u, j), and the expression is: where: P(u, j) represents the personalized preference matching score of user u for candidate product j; u k represents the value of the kth feature in the user fusion feature vector; v k represents the value of the kth feature in the product feature vector; p is the dimension of the feature vector;

[0103] S53: Perform a secondary re-sorting on the primary sorted list according to the calculated personalized matching score P(u, j) to generate a personalized recommendation list, and the expression for the secondary re-sorting is:

[0104] L(u) = Sort{P(u, j)}, j ∈ J, where, L(u) is the final personalized recommendation list of user u; J is the candidate recommended product set after primary sorting; Sort{·} is a function that sorts from high to low according to the matching score; the above steps realize the balanced optimization of the overall marketing goal of the platform and the individual needs of users through a two-level sorting mechanism combining operation strategy weights and user personalized preferences, improve the accuracy and personalized experience of the recommendation list, and ensure that the recommendation strategy is more practical and effective.

[0105] S6 specifically includes:

[0106] S61: Push the personalized recommendation list generated by S5 to the user terminal interface in the form of interactive messages through the data communication interface of the power marketing management information platform, and record the browsing behavior of the user terminal for the recommended products;

[0107] S62: Calculate the click-through rate and purchase conversion rate of the user for each recommended product according to the browsing records of the user terminal for the recommended products;

[0108] The specific calculation formulas are as follows:

[0109] Calculation formula for click-through rate:

[0110] Calculation formula for purchase conversion rate:

[0111] Wherein, CTR(j) is the click-through rate of product j; CR(j) is the purchase conversion rate of product j; N click (j) is the number of user clicks on product j; N show (j) is the number of times product j is shown to users; N buy (j) is the number of times users actually purchase product j;

[0112] S63: According to the changes in the click-through rate CTR(j) and purchase conversion rate CR(j) of the product, the gradient optimization method is used to dynamically adjust the trade-off coefficient γ between the promotion coefficient and the historical sales popularity coefficient in 55; and then the adjusted trade-off coefficient is transmitted back to the two-level sorting algorithm in step S5 in real time to re-adjust the generation logic parameters of the subsequent recommendation list; the above steps dynamically optimize the recommendation logic parameters by collecting the click and purchase feedback behaviors of users in real time, realize the adaptive update of the product recommendation strategy, effectively improve the continuous accuracy of the personalized recommendation results and user satisfaction, and ensure that the recommendation method has good dynamic response ability and actual marketing effect.

[0113] S63 specifically includes:

[0114] S631: Based on the click-through rate and purchase conversion rate data within two consecutive statistical periods, calculate the click-through rate change ΔCTR(j) and purchase conversion rate change ΔCR(j) of the recommended product respectively. The calculation formulas are:

[0115] Click-through rate change:

[0116] Purchase conversion rate change:

[0117] Wherein, is the click-through rate of product j in the current period; is the click-through rate of product j in the previous period; is the purchase conversion rate of product j in the current period; is the purchase conversion rate of product j in the previous period;

[0118] S632: Construct a loss function L(γ) for parameter adjustment with the click-through rate change and purchase conversion rate change. The specific expression is: L(γ) = λ × (ΔCTR(j)) 2 + (1 - λ) × (ΔCR(j))2 where \(L(\gamma)\) represents the loss function value of the trade-off coefficient \(\gamma\); \(\lambda\) is the adjustment ratio coefficient of the click-through rate change and the purchase conversion rate change, and its value range is \([0, 1]\), which is set by the operation strategy;

[0119] S633: Calculate the gradient of the trade-off coefficient \(\gamma\) based on the loss function \(L(\gamma)\) to obtain the gradient value of the loss function with respect to the trade-off coefficient \(\gamma\). The gradient calculation formula is:

[0120] where, is the gradient value of the loss function \(L(\gamma)\) with respect to the trade-off coefficient \(\gamma\); represents the partial derivative of the click-through rate with respect to the trade-off coefficient \(\gamma\); represents the partial derivative of the purchase conversion rate with respect to the trade-off coefficient \(\gamma\);

[0121] S634: Update the trade-off coefficient according to the gradient descent optimization method. The calculation formula is: where \(\gamma\) new is the updated trade-off coefficient; \(\gamma\) old is the trade-off coefficient of the previous cycle; \(\eta\) is the learning rate, and its value range is \((0, 1)\), which is used to control the amplitude of each update. Through the above steps, by establishing a clear loss function and gradient optimization update mechanism, the change of user feedback can be captured in real time, the trade-off coefficient can be dynamically adjusted, the refinement and adaptive update of the recommendation strategy parameters can be realized, the matching degree between the recommendation result and the user's real-time preference can be effectively improved, and the marketing effect of the platform can be enhanced.

[0122] The present invention covers any substitutions, modifications, equivalent methods, and solutions made within the essence and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without these detailed descriptions. In addition, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0123] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A product recommendation method for an electric power marketing management information platform, characterized in that, It includes the following steps: S1: Collect user data from the power marketing management information platform, including user historical purchase records, power consumption data, and user feedback data; S2: Standardize the user data collected in S1 to generate standardized user data; S3: Build a dynamic user portrait based on the standardized user data, where the dynamic user portrait includes a static attribute layer and a dynamic behavior layer; S4: Match the dynamic user portrait with the power product database to generate a set of candidate recommended products; S5: Combine the real-time operation strategy data of the power marketing platform to perform two-level sorting on the set of candidate recommended products to generate a personalized recommendation list; S6: Push the personalized recommendation list generated in S5 to the user terminal, and dynamically adjust the recommendation logic parameters in the two-level sorting based on the click-through rate and purchase conversion rate of the recommended products by the user.

2. The product recommendation method for an electric power marketing management information platform according to claim 1, wherein The specific content of S1 includes: S11: Extract user historical purchase records from the marketing business system of the power marketing management information platform, where the user historical purchase records include user identification, types of purchased power products, purchase time, and corresponding power consumption equipment codes; S12: Obtain user power consumption data from the power metering automation system, where the power consumption data includes power consumption time series data divided by billing cycles, peak-to-valley power consumption ratios, and perform associated matching with user identification; S13: Collect user feedback data from the power customer service system, where the user feedback data includes text feedback content actively submitted by users, text of customer service call recordings, and corresponding user satisfaction ratings, and the text feedback content is stored in an associated manner with user identification.

3. A product recommendation method for an electric power marketing management information platform according to claim 1, characterized in that, The specific content of S2 includes: S21: Perform number mapping conversion on the types of power products and power consumption equipment codes involved in the user historical purchase records collected in S1 to uniformly convert them into fixed-length numerical feature vectors; S22: Perform periodic normalization on the power consumption time series data in the user power consumption data, using the maximum-minimum normalization method to normalize the power consumption in each period to the [0,1] interval; S23: Use the Z-score standardization method for the peak-to-valley power consumption ratio to calculate its standard deviation offset value in the overall distribution; S24: Convert non-standard power terms in the user feedback data by extracting power-related words from the user feedback text and matching the words with a preset power standard term library to replace them with corresponding standard terms; S25: Concatenate and integrate all the feature vectors processed by S21 to S24 according to user identification to generate standardized user data with unified dimensions and consistent formats.

4. A product recommendation method for an electric power marketing management information platform according to claim 1, characterized in that The specific content of S3 includes: S31: Select information reflecting the basic characteristics of users in the standardized user data to build a static attribute layer, where the static attribute layer includes the geographical region to which the user belongs, user industry classification, and types of commonly used power consumption equipment of the user; S32: Select information that can reflect the consumption preferences and power consumption behavior patterns of users in the standardized user data to build a dynamic behavior layer, where the dynamic behavior layer includes the purchase frequency within a continuous time period, peak-to-valley power consumption fluctuation patterns, and feedback evaluation trends. S33: Perform an association mapping between the static attribute layer and the dynamic behavior layer, establish a cross-layer connection relationship through the user identifier, so that data from different feature dimensions form a multi-dimensional attribute representation with a unified structure in the same user portrait, thereby generating a dynamic user portrait.

5. A product recommendation method for an electric power marketing management information platform according to claim 4, characterized in that The specific steps of S33 are as follows: S331: Correspondingly match the data of the static attribute layer and the dynamic behavior layer according to the user identifier, and construct an initial set of mapping pairs, which is specifically represented as a triple <U, A, B>, where U represents the user identifier; A represents the static attribute vector; B represents the dynamic behavior vector. S332: Concatenate vectors A and B to construct a fused feature vector F, whose expression is: F = [α1·a1, α2·a2,..., α i ·a i , β1·b1, β2·b2,..., β j ·b j , where: α i represents the weight factor of static attribute a i ; β j represents the weight factor of dynamic behavior b j . S333: Bind the fused feature vector F to the user identifier U to form the final dynamic user portrait D, and the expression is: D = {U, F}.

6. The product recommendation method for an electric power marketing management information platform according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Extract the fused feature vector of the user based on the dynamic user portrait, and extract the attribute feature vectors of various power products from the power product database. S42: Use a feature similarity measurement method to calculate the matching similarity between the user fused feature vector and the power product attribute feature vector, and obtain the product similarity score for each user. S43: Perform a threshold screening on the product similarity scores, and initially screen the power products that meet the similarity threshold conditions to form an initial product set. S44: Remove duplicates from the power products in the initial product set to form a candidate recommended product set for each user.

7. A product recommendation method for an electric power marketing management information platform according to claim 6, characterized in that, The specific steps of S42 are as follows: S421: Normalize the fused feature vector of the user and the attribute feature vector of the power product respectively to generate a vector form with a unified numerical range. S422: Calculate the matching similarity between the normalized user fusion feature vector and the power product attribute feature vector using the cosine similarity method. The specific calculation formula is as follows: Where: Sim is the similarity score between the user fusion feature vector and the power product attribute feature vector; u k Is the k-th eigenvalue in the user fusion feature vector; v k Is the k-th eigenvalue in the power product attribute feature vector; p is the dimension of the feature vector.

8. A product recommendation method for a power marketing management information platform according to claim 1, characterized in that The specific steps of S5 are as follows: S51: Obtain the promotion weight of power products based on real-time operation strategy data, and conduct a primary ranking of the products in the candidate recommended product set according to the promotion weight to generate a primary ranking list; the calculation formula for the promotion weight is: W op (j) = γ × Q(j) + (1 - γ) × R(j), where W op (j) represents the primary ranking promotion weight of candidate product j; Q(j) is the current promotion coefficient of candidate product j defined in the real-time operation strategy; R(j) is the historical sales popularity coefficient of candidate product j in the operation strategy; γ is the trade-off coefficient between the promotion coefficient and the historical sales popularity coefficient; S52: Calculate the user preference matching degree according to the user fused feature vector and the product feature vector in the primary sorted list, and generate a personalized matching score P(u, j). S53: Perform a secondary re-sorting on the primary sorted list according to the calculated personalized matching score P(u, j) to generate a personalized recommendation list. The expression for the secondary re-sorting is: L(u) = Sort{P(u, j)}, j ∈ J, where L(u) is the final personalized recommendation list of user u; J is the candidate recommended product set after primary sorting; Sort{·} is a function that sorts in descending order according to the matching score.

9. A product recommendation method for an electric power marketing management information platform according to claim 1, characterized in that, The specific steps of S6 are as follows: S61: Push the personalized recommendation list generated by S5 to the user terminal interface in the form of an interactive message through the data communication interface of the power marketing management information platform, and record the browsing behavior of the user terminal for the recommended products. S62: Calculate the click-through rate and purchase conversion rate of the user for each recommended product according to the browsing records of the user terminal for the recommended products. S63: Dynamically adjust the trade-off coefficient γ between the promotion coefficient and the historical sales popularity coefficient in S5 according to the changes in the click-through rate CTR(j) and purchase conversion rate CR(j) of the product by using the gradient optimization method.

10. A product recommendation method for a power marketing management information platform according to claim 9, characterized in that, The specific steps of S63 are as follows: S631: Based on the click-through rate and purchase conversion rate data in two consecutive statistical periods, calculate the click-through rate change amount ΔCTR(j) and purchase conversion rate change amount ΔCR(j) of the recommended product respectively. S632: Construct a loss function \(L(\gamma)\) for parameter adjustment using the change in click-through rate and the change in purchase conversion rate; S633: Calculate the gradient of the trade-off coefficient \(\gamma\) based on the loss function \(L(\gamma)\) to obtain the gradient value of the loss function with respect to the trade-off coefficient \(\gamma\). The gradient calculation formula is: Among them, is the gradient value of the loss function L(γ) with respect to the trade-off coefficient γ; represents the partial derivative of the click-through rate with respect to the trade-off coefficient γ; represents the partial derivative of the purchase conversion rate with respect to the trade-off coefficient γ; S634: Update the trade-off coefficient according to the gradient descent optimization method, and the calculation formula is: where γ new is the updated trade-off coefficient; γ old is the trade-off coefficient of the previous cycle; η is the learning rate.

Images