Digital marketing recommendation method and system based on big data

Abstract

The invention discloses a digital marketing recommendation method and system based on big data, and relates to the technical field of commodity recommendation. The method comprises the steps of obtaining original data of an e-commerce platform, obtaining a structured data set through preprocessing, performing fuzzy clustering on a scoring matrix, generating and iteratively updating a membership matrix, reducing the influence of initial clustering deviation, calculating clustering attribution in combination with user geographic information, and correcting scores according to the clustering attribution to obtain a final scoring matrix. Inputting the matrix and the user and commodity feature vectors into a recommendation model, and outputting a commodity recommendation result; clustering multiple membership degrees of the users by adopting fuzzy clustering, and adapting to complex overlapping characteristics of user preferences under sparse data of the distributors; the initial deviation is reduced by iteratively optimizing the membership degree, attribution is determined in combination with geographic information, and the clustering fitness and distinction degree are improved; the commodity recommendation model effectively solves the problem of low accuracy of the commodity recommendation result under the condition of reducing the calculation complexity.

Description

Technical Field

[0001] This invention belongs to the field of product recommendation technology, specifically relating to a digital marketing recommendation method and system based on big data. Background Technology

[0002] The e-commerce industry is currently experiencing explosive growth in both user base and product categories. Traditional recommendation models, relying on single-dimensional data, are susceptible to limitations such as data sparsity and rating bias, resulting in low accuracy and conversion rates. Massive amounts of big data resources, including user behavior, product attributes, and geographic information, have not been fully explored, making it difficult to capture potential user preferences and group characteristics. Against this backdrop, there is a need for big data-driven digital product recommendation technologies to accurately match supply and demand, improve user experience, and enhance platform operational efficiency.

[0003] The existing publication number (CN118365431A) discloses a product recommendation method and system for e-commerce platforms based on big data. This method includes analyzing the visual information of product images based on seller images, combining product descriptions and user reviews to analyze product characteristics and consumer sentiment, and generating product feature identification information. In this invention, by extracting product characteristic and consumer sentiment information and combining cross-validation of visual and textual information, the accuracy and completeness of product tags are enhanced, the descriptiveness and targeting of the product tagging system are optimized, and product demand is predicted by integrating user interaction data and a product tag database, combined with seasonal changes and historical sales data.

[0004] The above methods combine product descriptions and user reviews, and extract product characteristics and consumer sentiment information to predict product demand. However, some newly established e-commerce platforms may experience data sparsity. In such cases, the characteristic data of merchants and users are insufficient to support accurate predictions of product marketing, while the recommendation results output by traditional product recommendation models are relatively inaccurate. Summary of the Invention

[0005] The purpose of this invention is to solve the problem that it is difficult to support accurate prediction of product marketing in data-sparse scenarios, and the recommendation results output by traditional product recommendation models are not very accurate. Therefore, this invention proposes a digital marketing recommendation method and system based on big data.

[0006] In a first aspect of this invention, a digital marketing recommendation method based on big data is proposed, the method comprising: Obtain raw data from an e-commerce platform, and preprocess the raw data to obtain a structured dataset; the structured dataset includes a rating matrix, user feature vectors, product feature vectors, and a list of public product ratings; Fuzzy clustering is performed on the scoring matrix to generate an initial membership matrix, and the initial membership matrix is ​​iteratively updated to obtain the final membership matrix; Collect users' geographic information, and calculate the users' clustering results based on the users' geographic information and the final membership matrix; Based on the clustering results of users, the original user ratings and common product ratings are corrected to obtain corrected user and product ratings. The corrected ratings are then subjected to preset operations to obtain the final rating matrix. The final rating matrix, user feature vector, and product feature vector are input into the product recommendation model to output product recommendation results, which are then uploaded to the cloud.

[0007] By preprocessing raw data to generate a structured dataset, the problems of fragmented data dimensions and insufficient effective correlation are alleviated. Fuzzy clustering iteratively optimizes the membership matrix, and geographic information is combined to determine the user's cluster affiliation, making up for the lack of user preference features under sparse data. The rating is corrected based on the clustering results, weakening abnormal interference and improving the reliability of the rating matrix. Finally, the multi-dimensional data is input into the recommendation model and uploaded to the cloud, effectively improving the accuracy and stability of product recommendation in data-sparse scenarios.

[0008] Optionally, the initial membership matrix is ​​iteratively updated to obtain the final membership matrix, including: The rating matrix is ​​split into rows to obtain the rating vector for each user, and an initial membership matrix is ​​generated based on the rating vectors. The cluster centers after iteration are calculated using preset parameters based on the rating vectors of all users and the initial membership degree corresponding to each cluster. Calculate the distance from each user rating vector to each cluster center based on the iterated cluster centers, and calculate the membership degree after iteration based on the distance relationship and preset parameters; The difference between the membership degree after the iteration and the membership degree after the previous iteration is calculated. The difference value is compared with the iteration termination threshold. The iteration is continued or terminated according to the judgment result, and the final membership degree matrix is ​​output.

[0009] By grouping users with similar consumption preferences and behaviors using fuzzy clustering, the problem of data sparsity in personalized product recommendations can be effectively alleviated. At the same time, data complexity is reduced, key consumption characteristics and regional differences of user groups are preserved, and a precise user grouping basis is provided for subsequent rating conversion and consumption probability calculation, thereby improving the targeting and efficiency of recommendations.

[0010] Optionally, the user's clustering classification result is calculated based on the user's geographic information and the final membership matrix, including: Step 1: Select multiple users within the target area in the geographic information, and extract the membership degree of the users in the area to the target cluster center from the final membership matrix; the target cluster center is any one of multiple cluster centers, and the target area is any location in the geographic information. Step 2: Sort the membership degrees in ascending order to obtain a membership degree sequence, and select a preset proportion of membership degrees from the membership degree sequence as the membership degree threshold; Step 3: Compare the membership degree of the users in the target area in Step 1 with the membership degree threshold. If the membership degree of a user is greater than the membership degree threshold, then the user belongs to the target cluster center. If the membership degree of a user is not greater than the membership degree threshold, then the user does not belong to the target cluster center. Step 4: For each cluster center, perform steps 1 to 3 sequentially to obtain the clustering results for all users.

[0011] By leveraging the high similarity of consumption preferences among users in the same geographical area and using objective data to calculate thresholds to accurately classify user clusters, the clarity and reliability of clustering results are enhanced, while further mitigating data sparsity.

[0012] Optionally, the final rating matrix is ​​obtained by modifying the preset rating, including: The original ratings of users belonging to the target cluster center and the public ratings of products are calculated using preset correction parameters to obtain the corrected ratings. Preference judgment indicators are calculated based on users' original ratings and product public ratings, and preference levels are set based on these indicators. The number of users for each preference level was counted, and the preference probability for each preference level was calculated. The final rating matrix was constructed based on the adjusted rating, preference level, and preference probability.

[0013] This method addresses the problems of traditional recommendation models, such as relying solely on a single original score, ignoring the influence of public scores, failing to reflect differences in user preferences under the same score, and neglecting the repetitive nature of product consumption and diminishing marginal utility. It first corrects the original score to objectively reflect the balance between individual preferences and group consensus, then quantifies the dynamic impact of the number of consumptions on the probability of repeat consumption, and finally obtains a final score that better reflects the actual consumption scenario by correcting the product of the score and the consumption probability.

[0014] Optionally, the working principle of the product recommendation model is as follows: The input data is acquired and fed into the fully connected layer to output the first target feature. The first target feature is fed into the CNN module to obtain the second target feature, and activated by the ReLU function to obtain the third target feature. The second target feature is fed into two GNN modules in sequence to obtain the fourth target feature. The first, third and fourth target features are concatenated and fused to obtain the fifth target feature. The fifth target feature is processed by the LDA module to obtain the sixth target feature. The sixth target feature is then input into the channel attention module to obtain the first channel feature and the second channel feature. The first channel feature and the second channel feature are then fused with the sixth target feature and input into the spatial attention module to obtain the third channel feature and the fourth channel feature, respectively. The features of channel 1 and channel 3 are fused to obtain the features of channel 5, and the features of channel 2 and channel 4 are fused to obtain the features of channel 6. The features of channel 5 and channel 6 are concatenated and then input into the LSA module and the fully connected layer in sequence. After activation by the Sigmoid function, the recommendation result is output.

[0015] Compared to traditional recommendation models that only consider single-granularity sentiment analysis or rely solely on rating data, this model organically integrates cross-granularity sentiment analysis with the final rating matrix decomposition. It utilizes LDA (Latent Dirichlet Distribution) to extract comprehensive user preference features, and through spatial attention and channel attention mechanisms, assigns higher weights to key user features (such as geographic information and occupation) while suppressing irrelevant information to achieve a deep fusion of sentiment features and latent factors. This compensates for the lack of information in single rating data and solves the problem that single-granularity sentiment analysis cannot fully capture user preferences.

[0016] In a second aspect of this invention, a digital marketing recommendation system based on big data is proposed, the system comprising: The preprocessing module is used to acquire raw data from the e-commerce platform and preprocess the raw data to obtain a structured dataset; the structured dataset includes a rating matrix, user feature vectors, product feature vectors, and a list of public product ratings; The clustering module is used to perform fuzzy clustering on the scoring matrix to generate an initial membership matrix, and to iteratively update the initial membership matrix to obtain the final membership matrix. The results calculation module is used to collect the user's geographic information and calculate the user's clustering and affiliation results based on the user's geographic information and the final membership matrix. The correction module is used to correct the user's original rating and the product's common rating based on the user's clustering and affiliation results to obtain the corrected rating for the user and the product, and to perform preset operations on the corrected rating to obtain the final rating matrix. The upload module is used to input the final rating matrix, user feature vector, and product feature vector into the product recommendation model to output product recommendation results, and then upload the product recommendation results to the cloud.

[0017] Optionally, the clustering module includes: The splitting module is used to split the rating matrix by rows to obtain the rating vector for each user, and generate an initial membership matrix based on the rating vector; The first calculation module is used to calculate the cluster centers after iteration based on the rating vectors of all users and the initial membership degree corresponding to each cluster using preset parameters. The second calculation module is used to calculate the distance from each user rating vector to each cluster center based on the iterated cluster centers, and to calculate the membership degree after iteration based on the distance relationship and preset parameters. The judgment module is used to calculate the difference between the membership degree after the iteration and the membership degree after the previous iteration, judge the iteration based on the difference value and the iteration termination threshold, and execute the iteration to continue or terminate according to the judgment result, and output the final membership degree matrix.

[0018] Optionally, the result calculation module includes an attribution diagnosis module, wherein: The attribution diagnosis module is used to execute steps 1 to 4, and the process is as follows: Step 1: Select multiple users within the target area in the geographic information, and extract the membership degree of the users in the area to the target cluster center from the final membership matrix; the target cluster center is any one of multiple cluster centers, and the target area is any location in the geographic information. Step 2: Sort the membership degrees in ascending order to obtain a membership degree sequence, and select a preset proportion of membership degrees from the membership degree sequence as the membership degree threshold; Step 3: Compare the membership degree of the users in the target area in Step 1 with the membership degree threshold. If the membership degree of a user is greater than the membership degree threshold, then the user belongs to the target cluster center. If the membership degree of a user is not greater than the membership degree threshold, then the user does not belong to the target cluster center. Step 4: For each cluster center, perform steps 1 to 3 sequentially to obtain the clustering results for all users.

[0019] Optionally, the correction module includes: The third calculation module is used to calculate the corrected score from the original score of the user belonging to the target cluster center and the public score of the product using preset correction parameters. The rating setting module is used to calculate preference judgment indicators based on the user's original rating and the product's public rating, and to set preference levels based on the preference judgment indicators; The module is used to count the number of users for each preference level and calculate the preference probability for each preference level. Based on the corrected score, preference level, and preference probability, the final score matrix is ​​constructed.

[0020] Optionally, the upload module includes the working principle of the product recommendation model: The input data is acquired and fed into the fully connected layer to output the first target feature. The first target feature is fed into the CNN module to obtain the second target feature, and activated by the ReLU function to obtain the third target feature. The second target feature is fed into two GNN modules in sequence to obtain the fourth target feature. The first, third and fourth target features are concatenated and fused to obtain the fifth target feature. The fifth target feature is processed by the LDA module to obtain the sixth target feature. The sixth target feature is then input into the channel attention module to obtain the first channel feature and the second channel feature. The first channel feature and the second channel feature are then fused with the sixth target feature and input into the spatial attention module to obtain the third channel feature and the fourth channel feature, respectively. The features of channel 1 and channel 3 are fused to obtain the features of channel 5, and the features of channel 2 and channel 4 are fused to obtain the features of channel 6. The features of channel 5 and channel 6 are concatenated and then input into the LSA module and the fully connected layer in sequence. After activation by the Sigmoid function, the recommendation result is output.

[0021] The beneficial effects of this invention are: This invention proposes a digital marketing recommendation method based on big data. It uses fuzzy clustering to assign users to multiple clusters with different membership degrees, better aligning with the complexity and overlap of user preferences in sparse e-commerce data scenarios. It iterative updates optimize the accuracy of membership degree calculation, reducing initial clustering bias. Combining geographical information with the final membership degree matrix determines user cluster affiliation, effectively improving the alignment between clustering results and actual user consumption characteristics, enhancing cluster group differentiation, and avoiding problems of high cluster overlap and ambiguous group characteristics. The product recommendation model effectively solves the problems of inaccurate and inefficient product recommendation results while reducing computational complexity. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 A flowchart illustrating a digital marketing recommendation method based on big data, provided for an embodiment of the present invention; Figure 2 This is a structural diagram of the product recommendation model provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0025] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides a digital marketing recommendation method based on big data. See also... Figure 1 , Figure 1 A flowchart illustrating a big data-based digital marketing recommendation method provided in this embodiment of the invention. The method includes the following steps: S101: Obtain the raw data from the e-commerce platform and preprocess the raw data to obtain a structured dataset; S102, perform fuzzy clustering on the scoring matrix to generate an initial membership matrix, and iteratively update the initial membership matrix to obtain the final membership matrix; S103: Collect the user's geographic information and calculate the user's clustering result based on the user's geographic information and the final membership matrix; S104, Based on the user's clustering and classification results, the user's original rating and the product's common rating are corrected to obtain the corrected ratings for the user and the product. The corrected ratings are then subjected to preset operations to obtain the final rating matrix. S105 inputs the final rating matrix, user feature vector, and product feature vector into the product recommendation model to output the product recommendation results, and uploads the product recommendation results to the cloud.

[0027] The structured dataset contains a rating matrix, user feature vectors, product feature vectors, and a list of public product ratings.

[0028] This invention provides a big data-based digital marketing recommendation method that uses fuzzy clustering to assign users to multiple clusters with different membership degrees. This better reflects the complexity and overlap of user preferences in sparse e-commerce data scenarios. By iteratively updating and optimizing the membership degree calculation accuracy, the method reduces initial clustering bias. Combining geographical information with the final membership degree matrix, it determines the user's cluster affiliation, effectively improving the alignment between clustering results and actual user consumption characteristics, enhancing the distinguishability of cluster groups, and avoiding problems such as high cluster overlap and ambiguous group characteristics. The product recommendation model effectively solves the problem of inaccurate product recommendation results while reducing computational complexity.

[0029] In one implementation, the raw data includes user's original rating data, user's feature data, product feature data, and product public rating data; the original rating data includes personal ratings; the user's feature data includes: user ID, age, gender, geographical location, and occupation; the product feature data includes product ID, type, title, and attributes; and the product public rating data includes product public ratings.

[0030] In one implementation, the raw data is preprocessed to obtain a structured dataset, including: Integrate users' original rating data to construct an m×n dimension rating matrix; n is the number of users and m is the number of products; encode user features and product features to convert them into user feature vectors and product feature vectors; deduplicate and sort the common product rating data to obtain a list of common product ratings.

[0031] In one embodiment, iteratively updating the initial membership matrix to obtain the final membership matrix includes: The rating matrix is ​​split into rows to obtain the rating vector for each user, and the initial membership matrix is ​​generated based on the rating vector. The cluster centers after iteration are calculated using preset parameters based on the rating vectors of all users and the initial membership degree corresponding to each cluster. Calculate the distance from each user rating vector to each cluster center based on the iterated cluster centers, and calculate the membership degree after iteration based on the distance relationship and preset parameters; The difference between the membership degree after the iteration and the membership degree after the previous iteration is calculated. The difference value is compared with the iteration termination threshold. The iteration is continued or terminated according to the judgment result, and the final membership degree matrix is ​​output.

[0032] In one implementation, preset parameters are used to calculate the cluster centers and membership degrees after the first iteration. The calculation process is as follows: The initial membership matrix and initial cluster centers are generated by randomly assigning elements from the scoring vector. The formula for iteratively calculating the initial cluster centers is:

[0033] in, Let i be the cluster center of the t-th iteration. Let p be the rating vector of the j-th user; p is a preset parameter; N is the total number of users participating in the clustering; and T is the total number of iterations. The initial membership degree is calculated iteratively as follows:

[0034] in, Let be the membership degree of the j-th user to the i-th cluster in the t-th iteration; User rating vector Clustering centers in the (t-1)th iteration The Euclidean distance; This represents the i-th cluster center after the (t-1)-th iteration. Let r represent the k-th cluster center after the (t-1)-th iteration; r is the number of cluster centers generated in each iteration.

[0035] In one implementation, the process of determining the difference value based on the iteration termination threshold is as follows: like If so, the iteration continues; like If the iteration terminates, then the iteration ends; δ is the preset iteration termination threshold. Output the updated membership degree and replace the corresponding positions in the initial membership degree matrix to obtain the final membership degree matrix.

[0036] In one implementation, the clustering method breaks through the limitations of subjectively setting the membership threshold in traditional clustering methods. It combines user geographic information to calculate an objective threshold, abandons the rigid classification logic of hard clustering, and quantifies the membership degree of users to multiple clusters to more objectively reflect the ambiguity and overlap of user preferences in reality. This makes the clustering results more in line with actual consumption scenarios and enhances the reliability of subsequent recommendations.

[0037] In one embodiment, the user's clustering classification result is calculated based on the user's geographic information and the final membership matrix, including: Step 1: Select multiple users within the target area in the geographic information, and extract the membership degree of the users in the area to the target cluster center from the final membership matrix; the target cluster center is any one of multiple cluster centers, and the target area is any location in the geographic information. Step 2: Sort the membership degrees in ascending order to obtain a membership degree sequence, and select a preset proportion of membership degrees from the membership degree sequence as the membership degree threshold; Step 3: Compare the membership degree of the users in the target area in Step 1 with the membership degree threshold. If the membership degree of a user is greater than the membership degree threshold, then the user belongs to the target cluster center. If the membership degree of a user is not greater than the membership degree threshold, then the user does not belong to the target cluster center. Step 4: For each cluster center, perform steps 1 to 3 sequentially to obtain the clustering results for all users.

[0038] In one implementation, the preset ratio can be set to the top 20%. Introducing geographic information as a key dimension in the user clustering determination process compensates for the shortcomings of relying solely on the scoring matrix to cluster, which ignores differences in user geographic attributes, and addresses the limitations of existing technologies that rely on subjective human judgment for thresholds or do not consider regional consumption characteristics. In e-commerce scenarios, users' consumption preferences and product demands are often closely related to geographical regions. Combining geographic information with the final membership matrix to calculate clustering results allows the clustering results to better reflect users' actual consumption characteristics, while enhancing the distinguishability between different cluster groups. This avoids problems such as high cluster overlap and unclear group characteristics, effectively improving the objectivity and accuracy of clustering.

[0039] In one embodiment, obtaining the final score matrix by modifying the score preset operation includes: The original ratings of users belonging to the target cluster center and the public ratings of products are calculated using preset correction parameters to obtain the corrected ratings. Preference judgment indicators are calculated based on users' original ratings and product public ratings, and preference levels are set based on these indicators. The number of users for each preference level was counted, and the preference probability for each preference level was calculated. The final rating matrix was constructed based on the adjusted rating, preference level, and preference probability.

[0040] In one implementation, the corrected score and preference judgment index are calculated using preset correction parameters, as follows:

[0041] in, This represents the revised rating of the i-th user for the j-th product. Let be the preference index for the i-th user regarding the j-th product. This represents the original rating of the i-th user for the j-th product. Let be the public rating of the j-th product; α, β, and γ are all preset adjustable parameters, which can take values ​​of 0.053, 3, and 0.25, respectively; λ is the ratio of the maximum value of the public rating to the original rating, which can be 2 in the experiment; η is the shrinkage parameter, which can take a value of 0.5.

[0042] In one implementation, the preference level setting includes: Preference levels are set based on preference judgment indicators. The preference level is set according to the preset level threshold (ζ1,ζ2,ζ3,ζ4,ζ5,ζ6)=(1,1.5,2.25,3,3.75,4.5);

[0043] The above-mentioned level setting formula corresponds to the following from top to bottom: extreme aversion (-3), very aversion (-2), aversion (-1), indifference (0), like (1), very like (2), and extreme like (3).

[0044] In one implementation, the specific process of constructing the final scoring matrix includes:

[0045] in, This indicates that the j-th item is at the preference level. The probability of preference under; The total number of users evaluating the j-th product; This represents the number of users who belong to the q-th preference level among the total number of users who rated the j-th product; This represents the probability that the i-th user will purchase the j-th item again. This represents the final rating of the i-th user for the j-th product; a, b, c, and ω are all adjustable parameters, which were set to (-0.025, 0.9, 0.8, and 0.9) in the experiment. Indicates revised rating The maximum value of d is d, where d is the number of times the j-th item has been consumed. Construct an empty matrix, and fill the corresponding positions of the empty matrix with the final scores to obtain the final score matrix.

[0046] One implementation addresses the problems of traditional recommendations that rely solely on individual original ratings, ignore the influence of public ratings, and fail to reflect differences in user preferences under the same original ratings by constructing a final rating matrix. By quantitatively integrating two psychological effects, the original ratings are precisely corrected, and a more objective quantitative value of user preferences is output.

[0047] In one implementation, this method overcomes the limitations of existing technologies that only consider certain psychological effects or lack integration of rating and consumption probability systems. It quantifies and integrates the herd effect (individual preferences converging to public ratings) and the third-person effect (preserving individual true preferences) of consumption psychology through mathematical formulas. Based on public ratings and user original ratings, the weights of the two are reasonably allocated through parameters: public rating related items account for 25%, and original rating retained items account for 75%. Preference levels are divided by consumption probability, thereby quantifying the decreasing impact of consumption frequency on probability. This achieves the unity of user static preferences, dynamic consumption behavior, and psychological effects. Interval constraints ensure that the correction intervals of different original ratings are neither excessively overlapping nor too narrow, effectively achieving a balance between group consensus and individual differences, and accurately distinguishing user preference differences under the same original rating.

[0048] In one embodiment, the product recommendation model works as follows: The input data is acquired and fed into the fully connected layer to output the first target feature. The first target feature is fed into the CNN module to obtain the second target feature, and activated by the ReLU function to obtain the third target feature. The second target feature is fed into two GNN modules in sequence to obtain the fourth target feature. The first, third and fourth target features are concatenated and fused to obtain the fifth target feature. The fifth target feature is processed by the LDA module to obtain the sixth target feature. The sixth target feature is then input into the channel attention module to obtain the first channel feature and the second channel feature. The first channel feature and the second channel feature are then fused with the sixth target feature and input into the spatial attention module to obtain the third channel feature and the fourth channel feature, respectively. The features of channel 1 and channel 3 are fused to obtain the features of channel 5, and the features of channel 2 and channel 4 are fused to obtain the features of channel 6. The features of channel 5 and channel 6 are concatenated and then input into the LSA module and the fully connected layer in sequence. After activation by the Sigmoid function, the recommendation result is output.

[0049] In one implementation, the product recommendation module runs on a virtual machine on a server. The training hardware environment includes: CPU: Intel Core i9-9900KF, GPU: GeForce RTX 2080 Ti, and 64G of memory; the software environment includes: operating system: Windows Server 2019, Python 3.7, text processing: Nltk, LDA topic modeling: Gensim; and the dataset used is a self-built dataset.

[0050] In one implementation, a fully connected layer is used to comprehensively integrate and optimize the hyperparameters of CNN (Convolutional Neural Network) and GNN (Graph Neural Network), thereby accelerating the model convergence speed, improving the efficiency of feature selection, and avoiding interference from invalid features.

[0051] In one implementation, LSA (Latent Semantic Analysis) is used to simultaneously extract fine-grained (clause-level, such as words, phrases, and clauses) and coarse-grained (sentence-level) sentiment features of reviews. These features are then combined with latent factors from the final user and product rating matrix decomposition to accurately predict user ratings for items, thereby addressing the issues of data sparsity and poor interpretability.

[0052] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A digital marketing recommendation method based on big data, characterized in that, The method includes: Obtain raw data from an e-commerce platform, and preprocess the raw data to obtain a structured dataset; the structured dataset includes a rating matrix, user feature vectors, product feature vectors, and a list of public product ratings; Fuzzy clustering is performed on the scoring matrix to generate an initial membership matrix, and the initial membership matrix is ​​iteratively updated to obtain the final membership matrix; Collect users' geographic information, and calculate the users' clustering results based on the users' geographic information and the final membership matrix; Based on the clustering results of users, the original user ratings and common product ratings are corrected to obtain corrected user and product ratings. The corrected ratings are then subjected to preset operations to obtain the final rating matrix. The final rating matrix, user feature vector, and product feature vector are input into the product recommendation model to output product recommendation results, which are then uploaded to the cloud.

2. The digital marketing recommendation method based on big data according to claim 1, characterized in that, The initial membership matrix is ​​iteratively updated to obtain the final membership matrix, including: The rating matrix is ​​split into rows to obtain the rating vector for each user, and an initial membership matrix is ​​generated based on the rating vectors. The cluster centers after iteration are calculated using preset parameters based on the rating vectors of all users and the initial membership degree corresponding to each cluster. Calculate the distance from each user rating vector to each cluster center based on the iterated cluster centers, and calculate the membership degree after iteration based on the distance relationship and preset parameters; The difference between the membership degree after the iteration and the membership degree after the previous iteration is calculated. The difference value is compared with the iteration termination threshold. The iteration is continued or terminated according to the judgment result, and the final membership degree matrix is ​​output.

3. The digital marketing recommendation method based on big data according to claim 1, characterized in that, The user's clustering results are calculated based on the user's geographic information and the final membership matrix, including: Step 1: Select multiple users within the target area in the geographic information, and extract the membership degree of the users in the area to the target cluster center from the final membership matrix; the target cluster center is any one of multiple cluster centers, and the target area is any location in the geographic information. Step 2: Sort the membership degrees in ascending order to obtain a membership degree sequence, and select a preset proportion of membership degrees from the membership degree sequence as the membership degree threshold; Step 3: Compare the membership degree of the users in the target area in Step 1 with the membership degree threshold. If the membership degree of a user is greater than the membership degree threshold, then the user belongs to the target cluster center. If the membership degree of a user is not greater than the membership degree threshold, then the user does not belong to the target cluster center. Step 4: For each cluster center, perform steps 1 to 3 sequentially to obtain the clustering results for all users.

4. The digital marketing recommendation method based on big data according to claim 1, characterized in that, The final scoring matrix is ​​obtained by modifying the preset scoring method, including: The original ratings of users belonging to the target cluster center and the public ratings of products are calculated using preset correction parameters to obtain the corrected ratings. Preference judgment indicators are calculated based on users' original ratings and product public ratings, and preference levels are set based on these indicators. The number of users for each preference level was counted, and the preference probability for each preference level was calculated. The final rating matrix was constructed based on the adjusted rating, preference level, and preference probability.

5. The digital marketing recommendation method based on big data according to claim 1, characterized in that, The working principle of the product recommendation model is as follows: The input data is acquired and fed into the fully connected layer to output the first target feature. The first target feature is fed into the CNN module to obtain the second target feature, and activated by the ReLU function to obtain the third target feature. The second target feature is fed into two GNN modules in sequence to obtain the fourth target feature. The first, third and fourth target features are concatenated and fused to obtain the fifth target feature. The fifth target feature is processed by the LDA module to obtain the sixth target feature. The sixth target feature is then input into the channel attention module to obtain the first channel feature and the second channel feature. The first channel feature and the second channel feature are then fused with the sixth target feature and input into the spatial attention module to obtain the third channel feature and the fourth channel feature, respectively. The features of channel 1 and channel 3 are fused to obtain the features of channel 5, and the features of channel 2 and channel 4 are fused to obtain the features of channel 6. The features of channel 5 and channel 6 are concatenated and then input into the LSA module and the fully connected layer in sequence. After activation by the Sigmoid function, the recommendation result is output.

6. A digital marketing recommendation system based on big data, characterized in that, The system includes: The preprocessing module is used to acquire raw data from the e-commerce platform and preprocess the raw data to obtain a structured dataset; the structured dataset includes a rating matrix, user feature vectors, product feature vectors, and a list of public product ratings; The clustering module is used to perform fuzzy clustering on the scoring matrix to generate an initial membership matrix, and to iteratively update the initial membership matrix to obtain the final membership matrix. The results calculation module is used to collect the user's geographic information and calculate the user's clustering and affiliation results based on the user's geographic information and the final membership matrix. The correction module is used to correct the user's original rating and the product's common rating based on the user's clustering and affiliation results to obtain the corrected rating for the user and the product, and to perform preset operations on the corrected rating to obtain the final rating matrix. The upload module is used to input the final rating matrix, user feature vector, and product feature vector into the product recommendation model to output product recommendation results, and then upload the product recommendation results to the cloud.

7. A digital marketing recommendation system based on big data according to claim 6, characterized in that, The clustering module includes: The splitting module is used to split the rating matrix by rows to obtain the rating vector for each user, and generate an initial membership matrix based on the rating vector; The first calculation module is used to calculate the cluster centers after iteration based on the rating vectors of all users and the initial membership degree corresponding to each cluster using preset parameters. The second calculation module is used to calculate the distance from each user rating vector to each cluster center based on the iterated cluster centers, and to calculate the membership degree after iteration based on the distance relationship and preset parameters. The judgment module is used to calculate the difference between the membership degree after the iteration and the membership degree after the previous iteration, judge the iteration based on the difference value and the iteration termination threshold, and execute the iteration to continue or terminate according to the judgment result, and output the final membership degree matrix.

8. A digital marketing recommendation system based on big data according to claim 6, characterized in that, The result calculation module includes an attribution diagnosis module, wherein: The attribution diagnosis module is used to execute steps 1 to 4, and the process is as follows: Step 1: Select multiple users within the target area in the geographic information, and extract the membership degree of the users in the area to the target cluster center from the final membership matrix; the target cluster center is any one of multiple cluster centers, and the target area is any location in the geographic information. Step 2: Sort the membership degrees in ascending order to obtain a membership degree sequence, and select a preset proportion of membership degrees from the membership degree sequence as the membership degree threshold; Step 3: Compare the membership degree of the users in the target area in Step 1 with the membership degree threshold. If the membership degree of a user is greater than the membership degree threshold, then the user belongs to the target cluster center. If the membership degree of a user is not greater than the membership degree threshold, then the user does not belong to the target cluster center. Step 4: For each cluster center, perform steps 1 to 3 sequentially to obtain the clustering results for all users.

9. A digital marketing recommendation system based on big data according to claim 6, characterized in that, The correction module includes: The third calculation module is used to calculate the corrected score from the original score of the user belonging to the target cluster center and the public score of the product using preset correction parameters. The rating setting module is used to calculate preference judgment indicators based on the user's original rating and the product's public rating, and to set preference levels based on the preference judgment indicators; The module is used to count the number of users for each preference level and calculate the preference probability for each preference level. Based on the corrected score, preference level, and preference probability, the final score matrix is ​​constructed.

10. A digital marketing recommendation system based on big data according to claim 6, characterized in that, The upload module includes the working principle of the product recommendation model: The input data is acquired and fed into the fully connected layer to output the first target feature. The first target feature is fed into the CNN module to obtain the second target feature, and activated by the ReLU function to obtain the third target feature. The second target feature is fed into two GNN modules in sequence to obtain the fourth target feature. The first, third and fourth target features are concatenated and fused to obtain the fifth target feature. The fifth target feature is processed by the LDA module to obtain the sixth target feature. The sixth target feature is then input into the channel attention module to obtain the first channel feature and the second channel feature. The first channel feature and the second channel feature are then fused with the sixth target feature and input into the spatial attention module to obtain the third channel feature and the fourth channel feature, respectively. The features of channel 1 and channel 3 are fused to obtain the features of channel 5, and the features of channel 2 and channel 4 are fused to obtain the features of channel 6. The features of channel 5 and channel 6 are concatenated and then input into the LSA module and the fully connected layer in sequence. After activation by the Sigmoid function, the recommendation result is output.

Images