Marketing operation method based on big data analysis

Abstract

The invention relates to the technical field of marketing operation based on big data analysis, and discloses a marketing operation method based on big data analysis. Preprocessing the user-commodity interaction data to construct an initial interaction matrix; constructing a co-occurrence counting matrix; automatically calculating a sparse penalty coefficient and a convergence threshold; setting an optimization objective function and a soft threshold function; initializing a projection matrix and iteratively solving a sparse projection matrix through a coordinate descent algorithm; constructing a commodity sparse co-occurrence adjacency set; and generating a personalized recommendation result in combination with user historical behaviors. The analysis quality is improved through data preprocessing and denoising, commodity potential correlation is mined through co-occurrence modeling, adaptability is enhanced through automatic parameter adjustment, a projection matrix optimization result is sparse and interpretable, the structure sensing capacity of recommendation is improved through an adjacent structure, and finally high-precision and high-correlation recommendation is achieved. The structure definition, the calculation efficiency and the model generalization ability are excellent, and an efficient and stable personalized marketing recommendation system is constructed.

Description

Technical Field

[0001] The present invention relates to the field of marketing operations technology based on big data analysis, and specifically to a marketing operations method based on big data analysis. Background Art

[0002] With the rapid adoption of the internet and mobile devices, e-commerce and content platforms have accumulated massive amounts of user-product interaction data. Extracting high-value correlated information from this data for precision marketing and operations has become a hot topic in industry research. Existing techniques primarily include content-based recommendation, collaborative filtering-based recommendation, and model decomposition-based recommendation. Content-based recommendation analyzes product attributes to match user profiles, but struggles to maintain high accuracy for products with incomplete or difficult-to-normalize attributes. Memory-based collaborative filtering-based methods focus on user or product similarity. While their implementation principles are intuitive, the overhead of calculating similarity matrices or neighborhoods increases exponentially with the scale of users and products, making them difficult to meet real-time requirements. Matrix decomposition methods based on model decomposition compress user-product interactions into a low-dimensional space using latent semantic factors. However, they typically rely on manually set hyperparameters (such as regularization coefficients and learning rates) and lack adaptive processing strategies for sparse data, making them prone to overfitting and poor convergence.

[0003] For the aforementioned recommendation algorithms, researchers have introduced L1 or L2 regularization to control model complexity. However, the penalty coefficient often requires repeated tuning through cross-validation or manual experience, lacking an automated calculation mechanism. Others have employed iterative optimization methods such as coordinate descent to solve sparse matrix decomposition, but these methods still suffer from limitations such as unclear thresholds and low iteration efficiency in initialization, convergence determination, and large-scale data processing. Furthermore, traditional neighborhood selection often uses fixed empirical values, such as taking the top K similar items, ignoring the items' own interaction popularity and data distribution differences, resulting in a mismatch between popular and unpopular items in recommendation performance. Furthermore, user behavior in real-world business scenarios is highly heterogeneous and dynamic, making a single static model difficult to timely reflect changes in user interests. Furthermore, in the context of large-scale data, the interaction matrix is ​​often highly sparse. Directly operating on the original matrix consumes significant storage and computing resources and is prone to noise interference, reducing the stability and interpretability of recommendation results. Although some studies have attempted to extract richer co-occurrence relationships based on graph signal processing or spectral decomposition, these methods introduce complex graph Laplacian matrix construction or eigenvalue decomposition operations, further increasing the difficulty of system implementation and computational cost.

[0004] To this end, this case proposes a marketing operations approach based on big data analysis. Leveraging user-product interaction data, sparse matrix modeling and projection algorithms accurately extract the underlying preference relationships behind user behavior, thereby building a personalized recommendation system. The entire process encompasses data cleaning and preprocessing, co-occurrence relationship modeling, automatic parameter tuning, sparsity optimization, and finally, recommendation generation. This creates a marketing operations mechanism that is adaptable and highly accurate for large-scale data scenarios. Summary of the Invention

[0005] The present invention provides a marketing operation method based on big data analysis, which helps solve the problems mentioned in the above background technology.

[0006] The present invention provides the following technical solution: a marketing operation method based on big data analysis, comprising: Preprocess the original user-product interaction data and construct the initial interaction matrix; Construct a co-occurrence count matrix based on the initial interaction matrix; Automatically calculate the penalty coefficient for sparsification and the threshold for judging iterative convergence; Set the optimization objective function and soft threshold function; Initialize the projection matrix; Iteratively solve the sparse projection matrix through the coordinate descent algorithm; Construct a sparse co-occurrence adjacency set of products based on the sparse projection matrix; Combine user historical interaction data with sparse neighbor sets to generate personalized recommendation results.

[0007] Optionally, preprocessing the original user-product interaction data to construct an initial interaction matrix specifically includes: Get the user-product interaction record set from the business log: ; Count the number of unique users: ; Count the number of unique products: ; in, is a collection of interaction records, Represents a user There is interaction; Index for users, ; is the total number of users, which are different users who have at least one interaction The number of For product index, ; is the total number of products, which are different products that satisfy at least one interaction The number of Constructing a binary interaction matrix , specifically: , ; in, is a binary matrix with the number of rows , the number of columns is ; is a binary matrix In the Rank Elements of the column; Calculate the interaction rate per user: ;in, For the The total number of interactions per user; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding user row, after updating, the total number of new users is still recorded as , the corresponding updated matrix is ​​recorded as ; Calculate the interaction rate of each product: ;in, For the The total number of interactions with each product; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding product column, after updating, the total number of new products is still recorded as , the corresponding updated matrix is ​​recorded as .

[0008] Optionally, constructing a co-occurrence count matrix based on the initial interaction matrix specifically includes: Initialize the co-occurrence count matrix as ;in, is the co-occurrence count matrix, with the number of rows and columns being ; is a set of non-negative integers; is the co-occurrence count matrix Rank Column elements; For each user , take its interactive product set ;in, For the The index collection of products that the user has interacted with; For all ordered pairs implement ; After the completion, the symmetric matrix is ​​output .

[0009] Optionally, the automatic calculation of the penalty coefficient for the sparsification process and the threshold for determining iterative convergence specifically includes: Calculate the sum of off-diagonal co-occurrences, specifically: , ;in, is the sum of all off-diagonal elements; is the total number of off-diagonal elements; Assume the sparse penalty coefficient is ;in, is the L1 sparse penalty coefficient; Compute the diagonal sum and mean: , ;in, is the sum of the diagonal elements, that is, the sum of the total number of interactions of all products themselves; is the average value of diagonal elements; Assume that the iterative convergence judgment threshold is .

[0010] Optionally, the setting of the optimization objective function and the soft threshold function specifically includes: Assume the objective function is: ;in, is the projection matrix to be determined; , is the square of the Frobenius norm; for The absolute value of is the projection matrix to be determined Rank Column elements; Assume the soft threshold function is: ;in, is the soft threshold function; is a symbolic function, if , then +1, if , then -1; is the threshold parameter.

[0011] Optionally, the initializing the projection matrix specifically includes: Initialize the projection matrix: Set the iteration count ,make ;in, For the The projection matrix obtained by round iteration is for The projection matrix at time .

[0012] Optionally, the iteratively solving the sparse projection matrix by a coordinate descent algorithm specifically includes: When satisfied , perform the following steps: S1. Calculate the residual accumulation: ;in, For fixed lines and column When , the cumulative error term except its own diagonal term; S2. Calculate the original correlation: ;in, is the residual correlation measure before updating; S3. Coordinate update: ; S4, iteration count increment: ; Repeat steps S1 to S4 until convergence.

[0013] Optionally, constructing a sparse co-occurrence adjacency set of products based on a sparse projection matrix specifically includes: For each product Calculate interaction: ;in, For commodities Total number of interactions with itself; Calculate the number of adaptive neighbors: ;in, For products The number of neighbors to retain; is the ceiling function; For each product Extract the set of non-zero projection column indices and press Sort in descending order; Take the front Items constitute a neighbor set ;in, For commodities The neighbor set of ,is used to recommend candidates.

[0014] Optionally, combining user historical interaction data with a sparse adjacency set to generate personalized recommendation results specifically includes: For each user Construct a history interaction set: ;in, For users Interacted product index set; Merge neighbor sets ;in, For users The candidate product set; For each candidate product Calculate the recommendation score, specifically: ;in, For users For candidate products The cumulative projection correlation of Assume the recommended length is ;in, The number of recommended products output to the user; For users The number of historical interactive products; according to Before descending selection Item, recommended to users .

[0015] The present invention has the following beneficial effects: 1. Structure the original business logs into a standardized binary interaction matrix, and use the median threshold to filter out low-frequency users and products. This method achieves an efficient transition from unstructured data to structured analysis, and automatically adapts to the data distribution characteristics through a "median denoising" mechanism. In existing recommendation systems, the "cold start" problem has always been a major obstacle to the accuracy and effectiveness of recommendations. Traditional methods often use fixed thresholds or manual rules, which are difficult to adapt to different data scenarios. In contrast, this solution dynamically filters invalid data through the median of interaction frequency, effectively improving the quality of the interaction matrix, reducing the impact of noise, and enhancing the stability of subsequent calculations and the credibility of recommendations.

[0016] 2. Convert the user-item interaction matrix into item-item co-occurrence relationships. A symmetric co-occurrence matrix is ​​constructed using the co-occurrence frequencies of all item pairs in user behavior, providing a quantitative basis for subsequent correlation mining. Conventional collaborative filtering typically only measures rating similarity or user interest similarity, making it difficult to characterize direct common connections between items. The co-occurrence count matrix introduced in this paper not only enhances the ability to model the structure between items but also ensures symmetry and interpretability through ordered pair processing, making it an important intermediate layer in the transition to graph-structured recommendation systems.

[0017] 3. By automatically generating a sparse penalty coefficient from the mean of off-diagonal elements and a convergence threshold from the mean of diagonal elements, this technology achieves "adaptive tuning" without the need for manual parameter setting. The algorithm exhibits excellent generalization and stability, particularly in the context of dynamically changing big data. Traditional machine learning models often require manual hyperparameter adjustment, which is time-consuming and labor-intensive, and results in unstable accuracy. In contrast, this technology uses statistics to drive adaptive adjustment strategies, taking into account both global trends and local changes. This truly implements a closed-loop "automatic modeling - automatic parameter adjustment - automatic convergence" approach, enhancing the system's intelligence and industrial applicability.

[0018] 4. We propose an objective function that combines the Frobenius norm and L1 regularization term to ensure solution sparsity while controlling reconstruction error. A companion soft threshold function is used for sparse filtering during coordinate updates, balancing data fidelity and result compression. Traditional recommendation systems often use L2 regularization or high-dimensional dense weight matrices, resulting in difficult-to-interpret recommendation models and poor generalization. This method draws on the principles of sparse coding to significantly compress model parameters while ensuring fitting accuracy. This improves the transparency of the recommendation path and facilitates subsequent strategic interpretation and intervention by the business team.

[0019] 5. This paper uses a coordinate descent method to optimize the projection matrix elements point by point, and introduces a residual term decomposition and soft threshold fusion processing strategy to ensure convergence and result stability. In high-dimensional sparse matrix scenarios, gradient descent-based optimization methods are susceptible to local minima and have high computational resource overhead. In contrast, the coordinate descent method, through dimension-by-dimensional updates and local error minimization, is more stable and has better interpretability and debuggability, making it an ideal solution path for industrial-grade recommendation systems.

[0020] 6. By adaptively setting the number of neighbors and dynamically adjusting the number of retained neighbors based on each product's interactivity, we ensure that popular products have a richer set of neighbors, while unpopular products maintain reasonable accuracy, avoiding overfitting or oversparseness. Compared to the traditional fixed K-nearest neighbor strategy, this method adaptively sets the neighbor set based on product interaction behavior, making the constructed product network more consistent with actual consumer behavior, with greater dynamic scalability and recommendation flexibility. It is particularly suitable for scenarios such as e-commerce and video platforms where the number of products and content is rapidly increasing.

[0021] 7. Based on the neighbor sets of multiple interest-based products, sparse weighted scores are integrated to generate candidate sets. This supports modeling multiple user interests, improving the diversity and relevance of recommendations. This is more flexible and accurate than single-interest clustering or single-path recommendation methods.

[0022] 8. Adopting a clustering strategy based on the seed product neighbor set, combining projection weights to calculate recommendation scores, and automatically determining recommendation length based on user interaction history, this approach balances the richness of recommended content with user acceptance. Traditional recommendation systems suffer from monotonous recommended content and low user satisfaction. This invention introduces a "candidate adjacency fusion" mechanism and an "adaptive recommendation length" mechanism into the recommendation generation process, improving recommendation diversity and matching, ensuring personalized recommendations are more aligned with actual user preferences and the platform's business objectives. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] Example, see Figure 1 , a marketing operation method based on big data analysis, including: Preprocess the original user-product interaction data and construct the initial interaction matrix; Construct a co-occurrence count matrix based on the initial interaction matrix; Automatically calculate the penalty coefficient for sparsification and the threshold for judging iterative convergence; Set the optimization objective function and soft threshold function; Initialize the projection matrix; Iteratively solve the sparse projection matrix through the coordinate descent algorithm; Construct a sparse co-occurrence adjacency set of products based on the sparse projection matrix; Combine user historical interaction data with sparse neighbor sets to generate personalized recommendation results.

[0026] By preprocessing raw interaction data, the system can filter out noise and abnormal behavior data, effectively improving the accuracy of subsequent analysis. By constructing a co-occurrence count matrix, it discovers potential commonalities between products, helping to capture implicit user preferences. By automatically calculating sparse parameters and convergence thresholds, the algorithm adapts to different data scenarios. Coordinate descent is combined with soft thresholding to generate a sparse, interpretable projection matrix. On this basis, product adjacency relationships are established, giving the recommendation system good structural perception capabilities. Finally, by integrating historical behavior with graph structure for personalized recommendations, it improves recommendation accuracy and commercial conversion rate. Compared with traditional methods, this method reduces modeling complexity, improves execution efficiency and model generalization capabilities, and is a well-structured and clear-cut intelligent marketing technology framework.

[0027] The preprocessing of the original user-product interaction data to construct the initial interaction matrix specifically includes: Get the user-product interaction record set from the business log: Extract “who viewed / bought what” in the original business log into a structured set of interaction pairs as the basis for subsequent analysis; Count the number of unique users: ; Determine the total number of users participating in the recommendation system for subsequent matrix dimension definition; Count the number of unique products: ; Determine the total amount of recommended products in the system for subsequent matrix dimension definition; in, is a collection of interaction records, Represents a user There is interaction; Index for users, ; is the total number of users, which are different users who have at least one interaction The number of For product index, ; is the total number of products, which are different products that satisfy at least one interaction The number of Constructing a binary interaction matrix , specifically: , ; Map the interaction pairs into matrix form to facilitate subsequent matrix operations; in, is a binary matrix with the number of rows , the number of columns is ; is a binary matrix In the Rank Elements of the column; Calculate the interaction rate per user: ;in, For the The total number of interactions per user; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding user row, after updating, the total number of new users is still recorded as , the corresponding updated matrix is ​​recorded as ; Remove "cold users" with too little interaction to ensure data quality and computing efficiency; Calculate the interaction rate of each product: ;in, For the The total number of interactions with each product; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding product column, after updating, the total number of new products is still recorded as , the corresponding updated matrix is ​​recorded as ; Remove “cold products” with too few clicks / purchases to reduce noise.

[0028] A method is proposed to extract structured data from user-item interaction logs and construct a binary interaction matrix, while simultaneously filtering out low-frequency users and items based on the median. Through log extraction and structured processing, unstructured raw business data is converted into a matrix representation, laying the foundation for subsequent machine learning modeling. The number of unique users and items is counted, and the matrix dimensions are set accordingly, ensuring scale consistency in data processing. The median filtering mechanism outperforms fixed threshold methods, more effectively removing "cold start" objects and reducing the impact of sparse matrix noise on modeling. It also ensures that the users and items retained in the data have a certain level of activity, which helps improve the convergence speed, stability, and prediction quality of model training. In practical applications, this process is simple and efficient, with good replicability and cross-platform applicability.

[0029] The constructing of the co-occurrence count matrix based on the initial interaction matrix specifically includes: Initialize the co-occurrence count matrix as ;in, is the co-occurrence count matrix, with the number of rows and columns being ; is a set of non-negative integers; is the co-occurrence count matrix Rank Column elements; create a blank co-occurrence count matrix; For each user , take its interactive product set ;in, For the The index collection of products that the user has interacted with; For all ordered pairs implement ; Count the number of times any two products are interacted with by the same user; After the completion, the symmetric matrix is ​​output ; Ensure that the count matrix is ​​bilaterally symmetric and meets the co-occurrence definition.

[0030] This method converts user-item interaction information into item-item co-occurrence relationships, generating a symmetric co-occurrence count matrix. By traversing the user-interacted item set and counting the co-occurrence counts of item pairs, it is possible to establish correlations between items along the user behavior dimension. This item-item co-occurrence relationship is more stable than user-user co-occurrence and is particularly suitable for scenarios where the number of users far outnumbers the number of items. This matrix construction method fully considers the balance between symmetry and data density, laying the foundation for constructing an item association graph. In recommendation systems, item co-occurrence is a key metric for measuring similarity and complementarity. It helps the system identify potentially related items, improves the relevance and diversity of recommended content, and significantly optimizes the personalization and commercial value of recommendations.

[0031] The automatic calculation of the penalty coefficient for the sparsification process and the threshold for judging iterative convergence specifically includes: Calculate the sum of off-diagonal co-occurrences, specifically: , ;in, is the sum of all off-diagonal elements; is the total number of off-diagonal elements; sums up the number of co-occurrences of all product pairs to measure the overall co-occurrence intensity; Assume the sparse penalty coefficient is ;in, It is the L1 sparse penalty coefficient; the penalty intensity is automatically determined based on the average co-occurrence frequency, without manual adjustment; Compute the diagonal sum and mean: , ;in, is the sum of the diagonal elements, that is, the sum of the total number of interactions of all products themselves; is the average value of diagonal elements; Assume that the iterative convergence judgment threshold is ; The iteration convergence standard is automatically set based on the product’s own interaction mean.

[0032] By calculating diagonal and off-diagonal statistics of the co-occurrence matrix, the penalty coefficient for sparsification and the convergence threshold are automatically generated. Traditional recommendation algorithms often rely on empirically determined sparsity parameters and convergence criteria, which can easily lead to overfitting or overconservatism. By analyzing the strength of both item-item and item-item interactions in the co-occurrence matrix and calculating the average for adaptive setting, this approach not only reduces parameter tuning costs but also improves the rationality of parameter settings. The introduction of the penalty coefficient controls model complexity, while the convergence threshold ensures the stringency of the iteration stopping criteria, thereby ensuring the model strikes an optimal balance between accuracy and efficiency. This self-tuning parameter mechanism is applicable to a wide range of data sizes and significantly enhances model robustness and deployment efficiency.

[0033] The setting of the optimization objective function and the soft threshold function specifically includes: Assume the objective function is: ;in, is the projection matrix to be determined; , is the square of the Frobenius norm; for The absolute value of is the projection matrix to be determined Rank Column elements; Find a balance between reconstruction error and sparsity to generate a projection matrix that can reconstruct the behavior while being concise enough; Assume the soft threshold function is: ;in, is the soft threshold function; is a symbolic function, if , then +1, if , then -1; is the threshold parameter; During the coordinate descent update, the projected elements are "denoised" and kept sparse.

[0034] A combined objective function including the Frobenius norm and the L1 norm was designed, and a corresponding soft threshold function was constructed for sparse optimization. Introducing the Frobenius norm in the optimization objective to minimize the reconstruction error and using L1 regularization to enhance the sparsity of the solution is a typical sparse modeling approach in machine learning. Furthermore, a soft threshold function was introduced to control the degree of "denoising" in each coordinate update, effectively improving the model's anti-interference ability. Compared with ordinary L2 regularization, this construction can significantly compress the model size, facilitating deployment in resource-constrained scenarios (such as mobile recommendation systems). The output model also has good interpretability, which helps to trace and optimize recommendation logic in commercial scenarios.

[0035] The initialization of the projection matrix specifically includes: Initialize the projection matrix: Set the iteration count ,make ;in, For the The projection matrix obtained by round iteration is for The projection matrix at the time of ; provides a zero matrix starting point for the iterative process.

[0036] A method for initializing the projection matrix is ​​proposed, starting with a zero matrix or the identity matrix and setting an iteration counter. Initialization plays a critical role in the iterative optimization process. Choosing a zero matrix as the starting point prevents biased results from initial guesses and facilitates the gradual evolution of sparsity during structural optimization. Setting an iteration counter ensures controllable progress during model training and facilitates subsequent convergence assessment and performance monitoring. This concise and efficient design facilitates code implementation and multi-environment deployment, making it a key building block for fast cold start and online recommendation.

[0037] The iterative solution of the sparse projection matrix by the coordinate descent algorithm specifically includes: When satisfied , execute the following steps to determine whether to continue iterating: S1. Calculate the residual accumulation: ;in, For fixed lines and column When , the cumulative error term except its own diagonal term; calculate the error term for the first Cumulative error of coordinates; S2. Calculate the original correlation: ;in, It is the residual correlation measure before updating; it reflects the degree of fit between the current coordinates and the data; S3. Coordinate update: ; Use soft thresholding to adjust renormalization to ensure sparseness and avoid division by zero; like For commodities If it has never been interacted with, it is set to zero to prevent division by zero errors; S4, iteration count increment: ; Prepare for the next round of updates until convergence is met; Repeat steps S1 to S4 until convergence.

[0038] The sparse projection matrix is ​​gradually solved based on the coordinate descent algorithm, including residual calculation, related quantity evaluation, coordinate update and iterative control. The coordinate descent algorithm is widely used in solving high-dimensional sparse problems due to its high computational efficiency and simple implementation. In the present invention, the update direction is finely controlled by the residual and related quantity, combined with the soft threshold adjustment mechanism, which can effectively avoid the local optimal point dilemma and achieve fine-grained sparse control. An error tolerance mechanism is introduced in the iteration, such as handling the zero division case, to make the algorithm more robust. Overall, this method achieves high-quality optimization results with a relatively low computational overhead, and is particularly suitable for deployment in real-time personalized recommendation systems.

[0039] The step of constructing a sparse co-occurrence adjacency set of commodities based on the sparse projection matrix specifically includes: For each product Calculate interaction: ;in, For commodities The total number of interactions with itself; obtain the frequency of each product's own appearance for adaptive neighbor number; Calculate the number of adaptive neighbors: ;in, For products The number of neighbors to retain; is a rounding-up function; how many neighbors should be retained for the product based on its popularity; For each product Extract the set of non-zero projection column indices and press Sort in descending order; filter out products There are candidate neighbors with connections; Take the front Items constitute a neighbor set ;in, For commodities The neighbor set of is used to recommend candidates; based on sparse projection, only the strongest edges to build high-quality adjacencies.

[0040] Based on the projection matrix results and the frequency of product interactions, a sparse adjacency set is constructed for each product, achieving personalized adjacency graph optimization. An adaptive adjacency quantity mechanism applies differentiated adjacency strategies to products with different interaction intensities, avoiding overload of adjacency information for popular products and insufficient information for unpopular products. In actual recommendations, this structure helps form a "sparse but high-quality" product graph, significantly improving recommendation accuracy and system response speed. Furthermore, the product adjacency graph can be used for subsequent graph embedding, community analysis, and cross-category marketing, providing key data support for scalable recommendation architectures.

[0041] The method combines user historical interaction data with sparse adjacent sets to generate personalized recommendation results, specifically including: For each user Construct a history interaction set: ;in, For users Interacted product index set; identify users Interest seeds; Merge neighbor sets ;in, For users A set of candidate products; summarize the neighbors of all seed products to form a candidate pool; For each candidate product Calculate the recommendation score, specifically: ;in, For users For candidate products The cumulative projection correlation of ; the contribution of multiple seeds to the same candidate is integrated to measure the recommendation priority; Assume the recommended length is ;in, The number of recommended products output to the user; For users The number of historical interactive products; the length of the recommendation list is adaptively determined based on user activity; according to Before descending selection Item, recommended to users ; Present the most relevant product list to the user.

[0042] This strategy aggregates neighboring items of items a user has interacted with to generate a pool of recommendation candidates, which are then scored based on a projection matrix and recommended to the user. By integrating multiple item neighbors to construct an extended user interest space and quantifying recommendation strength using projection-related metrics, it offers a recommendation strategy that balances accuracy and diversity. The introduction of a mechanism for adaptively setting recommendation length based on the number of user interactions helps improve the quality of recommendations for long-tail users while ensuring the commercial value of the recommendations. This simple and efficient strategy is widely applicable to personalized marketing systems in e-commerce, content distribution, video platforms, and other fields.

[0043] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0044] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A marketing operation method based on big data analysis, characterized in that: include: Preprocess the original user-product interaction data and construct the initial interaction matrix; Construct a co-occurrence count matrix based on the initial interaction matrix; Automatically calculate the penalty coefficient for sparsification and the threshold for judging iterative convergence; Set the optimization objective function and soft threshold function; Initialize the projection matrix; Iteratively solve the sparse projection matrix through the coordinate descent algorithm; Construct a sparse co-occurrence adjacency set of products based on the sparse projection matrix; Combine user historical interaction data with sparse neighbor sets to generate personalized recommendation results.

2. A marketing operation method based on big data analysis according to claim 1, characterized in that: The preprocessing of the original user-product interaction data to construct the initial interaction matrix specifically includes: Get the user-product interaction record set from the business log: ； Count the number of unique users: ; Count the number of unique products: ; in, is a collection of interaction records, Represents a user There is interaction; Index for users, ; is the total number of users, which are different users who have at least one interaction The number of For product index, ; is the total number of products, which are different products that satisfy at least one interaction The number of Constructing a binary interaction matrix , specifically: ， ； in, is a binary matrix with the number of rows , the number of columns is ; is a binary matrix In the Rank Elements of the column; Calculate the interaction rate per user: ;in, For the The total number of interactions per user; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding user row, after updating, the total number of new users is still recorded as , the corresponding updated matrix is ​​recorded as ; Calculate the interaction rate of each product: ;in, For the The total number of interactions with each product; For all After sorting, take the value in the middle and record it as the median ; reserve The corresponding product column, after updating, the total number of new products is still recorded as , the corresponding updated matrix is ​​recorded as .

3. The marketing operation method based on big data analysis according to claim 2, characterized in that: The constructing of the co-occurrence count matrix based on the initial interaction matrix specifically includes: Initialize the co-occurrence count matrix as ;in, is the co-occurrence count matrix, with the number of rows and columns being ; is a set of non-negative integers; is the co-occurrence count matrix Rank Column elements; For each user , take its interactive product set ;in, For the The index collection of products that the user has interacted with; For all ordered pairs implement ; After the completion, the symmetric matrix is ​​output .

4. The marketing operation method based on big data analysis according to claim 3, characterized in that: The automatic calculation of the penalty coefficient for the sparsification process and the threshold for judging iterative convergence specifically includes: Calculate the sum of off-diagonal co-occurrences, specifically: , ;in, is the sum of all off-diagonal elements; is the total number of off-diagonal elements; Assume the sparse penalty coefficient is ;in, is the L1 sparse penalty coefficient; Compute the diagonal sum and mean: , ;in, is the sum of the diagonal elements, that is, the sum of the total number of interactions of all products themselves; is the average value of diagonal elements; Assume that the iterative convergence judgment threshold is .

5. The marketing operation method based on big data analysis according to claim 4, characterized in that: The setting of the optimization objective function and the soft threshold function specifically includes: Assume the objective function is: ;in, is the projection matrix to be determined; , is the square of the Frobenius norm; for The absolute value of is the projection matrix to be determined Rank Column elements; Assume the soft threshold function is: ;in, is the soft threshold function; is a symbolic function, if , then +1, if , then -1; is the threshold parameter.

6. The marketing operation method based on big data analysis according to claim 5, characterized in that: The initialization of the projection matrix specifically includes: Initialize the projection matrix: Set the iteration count ,make ;in, For the The projection matrix obtained by round iteration is for The projection matrix at time .

7. The marketing operation method based on big data analysis according to claim 6, characterized in that: The iterative solution of the sparse projection matrix by the coordinate descent algorithm specifically includes: When satisfied , perform the following steps: S1. Calculate the residual accumulation: ;in, For fixed lines and column When , the cumulative error term except its own diagonal term; S2. Calculate the original correlation: ;in, is the residual correlation measure before updating; S3. Coordinate update: ； S4, iteration count increment: ; Repeat steps S1 to S4 until convergence.

8. The marketing operation method based on big data analysis according to claim 7, characterized in that: The step of constructing a sparse co-occurrence adjacency set of commodities based on the sparse projection matrix specifically includes: For each product Calculate interaction: ;in, For commodities Total number of interactions with itself; Calculate the number of adaptive neighbors: ;in, For products The number of neighbors to retain; is the ceiling function; For each product Extract the set of non-zero projection column indices and press Sort in descending order; Take the front Items constitute a neighbor set ;in, For commodities The neighbor set of ,is used to recommend candidates.

9. The marketing operation method based on big data analysis according to claim 8, characterized in that: The method combines user historical interaction data with sparse adjacent sets to generate personalized recommendation results, specifically including: For each user Construct a history interaction set: ;in, For users Interacted product index set; Merge neighbor sets ;in, For users The candidate product set; For each candidate product Calculate the recommendation score, specifically: ;in, For users For candidate products The cumulative projection correlation of Assume the recommended length is ;in, The number of recommended products output to the user; For users The number of historical interactive products; according to Before descending selection Item, recommended to users .

Images