Insurance advertisement putting method based on big data analysis

Abstract

The invention relates to the technical field of insurance advertisement putting based on big data analysis, and discloses an insurance advertisement putting method based on big data analysis. According to the method, a user, channel and theme ternary heterogeneous graph is constructed, a budget-driven high-priority proportion and a balance coefficient are set, a minimum cutting and scale balance composite cost function is accurately defined, an initial sub-graph is generated based on global association strength sorting, and the initial sub-graph is subjected to iterative optimization through greedy node flipping and random local exchange. And finally, extracting high-priority sub-graph nodes and hierarchically distributing channels and user budget. Through the process, the problems that a single model is difficult to give consideration to a multi-dimensional relation, signal fusion is one-sided, budget and value are difficult to balance, single-target optimization is easy to sink into local optimum, budget distribution is extensive and the like are solved, multi-target integrated optimization, efficient convergence, global quality guarantee and multi-level accurate budget landing are realized, and advertisement putting accuracy and ROI are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of insurance advertising technology based on big data analysis, specifically to a method for insurance advertising based on big data analysis. Background Technology

[0002] With the rapid popularization of the internet and mobile devices, the advertising methods in the insurance industry are shifting from traditional offline media such as television and newspapers to precise online targeting based on big data. This involves collecting users' basic attributes and historical interaction behavior, using algorithms such as K-means and hierarchical clustering to segment users, and then setting corresponding targeting strategies for different tag groups. While this method provides some guidance for building a tagging system, it often only combines one or a few dimensions and lacks the ability to integrate multi-source heterogeneous information. Mainstream advertising exchanges use real-time bidding mechanisms, using the historical click-through rate or conversion rate of user-channel pairs as the basis for bidding, and allocating budgets through programmatic bidding. Although this method can improve the efficiency of targeting at the bidding level, it ignores the differences in content themes and the availability of channel resources, resulting in an incomplete assessment of the value of the same user across different themes or channels. Some advertising systems borrow e-commerce recommendation technology, using collaborative filtering or deep learning CTR prediction models to predict the user-ad creative matching degree and recommend ads most likely to generate clicks or conversions based on the prediction results. While these methods have achieved some breakthroughs in ad matching accuracy, they remain limited to a binary relationship between users and ad creatives, lacking more detailed modeling at the channel and theme levels. To balance cost and effectiveness across different channels, existing methods often employ linear programming or heuristic algorithms to allocate budgets to each channel, which then handles the campaigns within each channel. Because budget allocation and target coverage are often disconnected, and there is a lack of in-depth analysis of the interaction between channels, themes, and users, the efficiency of budget allocation both between and within channels is unsatisfactory.

[0003] Modeling solely at the user-materials or user-channel level fails to integrate the three key elements of "users, channels, and themes" into a single graph structure for comprehensive analysis, making it difficult to discover the optimal ad mix. Most methods rely on single metrics such as CTR, CVR, or dwell time, neglecting the synergistic effects of multi-source information such as theme popularity timeliness, channel remaining inventory, and cross-relationship coverage, resulting in a one-sided definition of edge weights. Traditional segmentation or clustering algorithms often only consider minimizing segmentation costs or maximizing intra-cluster similarity, failing to incorporate sub-cluster size constraints and value maximization objectives into a unified optimization framework, thus hindering flexible control over ad scale and balance. Commonly used methods such as spectral segmentation and greedy clustering lack effective global escape mechanisms during iteration, easily getting trapped in local optima and neglecting the discovery of potentially high-value nodes. Existing budget management is mostly single-layer allocation, lacking refined secondary allocation strategies at both the channel and user levels, making it difficult to truly concentrate resources on the most promising user-channel pairs.

[0004] Therefore, this proposal aims to develop a big data analytics-based method for insurance advertising. This solution constructs a ternary heterogeneous graph of "user-channel-theme," defines multi-dimensional and normalizable edge weights of association strength, and aims for a combined cost of "minimum cut + scale balancing." It employs a dual optimization approach of greedy flipping and random swapping to ultimately extract a high-priority subgraph. Nodes in this subgraph represent the most worthwhile user, channel, and theme combinations to prioritize within a limited budget. Then, through a two-layer proportional allocation, the advertising budget is precisely allocated to each user-channel pair. Summary of the Invention

[0005] This invention provides a method for insurance advertising based on big data analysis, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a method for insurance advertising placement based on big data analysis, comprising: Construct a heterogeneous ternary graph and define nodes, wherein the heterogeneous ternary graph includes user nodes, channel nodes, and topic nodes; Calculate the association strength between nodes and assign weights to the edges in the graph based on the association strength; Set the proportion and balance coefficient of the high-priority subgraph to control the size and balance of high-priority nodes; A precise composite cost function is defined, which comprehensively considers the cost of cutting the subgraph and the cost of balancing the scale. Generate an initial subgraph candidate set based on node strength; The initial subgraph is iteratively optimized using a greedy node flipping optimization method. The optimized subgraph is then refined through local swapping. Extract high-priority nodes from the final optimized subgraph and allocate the campaign budget based on the set of high-priority nodes.

[0007] Optionally, the construction of the heterogeneous ternary graph and the definition of nodes specifically includes: Set the user node set as ;in, For the complete set of user nodes; For the first One user node; Index for user nodes; Total number of user nodes; Set the channel node set as ;in, A complete set of channel nodes; For the first One channel node; For channel node indexing; Total number of channel nodes; Set the topic node set as ;in, The complete set of thematic nodes; For the first Each theme node; Index for topic nodes; Total number of topic nodes; Construct the complete set of nodes as , ;in, Let be the complete set of all nodes in the graph; This represents the total number of nodes in the entire set. Set the edge set as ;in, It is a set of node pairs, including only three types of associations: user-channel. User-Topic and Channels - Theme .

[0008] Optionally, the calculation of the association strength between nodes and the assignment of weights to edges in the graph based on the association strength specifically includes: Set user channel side weights to ;in, For user nodes With channel nodes The strength of the correlation between them; For users In channels Number of exposures; For users In channels Number of clicks; For users In channels The number of conversions; and These are click and conversion weights, respectively. ; Set the user theme weight to ;in, For users On the topic The strength of the association; For users On the topic The total time spent on the relevant content, in seconds; For users Total cumulative time spent on all topics, in seconds; For all users' last comment on the topic The timestamp of the action; This is the current system timestamp; The time decay constant; and These are the weights for stay and timeliness, respectively. ; Set the channel theme edge weight as ;in, For channels On the topic The strength of the association; For channels Previous topic Historical exposure count; For channels Total historical exposure for all topics; For channels The remaining available exposure for this period; For channels Total exposure quota for this period; and These are coverage and inventory weight, respectively. .

[0009] Optionally, setting the proportion and balance coefficient of the high-priority subgraph specifically includes: Set the proportion of high-priority nodes to ;in, The proportion of high-priority subgraphs to all nodes; The target proportion coefficient is set in advance by the budget; It is a rounding function; Set the balance regularization coefficient to ;in, This is a penalty coefficient for balancing scale. The scaling factor set by the user; It is the maximum value among all edge weights.

[0010] Optionally, the precise definition of the composite cost function specifically includes: Define the cost of subgraph cutting as ;in, , is the set of high-priority subgraph nodes to be determined; For the complete series Remove The remaining set of nodes after; For nodes With nodes The boundary weight between; Define the cost of subgraph size balancing as ;in, for The number of nodes; The ideal number of high-priority nodes; The total cost is .

[0011] Optionally, the step of generating an initial subgraph candidate set based on node strength specifically includes: The strength of each node is calculated as follows: ;in, For nodes The sum of global correlation strengths; For nodes With nodes The boundary weights between them; Sort by intensity in descending order and select the top... The nodes form the initial subgraph, specifically: .

[0012] Optionally, the iterative optimization of the initial subgraph using a greedy node flipping optimization method specifically includes: S1, For each node The calculation of the flipping incremental cost is as follows: ;in, To flip the node Change in total cost afterwards; For the first Subgraph after the next iteration; If Remove from the list; otherwise, add to the list. S2. If there is an incremental cost for node flipping. Then execute ;like Then keep ; Repeat steps S1 and S2 until all are... All .

[0013] Optionally, the local exchange refinement of the optimized subgraph specifically includes: S3, in the current situation and Randomly select one pair from each ,in ; S4. Calculate the incremental cost of the exchange, specifically: ;in, To remove simultaneously ,join in The change in cost afterward; S5, if Then at the same time Remove and The move-in, specifically: Otherwise, remain unchanged; Repeat steps S3 to S5 This is to escape the local optimum.

[0014] Optionally, the step of extracting high-priority nodes from the final optimized subgraph and allocating the delivery budget based on the set of high-priority nodes specifically includes: Extract the final high-priority subgraph as follows ;in, This is the final set of high-priority nodes; Separate subsets, specifically: high-priority user set High-priority channel collection High-priority topic set ; Channel-level budget allocation is as follows: ;in, To be allocated to channels Budget; This is the total advertising budget for this period; User channels are redistributed in a secondary manner, specifically as follows: ;in, For final allocation to users In channels Budget; Will As a user In channels The advertising budget.

[0015] The present invention has the following beneficial effects: 1. Constructing a ternary heterogeneous graph instead of the common binary model, the user-item model enables the interconnected analysis of users, channels, and themes. This enhances the model's expressive power, allowing for complex scenarios where a user is most interested in a particular theme across different channels.

[0016] 2. Three weighted metrics integrate multi-source signals such as click-through rate (CTR), conversion rate (CTR), dwell time, behavioral timeliness, historical coverage, and remaining inventory. Each weight is normalized to [0,1] to eliminate differences in the dimensions of each metric and ensure that all related signals can be weighted and integrated. Existing bidding models mostly focus on CTR or CVR, lacking dynamic consideration of topic newness and remaining inventory. This solution takes into account multi-dimensional information and performs time decay processing.

[0017] 3. Set the proportion of high-priority nodes With balance regularization coefficient The coupling mechanism supports the injection of business strategies on the one hand, and standardizes the deviation of subgraph size on the other. It integrates "budget coverage ratio" and "subgraph size constraints" into the same calculation framework in one step, avoiding post-hoc parameter tuning. Budget allocation is often disconnected between market strategy and technical optimization; this solution couples the two to ensure a balance between scale and value.

[0018] 4. Balance the principles of "minimizing the severing of high-value associations" and "minimizing subgroup size bias" through a balanced regularization coefficient. Seamless integration for a single objective. This allows the algorithm to optimize two key metrics simultaneously, avoiding the pursuit of the densest subgroups while ignoring scale, or the pursuit of scale accuracy while sacrificing value.

[0019] 5. Using the "global association strength" of nodes as the initial seed, this approach considers the overall association degree of each node with the entire set. This improves the quality of the starting point for optimization, accelerates convergence, and avoids suboptimal solutions caused by random or uniform initialization. Traditional graph partitioning often uses random or uniform initialization; this scheme selects the optimal initialization based on objective strength indicators.

[0020] 6. Calculate the "flipping incremental cost" for each node. "And it is only executed when the total cost can be reduced, forming a pure locally optimal descent chain. This ensures that each modification takes a step closer to the overall goal, and the optimization process is transparent and controllable."

[0021] 7. Introduce "random" in fixed rounds "Mechanism, through comparison" The method for determining whether a transaction is beneficial combines exploratory and convergent approaches. It effectively escapes local optima, increasing the probability of reaching the global optimum, while ensuring that swaps only occur when they are beneficial, thus reducing unnecessary computations.

[0022] 8. Two-tier allocation: First, allocate the budget to channels based on the overall "user-channel" correlation strength, then further subdivide it to users based on the same correlation strength, achieving tiered refinement. This ensures that the overall budget is maximized and allocated to the most valuable channels, while avoiding "internal waste" caused by even allocation within channels. Common bidding platforms can only control at a single dimension level; this solution performs multi-dimensional allocation at the budget level, making it more suitable for large-scale budget management and strategy execution. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process 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. 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.

[0025] Example, refer to Figure 1 A method for insurance advertising based on big data analysis, comprising: Construct a heterogeneous ternary graph and define nodes, wherein the heterogeneous ternary graph includes user nodes, channel nodes, and topic nodes; Calculate the association strength between nodes and assign weights to the edges in the graph based on the association strength; Set the proportion and balance coefficient of the high-priority subgraph to control the size and balance of high-priority nodes; A precise composite cost function is defined, which comprehensively considers the cost of cutting the subgraph and the cost of balancing the scale. Generate an initial subgraph candidate set based on node strength; The initial subgraph is iteratively optimized using a greedy node flipping optimization method to reduce the total cost; The optimized subgraph is refined by local swapping in order to escape local optima; Extract high-priority nodes from the final optimized subgraph and allocate the campaign budget based on the set of high-priority nodes.

[0026] By constructing a heterogeneous ternary graph and defining user, channel, and theme nodes, this approach addresses the challenge of traditional single user-advertisement or user-channel models simultaneously considering the multidimensional relationships between content and channels. It achieves a comprehensive portrayal of "who is most interested in which topics, where," within a single framework, laying a solid foundation for subsequent accurate correlation calculations. By calculating the correlation strength of user-channel, user-theme, and channel-theme edges and assigning edge weights, it overcomes the problem that relying solely on click-through rate or dwell time cannot reflect the comprehensive value of multi-source signals. This achieves a unified and integrated measurement of multidimensional indicators such as click-through rate, conversion rate, interest depth, timeliness, and inventory availability, significantly improving the accuracy of correlation assessment. By setting a high-priority subgraph ratio and balancing coefficient, it solves the problem of often balancing budget coverage and campaign value. This allows for controllable and precise limitation of the campaign node scale within the target range, maintaining a balance between subgraph size and value, thus avoiding both excessive budget dispersion and excessive resource concentration. By precisely defining a composite cost function that includes both minimum cut cost and size balancing cost, this approach addresses the limitations of traditional minimum cut or clustering algorithms, which can only optimize for a single objective and cannot simultaneously consider subgraph size constraints. It achieves an automatic balance between "preserving high-value associations" and "maintaining subgraph size" under a unified objective, realizing multi-objective integrated optimization. The step of generating an initial subgraph candidate set by sorting nodes according to their global association strength solves the problem of slow convergence and unstable quality caused by random or uniform initialization. It achieves the effect of using the most promising node combination as a starting point, accelerating the optimization process, and improving the quality of the initial solution, effectively shortening the overall computation time. The step of iteratively optimizing the initial subgraph using a greedy node flipping optimization method solves the problems of lacking a controllable iteration direction and difficulty in quantifying single-step improvements. It achieves the effect of accurately evaluating the total cost increment for each node flip and executing only when reducing costs, thus ensuring a stable optimization process and a clear convergence path. By refining the optimization results through a fixed number of random local swap operations, this approach addresses the issues of greedy optimization being prone to local optima and lacking a global escape mechanism. It proactively attempts better node combinations and escapes local optima at controllable swap times, further improving the global quality and stability of the final subgraph. By extracting high-priority subgraph nodes and allocating channel and user budgets hierarchically, it solves the problem of low resource utilization efficiency caused by traditional one-size-fits-all or single-layer budget allocation. This achieves the effect of first tilting channel budgets according to the overall "user-channel" correlation strength, and then finely allocating them to users according to the internal correlation strength of channels. This results in multi-dimensional and precise budget implementation and a significant improvement in ROI.

[0027] The construction of the heterogeneous ternary graph and the definition of nodes specifically include: Set the user node set as ;in, This is the complete set of user nodes, used to represent the set of all users to be targeted. For the first One user node; Index for user nodes; Determine the total number of user nodes; identify the target user range for ad placement, and establish basic nodes for subsequent graph modeling. Set the channel node set as ;in, This represents the complete set of channel nodes, indicating all available advertising channels, such as apps, websites, and social media platforms. For the first One channel node; For channel node indexing; This represents the total number of channel nodes; all distribution channels must be clearly identified for subsequent calculations of user-channel relationships. Set the topic node set as ;in, This represents the complete set of topic nodes, indicating the collection of topic categories involved in insurance advertisements, such as health insurance, car insurance, etc. For the first Each theme node; Index for topic nodes; The total number of topic nodes; lists the topic categories involved in the advertisement, providing nodes for user-topic association; Construct the complete set of nodes as , ;in, Let be the complete set of all nodes in the graph; This represents the total number of nodes in the entire set; all node types are merged to form the entire set of the analysis object. Set the edge set as ;in, It is a set of node pairs, including only three types of associations: user-channel. User-Topic and Channels - Theme Clearly define the edge types that require weight calculation in the graph model to prepare for assigning association strength values.

[0028] By setting up user node sets as base nodes, the problem of accurately defining the target user scope in traditional advertising models is solved. This achieves the effect of clearly defining all users to be targeted within the same analytical framework and ensuring that no potential customers are missed. By setting up channel node sets for advertising channel modeling, the problem of lacking unified management for multi-channel advertising and difficulty in simultaneously measuring the value of heterogeneous channels such as apps, websites, and social media platforms is solved. This achieves the effect of consistent and unified modeling of resources across channels and provides an accurate foundation for subsequent correlation calculations. By setting up topic node sets for advertising topic classification, the problem of heterogeneous content and lack of topic-level differentiation is solved. This achieves the effect of fine-grained modeling of different topics such as health insurance, car insurance, and pension insurance involved in insurance advertising and clarifying the source of users' interest in specific topics. By constructing a complete set of nodes for complete node fusion, the problem of traditional separate models operating in isolation and being difficult to cross-use user, channel, and topic information is solved. This achieves the effect of unifying the three types of nodes into a single graph structure and realizing integrated analysis of multi-dimensional information. By setting the edge set as a step in defining relationships, the problem of not being able to fully capture the interactions of user-channel, user-topic, and channel-topic relationships is solved. This achieves the effect of clearly defining all edge types to be calculated in the graph model, laying a clear boundary for subsequent multi-dimensional association strength quantification. Through the precise definition of nodes and edges, the problem of various delivery elements being scattered and managed in different models, making it difficult to comprehensively optimize them in a single algorithm, is solved. This achieves the effect of constructing a heterogeneous ternary graph data structure that can be directly processed by the algorithm, providing high-quality input for subsequent composite cost optimization and subgraph cutting. Through the entire process of constructing the heterogeneous ternary graph, the problem of limited delivery decision-making perspective caused by relying only on single or binary relationships is solved, achieving the effect of deeper and more accurate insights and multi-dimensional delivery strategy formulation in the "user-channel-topic" ternary relationship dimension.

[0029] This step-by-step modeling process solves the problem of effectively integrating and uniformly optimizing large-scale heterogeneous data in precision targeting scenarios. It provides a standardized and scalable graph data foundation for subsequent weight assignment, subgraph extraction, and budget allocation, significantly improving advertising efficiency and ROI.

[0030] The calculation of the association strength between nodes and the assignment of weights to edges in the graph based on the association strength specifically includes: Set user channel side weights to ;in, For user nodes With channel nodes The strength of the correlation between them; For users In channels Number of exposures; For users In channels Number of clicks; For users In channels The number of conversions, such as purchases; and These are click and conversion weights, respectively. ; The larger the edge weight, the more it is biased towards the click-through rate signal. Users and channels with high click-through rates are more likely to be selected into high-priority sub-images, which are suitable for driving traffic demonstrations or brand exposure. The larger the margin, the more focused the weighting becomes on conversion rate, making it easier to prioritize high-quality user conversions and channels, which is suitable for refined investment and maximizing ROI; by combining click-through rate and conversion rate, the value of user investment on channels can be quantified; Set the user theme weight to ;in, For users On the topic The strength of the association; For users On the topic The total time spent on the relevant content, in seconds; For users Total cumulative time spent on all topics, in seconds; For all users' last comment on the topic The timestamp of the action; This is the current system timestamp; This is a time decay constant, controlling the influence of the most recent behavior. It should match the rhythm of the theme's life cycle. If the theme's popularity declines rapidly, such as with short-term promotions, Take the larger value to make the difference between old and new behaviors significant; if the theme is valid year-round, such as basic life insurance, Smaller values ​​are preferable, as they result in a gradual decrease in speed. Increase the weight of recent behaviors, while old behaviors are almost ignored, to quickly capture users with high activity in the short term; Reduced, historical interest is retained for a long time, balancing long-term potential interest with recent activity, suitable for long-term and stable themes; and These are the weights for stay and timeliness, respectively. ; The larger the size, the longer the stay. It contributes more, emphasizes the concentration of interest, and is suitable for content-driven topics; The larger the value, the stronger the impact of the behavior's timeliness, emphasizing novelty, and suitable for trending topics or insurance activities with high timeliness; combine users' interest in the topic with the topic's popularity and freshness to measure the user-topic association; Set the channel theme edge weight as ;in, For channels On the topic The strength of the association; For channels Previous topic Historical exposure count; For channels Total historical exposure for all topics; For channels The remaining available exposure for this period; For channels Total exposure quota for this period; and These are coverage and inventory weight, respectively. If channel coverage (exposure potential) is significantly higher than inventory, then it needs to be increased. To prevent over-allocation; if there is a surplus of inventory, then increase... Prioritize the use of high-coverage channels; The larger the budget, the more emphasis is placed on historical coverage, allocating the budget to channels with broad coverage, which is suitable when you want to quickly expand your reach; The larger the channel, the more emphasis is placed on remaining inventory, prioritizing the use of high-inventory channels to prevent inventory waste, which is suitable for scenarios with critical inventory levels or platform resource conservation; it measures the channel's suitability for the theme and available inventory to guide budget allocation.

[0031] By setting user-channel weighting, the limitations of relying solely on a single CTR or CVR to evaluate user value are addressed. This achieves the effect of balancing high-click-through-rate users with high-conversion-value customers, allowing for flexible balance between "driving traffic" and "driving conversions" in ad placements. Similarly, setting user-theme weighting overcomes the biases of traditional methods that only consider dwell time or focus solely on recent behavior. This balances long-term user interest and recent activity, enabling the targeting of core users with sustained interest in a theme while quickly capturing potential customers with short-term activity, thus improving theme matching accuracy. Adjustable time decay constant δ addresses the issue of large differences in theme lifecycles and the inability to apply a one-size-fits-all approach. This allows for rapid weight decay for short-term promotional themes and smooth decay for enduringly effective themes, enabling the system to flexibly adjust user-theme association strength based on theme expiration, improving the timing of ad placements. By setting up channel-theme edge weights, the problem of wasted budget or limited reach due to neglecting channel remaining inventory and historical coverage is solved. This achieves the effect of prioritizing the expansion of global reach when exposure potential is sufficient and prioritizing the use of remaining resources when inventory is tight, thereby significantly improving channel resource utilization efficiency and overall campaign coverage.

[0032] By unifying and integrating the three types of edge weights, the problem of inconsistent dimensions of multi-dimensional indicators and the difficulty in direct comparison and weighting is solved. This achieves the effect of comprehensively evaluating multiple source signals such as clicks, conversions, dwell time, timeliness, and inventory under the same dimension, thereby providing high-quality data input for the next stage of sub-image segmentation and budget allocation, and improving the accuracy and ROI of advertising.

[0033] The setting of the proportion and balance coefficient of the high-priority subgraph specifically includes: Set the proportion of high-priority nodes to ;in, The proportion of high-priority subgraphs to all nodes; This is a target proportion coefficient pre-set in the budget; it can be appropriately increased when the overall budget is large. Expand high-priority coverage; when budget is limited, A smaller approach is preferable; if the goal is to deeply reach the core audience, Smaller, more concentrated deployments are possible; however, rapid expansion may be required. It can be on the larger side; Increasing the size of the subgraph results in more nodes and wider coverage, but dilutes the budget per node. The smaller the subgraph, the fewer the nodes and the higher the concentration, making it suitable for in-depth operation of core users, but it may give up potential long-tail value; This is the floor function; it determines the mapping relationship between the number of high-priority nodes and the size of the entire set. Set the balance regularization coefficient to ;in, This is a scale balancing penalty coefficient used to adjust for subgraph size deviations. The severity of the punishment; The scaling factor set by the user controls the weight of the balance term; As the size increases, the penalty for scale deviation becomes stronger, and the number of nodes in the subgraph is more strictly affected. limit; Reducing the cost of segmentation (related value) is the primary optimization approach, which can produce subgraphs that deviate from the scale objective but have higher correlation. Set the maximum value among all edge weights; configure the subgraph size to deviate from the penalty strength to ensure the subgraph size is close to the maximum. .

[0034] By setting a high-priority node ratio α, the problem of directly mapping budget size and campaign coverage at the algorithmic level is solved, achieving the effect of automatically calculating the number of high-priority subgraph nodes based on the overall budget. In previous methods, campaign teams often had to manually estimate the number of core users to be covered based on budget and experience, lacking a clear mathematical mapping. This is addressed by introducing a target ratio coefficient pre-set by the budget. and use the floor function to round up By precisely mapping to the number of nodes, this solution can automatically adjust the size of high-priority subgraphs when the budget changes, avoiding over- or under-coverage caused by purely empirical methods. This ensures that the deployment neither exceeds the budget nor misses important nodes. Its adjustable mechanism solves the problem of balancing two drastically different advertising strategies—"deep reach" and "rapid expansion"—under the same algorithm mode, achieving the effect of flexibly switching advertising targets in different business scenarios. When enterprises want to deeply engage with their core audience, they can... Setting it to a smaller value makes the sub-graph nodes more concise and the budget for each node more focused; however, when a brand or new product needs to quickly expand its exposure, it can be... This mechanism increases and expands node coverage to achieve large-scale reach. It breaks the limitations of traditional binary choices, allowing the model to switch freely between "few but precise" and "broad and fast" modes. By applying the "round up" function, it solves the problem of improper rounding of small values ​​when calculating the number of high-priority nodes, ensuring that at least the target proportion of the population is covered and the number of nodes is an integer. Since the number of nodes must be an integer, simple rounding or rounding may result in the actual coverage proportion being lower than expected. Using round up ensures that the calculated result is not lower than... This approach effectively avoids the risk of "discounted budget coverage ratio" and improves the controllability and security of budget usage. By setting a balanced regularization coefficient μ, it solves the problem that optimizing solely based on scale or value can easily lead to subgraph scale imbalance or damage to associated value, achieving a dynamic trade-off between subgraph scale and node associated value. Traditional cutting algorithms, if they only pursue minimizing edge weight cutting costs, often generate subgraphs that are too large or too small; if they only pursue strict scale constraints, they may sacrifice associated value. This solution uses μ as the weight for scale deviation penalty, allowing the algorithm to consider both "minimizing cutting costs" and "scale approaching α|V|" in each iteration, thereby obtaining a priority subgraph that is both efficient and meets strategic scale requirements. Through the adjustable design of the scaling factor λ in μ, it solves the problem of not being able to freely balance between "strict scale" and "maximum value" according to current business needs, achieving the effect of flexibly adjusting the penalty intensity according to real-time needs. In some stages, market pressures or competitive landscapes may necessitate stricter control over the scale of ad placement. In such cases, λ can be appropriately increased to allow the penalty for scale deviation to dominate. In other stages, to pursue higher ROI or tap into long-tail value, λ can be decreased to prioritize related value. This adjustable mechanism significantly enhances the model's adaptability to complex and ever-changing business scenarios.

[0035] The precise definition of the composite cost function specifically includes: Define the cost of subgraph cutting as ;in, , is the set of high-priority subgraph nodes to be determined; For the complete series Remove The remaining set of nodes after; For nodes With nodes The edge weights between them; the sum of the cut-off association strengths between the high-priority subgraph and its complement is used as the segmentation cost; Define the cost of subgraph size balancing as ;in, for The number of nodes; The ideal number of high-priority nodes; the ratio of the number of nodes in the subgraph to the target number. To minimize deviations and ensure that the subgraph size is reasonable; The total cost is Considering the combined costs of cutting and scaling deviations, the optimization objective is to minimize this function.

[0036] Define the cost of subgraph cutting This approach addresses the problem of traditional cutting algorithms focusing only on the internal similarity of subgraphs while neglecting the crucial breakage costs associated with peripheral nodes. It achieves a quantitative assessment of the contribution of each edge weight between a high-priority subgraph and its complement, enabling precise preservation of core value chains during optimization and preventing the loss of potentially high-value targets due to underestimated cutting costs. This is achieved by defining a cost-balanced approach based on subgraph size. The steps address the common problem in minimum cut or community detection methods of generating too many or too few nodes, making it difficult to meet budget and policy size constraints. This achieves the effect of quantifying the deviation between the number of nodes in the subgraph and the target ratio, thus ensuring that high-priority subgraphs are neither too large, leading to budget dilution, nor too small, limiting coverage. This is achieved by establishing a composite cost function. This approach addresses the issue that single-objective optimization cannot simultaneously balance "value maximization" and "scale constraints," achieving automatic balancing of the two core metrics within the same optimization iteration. This allows the algorithm to consider value preservation and scale rationality simultaneously without post-processing parameter tuning, significantly improving overall deployment efficiency. This is achieved by introducing an adjustable balancing coefficient into the composite cost function. This mechanism addresses the issue of varying emphasis on subgraph size and associated value across different business stages, achieving the effect of flexibly adjusting the penalty intensity based on real-time needs. This allows for increased penalty intensity when strict control of coverage is required. When prioritizing the extraction of long-tail value, the amount can be reduced. This allows for strategy switching across various advertising scenarios. By using the composite cost function as a clear objective for subsequent greedy flipping and local exchange optimization, it solves the problem that conventional heuristic algorithms are prone to getting stuck in local optima because they cannot quantify "what is a better subgraph" during iteration. This provides a clear basis for comparison for each node adjustment and achieves steady convergence, thereby significantly improving the quality of high-priority subgraphs and the accuracy and ROI of the final advertising budget allocation.

[0037] The generation of the initial subgraph candidate set based on node strength specifically includes: The strength of each node is calculated as follows: ;in, For nodes The sum of global correlation strengths; For nodes With nodes Calculate the edge weights between nodes; calculate the global association strength of each node for initial sorting; Sort by intensity in descending order and select the top... The nodes form the initial subgraph, specifically: Choose the one with the highest intensity. Each node is used as an initial high-priority subgraph to ensure that the starting point meets the size requirements.

[0038] By calculating the global association strength of each node and sorting them in descending order of strength, this approach solves the problem that previous random or uniform initialization methods could not guarantee the quality of the initial subgraph, achieving the effect of starting optimization with the most promising clusters of nodes. Using the sum of edge weights between a node and all other nodes to measure "global influence" ensures that the initial subgraph naturally clusters the most critical nodes in the user-channel-topic three-dimensional relationship, thus avoiding blindly traversing invalid nodes and significantly improving the effectiveness of subsequent iterations. This is achieved by "selecting the nodes with the highest strength..." The step of "forming an initial subgraph from nodes" solves the problem of how to map the budget coverage ratio to the initial point selection size, achieving precise control over the initial subgraph size. This method, combined with a pre-set scaling factor, The rounding function ensures that the initial subgraph neither exceeds the budget coverage nor falls below the minimum coverage requirement, providing a stable starting point that meets business needs for the entire optimization process. This initial subgraph generation strategy solves the problem of slow convergence and easy trapping in suboptimal solutions caused by large differences in initial solutions in traditional clustering or segmentation algorithms, significantly accelerating convergence and increasing the probability of a globally optimal solution. Intensity ranking prioritizes the inclusion of the most important nodes, effectively reducing unnecessary adjustments and allowing subsequent greedy flips and local swaps to focus on small-scale fine-tuning rather than large-scale node additions and deletions. By introducing a global relational perspective at the initial stage, the limitation that local neighbor information often fails to reflect node importance is overcome, improving the stability and robustness of the solution. The initial subgraph inherently contains the strongest global node set, effectively resisting the impact of subsequent parameter fine-tuning or noisy data, ensuring that key targets are always prioritized.

[0039] This step generates a high-quality initial subgraph, solving the problem of significant computational resource waste in the early optimization stage of large-scale heterogeneous ternary graphs, and achieving a significant improvement in the overall efficiency of the algorithm and the utilization rate of system resources. Compared with random initialization, this method only requires one global strength calculation and simple sorting to quickly identify priority candidates for deployment, saving a lot of unnecessary work for subsequent iterations and reducing the consumption of memory and computation cycles.

[0040] The iterative optimization of the initial subgraph using the greedy node flipping optimization method specifically includes: S1, For each node The calculation of the flipping incremental cost is as follows: ;in, To flip the node Change in total cost afterwards; For the first Subgraph after the next iteration; If Remove the node if necessary, otherwise add it; evaluate the node's condition. The impact of adding or removing a subgraph on the total cost; S2. If there is an incremental cost for node flipping. Then execute ;like Then keep ; Repeat steps S1 and S2 until all are... All By making local optimum judgments, the subgraph structure is continuously optimized to reduce the total cost.

[0041] By calculating the incremental cost of flipping each node and determining whether to flip based on the increment, this approach solves the problem that traditional heuristics or random adjustments cannot quantify the impact of single-step operations on the overall goal. It achieves the effect of providing clear cost feedback and decision-making basis for each node addition or removal, ensuring the optimization process always progresses in the direction of decreasing total cost. This is achieved by only calculating the incremental cost... The node flipping step is executed only when necessary, avoiding blind or ineffective node movements. This solves the problem of wasted computational resources caused by useless operations in large-scale graph segmentation algorithms, significantly reducing the number of invalid iterations and improving overall optimization efficiency. By repeatedly performing incremental evaluation and flipping judgment after traversing all nodes in each round, the problem of early stopping or random termination caused by single-step greedy strategies is solved. This achieves a clear convergence path and well-defined termination conditions, ensuring that the final subgraph automatically stops when the total cost can no longer be reduced, enhancing the stability and reliability of the optimization. By adjusting only a single node through a local flipping strategy, the high computational complexity and uncontrollable results caused by large-scale global redistribution are solved. This results in a simple algorithm implementation, easy parallelization, and scalability to heterogeneous graphs with millions of nodes, improving usability and practicality in big data environments. By using local optimum judgment based on incremental cost, the problem of not being able to guarantee that each adjustment approaches the global optimum is solved. This achieves a gradual reduction in total cost, stable monotonic convergence, and ultimately convergence to a high-quality local optimum, providing solid mathematical support and reliable decision-making basis for subsequent budget allocation.

[0042] The local exchange refinement of the optimized subgraph specifically includes: S3, in the current situation and Randomly select one pair from each ,in Introducing randomness helps the algorithm escape greedy local optima. S4. Calculate the incremental cost of the exchange, specifically: ;in, To remove simultaneously ,join in The change in cost after quantization; quantization and simultaneous removal And join Changes in total cost afterward; S5, if Then at the same time Remove and The move-in, specifically: Otherwise, leave it unchanged; only swap if it reduces the total cost, further refining the subgraph; Repeat steps S3 to S5 This is to escape the local optimum.

[0043] Through the current subgraph and Randomly select one pair from each The steps described above address the problem that a simple greedy flipping algorithm is prone to getting stuck in a fixed pattern and failing to discover a globally better subgraph. It achieves the effect of expanding the search space and introducing new candidate combinations through random perturbation, thus providing the necessary detection opportunities for the algorithm to escape local optima. This is accomplished by calculating the incremental cost of the swap. The steps described above solve the problem of not being able to quantify the impact of "simultaneously removing one node and adding another" on the total cost in large-scale graph cutting, achieving the effect of accurately assessing the cost changes brought about by bidirectional swaps, and ensuring that each swap operation has a clear mathematical basis. This is achieved by only... When executed Remove and The conditional swapping step addresses the issue of high costs and convergence difficulties introduced by random swaps, ensuring that each local swap truly reduces the total cost, thereby further refining the subgraph structure and steadily approaching the optimal solution. By repeatedly performing a fixed number of random swaps, the limitations of a single swap in terms of scope and the difficulty of completely escaping local optima are overcome. This achieves the effect of comprehensively evaluating the global structure through multiple rounds of trying different swap pairs, significantly increasing the probability of obtaining a globally approximate optimal subgraph. By introducing a hybrid optimization framework of local swaps after greedy node flipping, the limitations of either a single greedy or single random search are overcome. This achieves the advantage of both rapidly reducing the total cost through flipping and achieving global escape through swapping, thus balancing convergence speed and solution quality.

[0044] Through the coordinated iteration of this local swap operation and the flipping strategy, the problem that high-priority subgraphs in heterogeneous ternary graphs are prone to unstable fluctuations under multi-objective constraints is solved, achieving stable convergence to a high-quality subgraph and providing a reliable set of nodes for subsequent accurate budget allocation.

[0045] The step of extracting high-priority nodes from the final optimized subgraph and allocating the deployment budget based on the set of high-priority nodes specifically includes: Extract the final high-priority subgraph as follows ;in, To determine the final set of high-priority nodes; Separate subsets, specifically: high-priority user set High-priority channel collection High-priority topic set Separate high-priority nodes by type to facilitate subsequent budget allocation; Channel-level budget allocation is as follows: ;in, To be allocated to channels Budget; This is the total budget for this advertising campaign; the overall budget is allocated to each channel based on the user-channel association strength ratio. User channels are redistributed in a secondary manner, specifically as follows: ;in, For final allocation to users In channels The budget is allocated within each channel based on the strength of the connection between the user and the channel, thus achieving refined targeting. Will As a user In channels Calculate the advertising budget; convert the calculation results into the actual budget and implement the plan.

[0046] By extracting the final high-priority subgraph This step addresses the lack of connection between algorithm output and delivery targets, achieving the goal of clearly identifying the optimal user-channel-theme node set. This step solidifies the results of previous iterations and optimizations, ensuring the system can accurately identify the most worthwhile core nodes for delivery, providing an accurate and stable set of targets for subsequent budget allocation, and avoiding the risk of misjudgment caused by fuzzy or manual screening. This is achieved by separating the high-priority user set. High-priority channel collection and high priority topic sets This approach addresses the challenge of accurately allocating resources amidst the complex management of multi-dimensional nodes, achieving a streamlined budget allocation process by dividing responsibilities according to type. Clearly categorizing nodes into users, channels, and themes facilitates the division of labor among departments (e.g., channel procurement, content planning, user operations) while preventing confusion during budget allocation, ensuring close collaboration and clear division of labor across all stages. This is achieved through channel-level budget allocation. This process addresses the inefficiency of traditional average or single-dimensional budget allocation, achieving a tiered allocation of the overall budget based on the actual strength of channel relationships. This step precisely measures the coverage value of each channel to the core user group based on user-channel relationship signals, allowing high-value channels to receive more budget and significantly improving channel ROI. This is achieved through secondary allocation of user channels. This step addresses the problem of broad budget allocation within channels, which fails to consider user differences, and achieves a more refined allocation of resources within the same channel. This step further allocates channel budgets based on user contribution, ensuring that high-value users receive higher visibility and reach frequency within the same channel, achieving truly precise marketing. This is achieved through the calculation of... As a step in the actual budgeting process, this step resolves the issue of poor interface integration between theoretical model results and the execution system, achieving the effect of quickly implementing algorithm output into the advertising delivery system. This step intuitively maps budget figures to specific delivery parameters, such as average daily bid and maximum number of impressions, realizing seamless integration from algorithm to system execution and improving delivery efficiency and operability.

[0047] This complete budget allocation loop solves the common problems of disconnect, resource waste, and inefficiency between campaign strategy design and execution, achieving a closed-loop effect from accurate target identification and efficient budget allocation to actual campaign implementation. This closed-loop process ensures seamless collaboration across the entire chain, from optimization model output to campaign system input, with each step working in tandem to balance budget, value, and scale constraints across different dimensions, resulting in systematic, refined, and highly efficient insurance advertising campaigns.

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0049] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for insurance advertising placement based on big data analysis, characterized in that, include: Construct a heterogeneous ternary graph and define nodes, wherein the heterogeneous ternary graph includes user nodes, channel nodes, and topic nodes; Calculate the association strength between nodes and assign weights to the edges in the graph based on the association strength; Set the proportion and balance coefficient of the high-priority subgraph to control the size and balance of high-priority nodes; A precise composite cost function is defined, which comprehensively considers the cost of cutting the subgraph and the cost of balancing the scale. Generate an initial subgraph candidate set based on node strength; The initial subgraph is iteratively optimized using a greedy node flipping optimization method. The optimized subgraph is then refined through local swapping. Extract high-priority nodes from the final optimized subgraph and allocate the campaign budget based on the set of high-priority nodes.

2. The insurance advertising placement method based on big data analysis according to claim 1, characterized in that, The construction of the heterogeneous ternary graph and the definition of nodes specifically include: Set the user node set as ;in, For the complete set of user nodes; For the first One user node; Index for user nodes; This represents the total number of user nodes. Set the channel node set as ;in, A complete set of channel nodes; For the first One channel node; For channel node indexing; Total number of channel nodes; Set the topic node set as ;in, The complete set of thematic nodes; For the first Each theme node; Index for topic nodes; Total number of topic nodes; Construct the complete set of nodes as , ;in, Let be the complete set of all nodes in the graph; This represents the total number of nodes in the entire set. Set the edge set as ;in, It is a set of node pairs, including only three types of associations: user-channel. User-Topic and Channels - Theme .

3. The insurance advertising placement method based on big data analysis according to claim 2, characterized in that, The calculation of the association strength between nodes and the assignment of weights to edges in the graph based on the association strength specifically includes: Set user channel side weights to ;in, For user nodes With channel nodes The strength of the correlation between them; For users In channels Number of exposures; For users In channels Number of clicks; For users In channels The number of conversions; and These are click and conversion weights, respectively. ; Set the user theme weight to ;in, For users On the topic The strength of the association; For users On the topic The total time spent on the relevant content, in seconds; For users Total cumulative time spent on all topics, in seconds; For all users' last comment on the topic The timestamp of the action; This is the current system timestamp; The time decay constant; and These are the weights for stay and timeliness, respectively. ; Set the channel theme edge weight as ;in, For channels On the topic The strength of the association; For channels Previous topic Historical exposure count; For channels Total historical exposure for all topics; For channels The remaining available exposure for this period; For channels Total exposure quota for this period; and These are coverage and inventory weight, respectively. .

4. The insurance advertising placement method based on big data analysis according to claim 3, characterized in that, The setting of the proportion and balance coefficient of the high-priority subgraph specifically includes: Set the proportion of high-priority nodes to ;in, The proportion of high-priority subgraphs to all nodes; The target proportion coefficient is set in advance by the budget; It is a rounding function; Set the balance regularization coefficient to ;in, This is a penalty coefficient for balancing scale. The scaling factor set by the user; It is the maximum value among all edge weights.

5. The insurance advertising placement method based on big data analysis according to claim 4, characterized in that, The precise definition of the composite cost function specifically includes: Define the cost of subgraph cutting as ;in, , is the set of high-priority subgraph nodes to be determined; For the complete series Remove The remaining set of nodes after; For nodes With nodes The boundary weights between them; Define the cost of subgraph size balancing as ;in, for The number of nodes; The ideal number of high-priority nodes; The total cost is .

6. The insurance advertising placement method based on big data analysis according to claim 5, characterized in that, The generation of the initial subgraph candidate set based on node strength specifically includes: The strength of each node is calculated as follows: ;in, For nodes The sum of global correlation strengths; For nodes With nodes The boundary weights between them; Sort by intensity in descending order and select the top... The nodes form the initial subgraph, specifically: 。 7. The insurance advertising placement method based on big data analysis according to claim 6, characterized in that, The iterative optimization of the initial subgraph using the greedy node flipping optimization method specifically includes: S1, For each node The calculation of the flipping incremental cost is as follows: ;in, To flip the node Change in total cost afterwards; For the first Subgraph after the next iteration; If Remove from the list; otherwise, add to the list. S2. If there is an incremental cost for node flipping. Then execute ;like Then keep ; Repeat steps S1 and S2 until all are... All .

8. The insurance advertising placement method based on big data analysis according to claim 7, characterized in that, The local exchange refinement of the optimized subgraph specifically includes: S3, in the current situation and Randomly select one pair from each ,in ; S4. Calculate the incremental cost of the exchange, specifically: ;in, To remove simultaneously ,join in The change in cost afterward; S5, if Then at the same time Remove and The move-in, specifically: Otherwise, remain unchanged; Repeat steps S3 to S5 This is to escape the local optimum.

9. The insurance advertising placement method based on big data analysis according to claim 8, characterized in that, The step of extracting high-priority nodes from the final optimized subgraph and allocating the deployment budget based on the set of high-priority nodes specifically includes: Extract the final high-priority subgraph as follows ;in, This is the final set of high-priority nodes; Separate subsets, specifically: high-priority user set High-priority channel collection High-priority topic set ; Channel-level budget allocation is as follows: ;in, To be allocated to channels Budget; This is the total advertising budget for this period; The user channel is redistributed in a secondary manner, specifically as follows: ;in, For final allocation to users In channels Budget; Will As a user In channels The advertising budget.

Images