An advertisement putting optimization method and system based on an e-commerce platform and a storage medium

Abstract

The application provides an e-commerce platform-based advertisement optimization method, system and storage medium. The method comprises the following steps: collecting and fusing multi-source heterogeneous advertisement data and commodity data, standardizing and cleaning the multi-source heterogeneous advertisement data and commodity data, and performing dimension aggregation processing to obtain an advertisement data view; extracting a commodity operation feature vector based on the advertisement data view, performing stage discrimination on the commodity operation feature vector through a preset classification model to obtain a target operation stage of the commodity; matching a corresponding advertisement optimization strategy engine from a preset strategy library according to the target operation stage, generating an advertisement execution instruction containing a budget configuration, a bidding interval and a keyword matching rule through the advertisement optimization strategy engine; and performing dynamic bidding calculation and optimization processing on the keywords according to the advertisement execution instruction to obtain fine advertisement regulation parameters meeting the advertisement constraints and executing the advertisement. The application can efficiently optimize the advertisements of the e-commerce platform at low cost.

Description

Technical Field

[0001] This application relates to the field of e-commerce platform technology, and in particular to an advertising optimization method, system and storage medium based on an e-commerce platform. Background Technology

[0002] With the rapid development of e-commerce, advertising on e-commerce platforms has become a core means for sellers to acquire traffic and increase sales. Mainstream e-commerce platforms such as Taobao and JD.com provide sellers with advertising systems based on keyword bidding mechanisms. Sellers participate in the competition for platform traffic by setting parameters such as promotion plans, ad units, keyword bids, and budgets. However, in actual operation, with the increase in the number of products, the scaling up of ad units, and the dynamic changes in the market environment, the complexity of advertising management has increased dramatically, and existing advertising management technologies have gradually revealed significant shortcomings.

[0003] Currently, e-commerce advertising management mainly relies on the following technical solutions: first, data display systems based on report analysis; second, automated advertising systems based on rule engines; and third, static bidding and keyword management technologies. While these solutions have assisted advertising management to some extent, significant technical problems remain when facing increasingly complex e-commerce operation scenarios. Specifically, the data processing, stage judgment, strategy generation, and execution control in existing technical solutions are fragmented, lacking an integrated intelligent advertising mechanism that can organically integrate multi-source heterogeneous data processing, automatic identification of product operation stages, dynamic matching of advertising strategies, and refined keyword control. This makes it difficult to guarantee real-time data consistency during advertising, relies on human experience for determining operation stages, fails to dynamically switch advertising strategies with the product lifecycle, and lags behind changes in the real-time competitive environment in bid and keyword adjustments. Consequently, the overall automation level and decision-making accuracy of advertising cannot meet the needs of efficient and precise advertising. This problem is particularly prominent in scenarios with massive amounts of products and multiple promotional plans running concurrently, severely restricting the improvement of advertising effectiveness and the reduction of human operation and maintenance costs. Summary of the Invention

[0004] To achieve efficient and cost-effective advertising delivery on e-commerce platforms, this application provides an advertising delivery optimization method, system, and storage medium based on e-commerce platforms.

[0005] Firstly, a method for optimizing advertising on e-commerce platforms is provided, including: Collect and integrate multi-source heterogeneous advertising data and product data from e-commerce platforms, and perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view; Based on the advertising data view, product operation feature vectors are extracted, and the target operation stage of the product is obtained by performing stage discrimination on the product operation feature vectors through a preset classification model. According to the target operation stage, the corresponding advertising strategy engine is matched from the preset strategy library, and the advertising strategy engine generates an advertising execution instruction containing budget configuration, bid range and keyword matching rules. Based on the advertising execution instructions, the keywords are dynamically bid and automatically optimized to obtain refined advertising control parameters that meet the placement constraints, and then the advertising is executed.

[0006] In some embodiments, the standardization, cleaning, and dimensional aggregation processing of the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view includes: The multi-source heterogeneous advertising data and the multi-source heterogeneous fields in the product data are mapped to a set of standard fields with unified semantics through a preset standardized data dictionary; Abnormal data values ​​in the standard field set are identified and removed using preset anomaly detection rules. The mean of the source data in the preceding time window is used to fill in the removed and the original missing data values ​​to obtain cleaned standard data with high confidence. Account structure data, campaign performance data, and product dimension data are extracted from the cleaned standard data, and the campaign performance data and product dimension data are subjected to rolling aggregation calculations at a preset period to generate aggregated derivative indicators with preset business entities as anchors. The account structure data and the aggregated derived metrics are written into an in-memory database built on an online analytics engine to form an ad delivery data view that supports multi-dimensional real-time queries.

[0007] In some embodiments, the step of determining the target operational stage of a product by using a preset classification model to classify the product's operational feature vector includes: The product operation feature vector is input into a preset classification model constructed by a rule engine and a gradient boosting model; The rule engine performs preset threshold logic judgments on the number of days on the shelf, sales fluctuation coefficient and sales trend in the product operation feature vector to obtain preliminary stage classification results. Meanwhile, the confidence score of each candidate operation stage is obtained by nonlinear cross-calculation of the sales time series of the near preset number of days, conversion rate fluctuation coefficient, sales trend, price competitiveness and traffic characteristics in the product operation feature vector through the gradient boosting model. The preliminary stage classification results and the confidence scores of each candidate operation stage are weighted and fused to obtain a comprehensive confidence score set. The operation stage corresponding to the highest comprehensive confidence score in the comprehensive confidence score set is determined as the target operation stage of the product.

[0008] In some embodiments, the step of matching the corresponding advertising delivery strategy engine from a preset strategy library according to the target operational stage includes: Obtain the stage identifier corresponding to the target operation stage, and retrieve the strategy configuration template that is bound to the stage identifier from the preset strategy library; Extract the budget allocation rules, bid boundary rules, and keyword matching pattern rules defined in the strategy template; The dynamic bidding boundary is obtained by using the sales trend and traffic features in the product operation feature vector to perform interval constraint correction on the bidding boundary rule; The budget allocation rules, the keyword matching mode rules, and the dynamic bidding boundaries are filled into the strategy configuration template to generate an advertising delivery strategy engine that is adapted to the target operational stage.

[0009] In some embodiments, the step of using the sales trend and traffic features in the product operation feature vector to perform interval constraint correction on the bidding boundary rule to obtain a dynamic bidding boundary includes: The sales momentum coefficient is obtained by calculating the ratio between the sales trend and the preset sales trend benchmark value. The flow momentum coefficient is obtained by calculating the ratio between the flow characteristic and a preset flow characteristic benchmark value; The boundary adjustment factor is obtained by weighted fusion calculation of the sales momentum coefficient and the flow momentum coefficient; Obtain the upper and lower bid limits defined in the bid boundary rules, and use the boundary adjustment factor to proportionally scale the upper and lower bid limits to obtain the corrected dynamic upper and lower bid limits. The interval formed by the dynamic upper and lower bid limits is determined as the dynamic bid boundary.

[0010] In some embodiments, generating ad delivery execution instructions containing budget configuration, bid range, and keyword matching rules through the ad delivery strategy engine includes: Obtain the baseline budget allocation ratio and budget ceiling defined in the strategy configuration template, use the sales trend and the traffic trend to weight and adjust the baseline budget allocation ratio to obtain the dynamic budget allocation weight, and calculate the target budget configuration based on the dynamic budget allocation weight and the budget ceiling. A bid range matching the target budget configuration is constructed based on the dynamic bid upper limit and the dynamic bid lower limit; The keyword matching mode type is extracted from the strategy configuration template, and compared with the traffic characteristics and the preset matching mode switching threshold. The target keyword matching rule is obtained by dynamically selecting the keyword matching mode type according to the comparison result. The target budget configuration, the bid range, and the target keyword matching rules are uniformly encapsulated into a structured ad delivery execution instruction.

[0011] In some embodiments, the step of performing dynamic bidding calculations and automatic optimization on keywords according to the advertising execution instructions to obtain refined advertising control parameters that meet the placement constraints and then executing the advertising placement includes: Obtain the real-time performance characteristics of each keyword in the ad delivery data view; The real-time performance features are dynamically evaluated to obtain the real-time bid valuation for each keyword, and the bid range is used to perform boundary constraint processing on the real-time bid valuation to obtain dynamic bidding parameters. The matching efficiency of the traffic quality characteristics of each keyword and the matching rules of the target keyword are evaluated to obtain the matching mode adjustment instruction; The search term data in the advertising delivery data view is processed to identify invalid words, resulting in a set of keywords to be filtered. The dynamic bidding parameters, the matching mode adjustment instructions, and the set of keywords to be filtered are integrated and encapsulated to generate refined advertising control parameters, and advertising is executed according to the refined advertising control parameters.

[0012] In some embodiments, the process of executing the ad delivery further includes: The system links and traces the performance data after ad placement with historical optimization operations, and uses online learning algorithms to incrementally update the preset classification model and ad placement strategy engine to form a closed-loop iteration for ad placement optimization.

[0013] Secondly, an advertising optimization system based on an e-commerce platform is provided, the system comprising: The preprocessing module is used to collect and integrate multi-source heterogeneous advertising data and product data from e-commerce platforms, and to perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view. The discrimination module is used to extract product operation feature vectors based on the advertising data view, and to determine the target operation stage of the product by performing stage discrimination on the product operation feature vectors through a preset classification model. The matching module is used to match the corresponding advertising strategy engine from the preset strategy library according to the target operation stage, and generate an advertising execution instruction containing budget configuration, bid range and keyword matching rules through the advertising strategy engine; The optimization module is used to perform dynamic bidding calculation and automatic optimization of keywords based on the advertising execution instructions to obtain refined advertising control parameters that meet the placement constraints and then execute the advertising.

[0014] Thirdly, a computer-readable storage medium is provided, on which a computer program capable of running on a processor is stored, wherein when the computer program is executed by the processor, it implements an advertising placement optimization method based on an e-commerce platform as described in the first aspect.

[0015] Using the above method, this application collects and integrates multi-source heterogeneous advertising and product data from e-commerce platforms. The multi-source heterogeneous advertising and product data undergoes standardized cleaning and dimensional aggregation to obtain a unified standard advertising delivery data view. Based on the advertising delivery data view, product operation feature vectors are extracted. A preset classification model is used to determine the target operation stage of the product through stage discrimination. According to the target operation stage, a corresponding advertising delivery strategy engine is matched from a preset strategy library. The advertising delivery strategy engine generates advertising delivery execution instructions containing budget configuration, bid range, and keyword matching rules. Based on the advertising delivery execution instructions, keywords are dynamically bid on and automatically optimized to obtain refined advertising control parameters that meet the delivery constraints, and then the advertising is delivered. This enables efficient and low-cost advertising delivery on e-commerce platforms. Attached Figure Description

[0016] Figure 1 This is a flowchart of the advertising optimization method based on e-commerce platforms provided in this application.

[0017] Figure 2 This is a block diagram of an advertising optimization system based on an e-commerce platform provided in this application. Detailed Implementation

[0018] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but is consistent with the broadest scope claimed in this application.

[0019] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating the advertising optimization method based on an e-commerce platform provided in this application. Figure 1 As shown, the advertising optimization method based on e-commerce platforms includes the following steps: Step S100: Collect and integrate multi-source heterogeneous advertising data and product data from e-commerce platforms, and perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising placement data view.

[0021] This application describes the optimization approach. Addressing the shortcomings of existing technologies, such as fragmented data and inefficient manual collection, it achieves intelligent collection and fusion of multi-source heterogeneous advertising and product data through the following specific implementation methods. First, a secure connection is established with at least one e-commerce platform via an API gateway, introducing an automatic data collection mode evaluation and switching mechanism. The optimization approach pre-sets three modes: synchronous call, asynchronous call, and near real-time streaming collection based on message queues. Upon initiation of collection, the system monitors in real time and automatically evaluates and selects the optimal collection mode based on the platform's API call restrictions, the number of advertising units in the account to be collected, and the historical data update frequency. Specifically, for large-scale data scenarios with a massive number of advertising units, the asynchronous collection mode based on message queues is triggered. Each collection task is encapsulated as a message and sent to a distributed message queue, where clustered consumer services process them concurrently and return the collection results asynchronously. This achieves peak smoothing and ensures the stability and real-time performance of data collection. For scenarios with smaller data volumes and lower real-time requirements, the system automatically switches to synchronous call mode to directly acquire data, reducing complexity and resource overhead. This allows for the automatic and efficient aggregation and fusion of advertising performance data, product dimension data, and competitive environment data that were originally scattered across different interfaces, breaking down data silos and providing a real-time, unified data foundation for subsequent processing.

[0022] Completing the above operations is insufficient to meet the needs of precise targeting. The resulting multi-source heterogeneous advertising and product data exhibits diverse characteristics, with inconsistent field naming, data formats, granularity, and quality. Directly using this data for strategy decisions will interfere with accuracy. Therefore, preprocessing is necessary. The standardization, cleaning, and dimensional aggregation of the multi-source heterogeneous advertising and product data to obtain a unified standard advertising data view includes the following steps: Step S101: Map the heterogeneous fields in the multi-source advertising data and product data to a set of standard fields with unified semantics through a preset standardized data dictionary.

[0023] Step S102: Abnormal data values ​​in the standard field set are identified and removed using preset abnormal detection rules. The removed and original missing data values ​​are then filled in using the mean of the same source data in the preceding time window to obtain cleaned standard data with high confidence.

[0024] Step S103: Extract account structure data, campaign performance data, and product dimension data from the cleaned standard data, and perform rolling aggregation calculations on the campaign performance data and product dimension data at a preset period to generate aggregated derivative indicators anchored to preset business entities.

[0025] Step S104: Write the account structure data and aggregated derived metrics into an in-memory database built on an online analytics engine to form an ad delivery data view that supports multi-dimensional real-time queries.

[0026] Internally, a standardized data dictionary is built for optimization. When multi-source heterogeneous advertising and product data floods in from API interfaces of different e-commerce platforms, the data cleaning and standardization engine captures this data in real time. For each record of multi-source heterogeneous advertising and product data, the engine calls the mapping rule table corresponding to that source in the dictionary, mapping the field names there to internally unified semantic tags. For example, regardless of whether the field returned by the original interface is "cost," "exhaustion," or the English word "cost," it will be unified into the standard field "total_cost" through dictionary mapping. This eliminates semantic differences between different data sources and lays a solid foundation for subsequent unified processing.

[0027] After the standardization process in step S101, although the data achieves semantic consistency, it may still contain outliers and missing values ​​due to network latency, platform system jitter, or data transmission errors. Therefore, deep cleaning is necessary to ensure data quality, employing a strategy combining dynamic anomaly detection and context-aware completion. Specifically, multi-dimensional anomaly detection rules are preset. For example, it monitors whether the single-click cost of the key indicator `total_cost` exceeds a preset multiple of the average single-click cost of the ad unit in the past 24 hours. If it does, it is marked as dirty data and removed. Simultaneously, the temporal integrity of the data stream is checked to identify any missing data within any time window. For the removed outliers and existing missing values, context-aware completion logic is executed: it does not fill with the entire dataset or zero values, but rather precisely locates the data from the previous complete time window that is from the same source as the current data point (i.e., from the same ad unit or the same promotion plan), calculates the mean value of the corresponding standard field within that window as the imputation value, and writes it into the missing position. If real-time click data is missing for a particular minute, the average click data per minute for that keyword in the previous hour will be used to fill in the gaps. All completion operations will be marked with a completion tag in the additional labels of the data records for weighting consideration by downstream models. Finally, high-confidence standard data are removed to eliminate obvious noise and provide reasonable estimates for missing values, ensuring maximum data consistency and integrity.

[0028] Although the data has been cleaned and supplemented, it still exists in the form of detailed records and cannot directly support insights into macro strategies. Further upgrades are needed to transform this information into high-level semantic information with business guidance value. This involves extracting three core data types from the cleaned standard dataset: first, account structure data, which constructs a tree-like hierarchical relationship between accounts, promotion plans, ad units, keywords, and creatives for subsequent attribution analysis; second, campaign performance data, including impressions, clicks, spending, conversions, and add-to-cart / favorites; and third, product-level data, which, in addition to basic inventory and sales, actively calculates and records higher-level features such as sales trends and traffic characteristics. After extraction, a rolling aggregation engine is activated, using a preset short time period as a window to continuously and overlappingly aggregate the campaign performance data and product-level data. This aggregation is not based on a time dimension, but rather uses a user-defined preset business entity (such as a 618 promotion plan or a product group) within the account structure as an anchor point, performing real-time summation and derivation calculations on the scattered data from all levels and units under that entity. Using a single promotional campaign as an anchor, the system continuously calculates aggregated metrics every second, including total spending over the past 24 hours, overall average cost per click (CPC), total return on investment (ROI), and moving average click-through rate (CTR). This provides a real-time overview of the overall return on investment from different business perspectives, offering rich, context-rich features for subsequent strategy models.

[0029] The value of the data generated through the above steps ultimately needs to be realized through efficient and flexible consumption methods. An in-memory database based on an Online Analytical Processing (OLAP) engine is used as the final carrier for data services. The structured account structure data and various aggregated derived indicators generated above are written in batches and in real-time to a pre-built OLAP in-memory data table via a high-performance data bus. This table adopts a wide table design, organized in the form of dimension columns (such as time, account, promotion plan, product, keywords, etc.) and indicator columns (such as total, average, ROI, etc.). Thanks to the data pre-aggregation, columnar storage, and vectorized computing capabilities of the OLAP engine, as well as the extreme read and write speed of in-memory computing, the view can achieve sub-second or even sub-second response times for complex multi-dimensional query requests. The final presented advertising data view is an advertising data view that supports multi-dimensional real-time queries, high consistency, and low latency.

[0030] Step S200: Extract product operation feature vectors based on advertising data view, and use a preset classification model to determine the target operation stage of the product by classifying the product operation feature vectors into stages.

[0031] After obtaining an ad placement data view that supports multi-dimensional real-time queries, it is necessary to extract multi-dimensional features that characterize the product's operational status to support subsequent stage judgments. Specifically, when extracting the product operation feature vector based on the ad placement data view, the product identifier is used as an index, and the pre-aggregated derived indicators in the view are directly queried through the online analysis and processing engine to obtain the product's listing days, sales trend, and traffic characteristics. Among them, the sales trend is an indicator reflecting the changes in product sales, calculated from sales data within a preset period, and the traffic characteristics are data indicators that comprehensively reflect changes in product traffic and user behavior. At the same time, the scrolling aggregation capability of this view is used to pull the daily sales and conversion volume of the past 7 days at a daily granularity, constructing a sales time series and calculating the daily conversion rate. Then, the conversion rate fluctuation coefficient is obtained by using the ratio of the standard deviation of the daily conversion rate to the mean. Features such as listing days, sales time series, conversion rate fluctuation coefficient, sales trend, price competitiveness, and traffic characteristics are concatenated in a fixed order to form the product operation feature vector. By directly reusing the high-level indicators already processed in the view, and performing only lightweight derivative calculations on some time-series features, an accurate, multi-dimensional, and real-time consistent representation of commodity operation features can be obtained, enabling subsequent models to make reliable stage judgments based on inputs that reflect the current market and price status.

[0032] After obtaining the advertising data view, it is necessary to extract multi-dimensional information that can characterize the product's operational status to support accurate identification of the product's operational stage. This involves constructing a multimodal ensemble classifier that combines a rule engine and a gradient boosting model as the preset classification model. This model organically combines deterministic rules accumulated through human experience with the fitting ability of machine learning models to complex nonlinear relationships, achieving automated and highly reliable inference of the product's operational stage. The process of determining the target operational stage of a product by using the preset classification model to classify the product's operational feature vectors includes the following steps: Step S201: Input the product operation feature vector into the preset classification model constructed by the rule engine and gradient boosting model.

[0033] Step S202: The rule engine performs a preset threshold logic judgment on the number of days on the shelf, sales fluctuation coefficient and traffic trend slope in the product operation feature vector to obtain the preliminary stage classification result.

[0034] In step S203, the confidence scores of each candidate operation stage are obtained by performing nonlinear cross-calculation on the sales time series of the near preset number of days, conversion rate fluctuation coefficient, sales trend, price competitiveness and traffic characteristics in the product operation feature vector through the gradient boosting model.

[0035] Step S204: The preliminary stage classification results and the confidence scores of each candidate operation stage are weighted and fused to obtain a comprehensive confidence score set, and the operation stage corresponding to the highest comprehensive confidence score in the comprehensive confidence score set is determined as the target operation stage of the product.

[0036] The aforementioned product operation feature vectors serve as unified input information, simultaneously fed into two parallel processing paths of a pre-defined classification model: one is a rule engine encapsulating expert experience, and the other is a gradient boosting model trained offline. This dual-path design enables the model to make rapid judgments using clear business logic, while also capturing subtle trend changes from historical data.

[0037] Within the rules engine pathway, a series of deterministic threshold rules based on business expert experience are pre-defined. The engine extracts three pieces of information with clear stage indications from the input product operation feature vector: number of days on the shelf, sales fluctuation coefficient, and sales trend, and performs numerical comparisons on each one. Specifically, the engine determines: if the number of days on the shelf is less than 30 days, the sales fluctuation coefficient is greater than 0.8, and the sales trend is less than 0.2, the new product testing period is output; if the average daily sales over the past 7 days are greater than 1.5 times the category average and the sales trend shows continuous growth, the main promotion / bestseller period is output. Through threshold judgments combining multiple conditions with AND / OR logic, the rules engine transforms the original numerical features into a discrete, preliminary stage classification result with clear business meaning. This result serves as prior knowledge for subsequent fusion decisions.

[0038] While the rule engine executes its judgments, a gradient boosting model on another path performs in-depth processing of the more dynamic and complex information in the product operation feature vector. This model incorporates an ensemble structure composed of multiple decision trees, performing non-linear cross-calculation on the input features. The processing involves the model receiving the sales time series of the past 7 days, conversion rate fluctuation coefficient, sales trend, price competitiveness, and traffic characteristics from the feature vector. This information is fed into each decision tree, and each tree, based on its learned splitting conditions, performs a series of yes / no judgments on the input features. For example, if price competitiveness is in a high range and traffic characteristics show an upward trend, the sample is ultimately guided to a leaf node. Each leaf node pre-records a prediction score. The model accumulates the scores from the leaf nodes of all decision trees to obtain an original score vector for different candidate operation stages. Subsequently, a normalized exponential function (such as the Softmax function) is used to mathematically transform the original score vector, mapping the accumulated score to probability values ​​between 0 and 1, generating confidence scores for each candidate operation stage (such as new product testing period, main best-selling product period, and long-tail clearance period). This enables the model to automatically capture the combined effect of high sales trends and rising traffic characteristics on the peak promotion period of best-selling products, or the direction of long-tail clearance period when conversion rates fluctuate wildly and sales trends are extremely low, generating accurate confidence assessments that include non-linear information.

[0039] Finally, a decision fusion process is performed to unify the heterogeneous results generated by the two pathways. First, the discrete preliminary stage classification results output by the rule engine are one-hot encoded, transforming them into a sparse vector with the same dimension as the candidate stage. For example, if the result is a new product testing period, the corresponding position is 1, and the rest are 0. Next, this sparse vector is weighted and summed with the confidence score vector output by the gradient boosting model. The weight coefficients can be pre-set based on the historical performance of the two models under different product categories. For example, in categories with sufficient data, a higher weight is assigned to the gradient boosting model. This calculation yields a comprehensive confidence score set that integrates prior rules and complex pattern recognition capabilities. Finally, this score set is traversed, and the operational stage with the highest score is selected as the current target operational stage of the product. For ambiguous cases where all stage scores are below a preset threshold, a fallback strategy is implemented, classifying the product as a stable operational period to ensure a stable and executable classification result is output. In this way, the rule engine transforms clear boundary features into preliminary classification results with business semantics, providing strong interpretability and reliability; the gradient boosting model performs nonlinear cross-calculation on multi-dimensional dynamic trend information, transforming the original features into confidence scores that can accurately reflect the transition state of the stage; finally, through weighted fusion, it avoids misjudging boundary cases due to rigid thresholds, and also avoids generating absurd conclusions in unfamiliar scenarios due to purely data-driven approaches, thereby improving the accuracy, robustness, and automation level of product operation stage determination, and laying a solid foundation for the accurate matching of subsequent differentiated strategies.

[0040] Step S300: Match the corresponding advertising strategy engine from the preset strategy library according to the target operation stage, and generate an advertising execution instruction that includes budget configuration, bid range and keyword matching rules through the advertising strategy engine.

[0041] The stage-based judgment results are then transformed into executable advertising strategies. This involves building a strategy configuration template library and dynamically adjusting strategy parameters using real-time data from product operation feature vectors, achieving refined matching and adaptive adjustment of the strategy. The process of matching the corresponding advertising strategy engine from the preset strategy library based on the target operational stage includes the following steps: Step S301: Obtain the stage identifier corresponding to the target operation stage, and retrieve the strategy configuration template that is bound to the stage identifier from the preset strategy library.

[0042] Step S302: Extract the budget allocation rules, bid boundary rules, and keyword matching pattern rules defined in the strategy template.

[0043] Step S303: Use the sales trend and traffic features in the product operation feature vector to perform interval constraint correction on the bidding boundary rule to obtain the dynamic bidding boundary.

[0044] Step S304: Fill the budget allocation rules, keyword matching mode rules and dynamic bidding boundaries into the strategy configuration template to generate an advertising delivery strategy engine that is adapted to the target operational stage.

[0045] The aforementioned stage identifiers are the operational stage enumeration values ​​corresponding to the highest comprehensive confidence score output in step S204, such as the new product testing period, the main promotion of best-selling products period, or the long-tail clearance period. The preset strategy library is a key-value pair relational mapping table, where the keys are the stage identifiers for each operational stage, and the values ​​are pre-configured strategy configuration templates. The strategy configuration template is a parameterized strategy blueprint preset for different operational stages, internally defining the initial values ​​and adjustment rules for core parameters such as budget allocation preferences, upper and lower bound constraints on bids, and keyword matching mode types to be adopted in that stage. For example, if the target operational stage identifier is the new product testing period, the core logic defined in the retrieved strategy configuration template is low cost and fast verification, including an initial small budget, a bid range of 50% to 70% of the industry average bid, and the use of broad match for keywords. If the stage identifier is the main promotion of best-selling products period, the template corresponds to a large budget, traffic grabbing, and ranking control logic, including increasing the budget cap to the preset peak, switching high-conversion keywords to exact match, and setting aggressive bid ranges for high-value keywords. By using stage identifiers as anchors for template-based matching, operational strategy knowledge is structured and reusable, reducing the computational overhead of regenerating strategy logic for each decision.

[0046] The strategy configuration template employs a hierarchical structure to define various campaign parameters. Parsing the retrieved templates, three core rule categories are extracted: budget allocation rules define the baseline budget allocation ratio and budget cap; bid boundary rules define a fixed bid range consisting of the bid cap and bid floor; and keyword matching mode rules define the keyword matching mode types to be used, such as exact match, phrase match, and broad match. These rules constitute the basic parameter framework for the campaign strategy during this operational phase. For example, during the long-tail clearance phase, the extracted budget allocation rules limit the daily budget to the minimum required to maintain current sales volume; the bid boundary rules define the bid cap to strictly anchor the maximum cost per click within the target profit margin; and the keyword matching mode rules define that only the most accurate long-tail keywords are retained and generate a negative keyword filter list.

[0047] The bidding boundary rules extracted from the strategy configuration template are static empirical values, which cannot perceive real-time sales pressure and demand changes of the product. Therefore, real-time data is introduced to dynamically adjust the bidding range. The process of using sales trends and traffic features from the product operation feature vector to adjust the bidding boundary rules to obtain the dynamic bidding boundary includes the following steps: Step S303-1: Calculate the sales momentum coefficient by comparing the sales trend with the preset sales trend benchmark value.

[0048] Step S303-2: Calculate the flow momentum coefficient by comparing the flow characteristics with the preset flow characteristic benchmark value.

[0049] Step S303-3: The boundary adjustment factor is obtained by weighted fusion calculation of the sales momentum coefficient and the flow momentum coefficient.

[0050] Step S303-4: Obtain the upper and lower limits of the bid as defined in the bid boundary rules. Use the boundary adjustment factor to scale the upper and lower limits of the bid respectively to obtain the corrected dynamic upper and lower limits of the bid. The interval formed by the dynamic upper and lower limits of the bid is determined as the dynamic bid boundary.

[0051] The preset sales trend benchmark value is the average sales trend of products in this category under normal sales conditions, derived from historical data statistics. The sales trend in the product operation feature vector is a trend indicator of sales changes within a preset period. Dividing the current sales trend by the sales trend benchmark value yields a ratio value, namely the sales momentum coefficient. If this coefficient is less than 1, it indicates that the current sales growth is slower than normal, resulting in sales backlog, and it is necessary to appropriately reduce the bid to prioritize profit margins; if the coefficient is greater than 1, it indicates that sales are growing rapidly, the product is selling well, and a higher bid can be accepted to capture more traffic.

[0052] Traffic characteristics are data indicators that comprehensively reflect changes in product traffic as a form of user behavior. The preset traffic characteristic benchmark value represents the natural traffic characteristics of this product category under stable market conditions. Dividing the current traffic characteristic by the traffic characteristic benchmark value yields the traffic momentum coefficient. When the traffic momentum coefficient is greater than 1, it indicates that the product is experiencing a higher-than-normal upward trend in traffic, and market popularity is rising. At this time, the bid ceiling should be widened to fully capture conversion opportunities within the rising momentum. When the coefficient is less than 1 or even negative, it indicates that traffic growth is weak or has entered a downward trend. In this case, the bid boundaries should be narrowed to control advertising costs.

[0053] Weights are set for the sales momentum coefficient and the traffic momentum coefficient, with the sum of the two weights being 1. Specific weight values ​​can be preset based on the differences in sales sensitivity and traffic sensitivity among different product categories. The boundary adjustment factor is obtained by multiplying the sales momentum coefficient by its weight and adding the traffic momentum coefficient by its weight. This adjustment factor comprehensively reflects the current supply-demand dynamics of the product: the pressure to clear sales volume and the market's enthusiasm for traffic simultaneously influence the adjustment decision of the bidding boundary, reducing the problem of biased adjustment direction caused by relying on only a single indicator.

[0054] The fixed values ​​of the upper and lower bid limits are retrieved from the aforementioned bid boundary rules. The upper bid limit is multiplied by a boundary adjustment factor to obtain the adjusted dynamic bid upper limit, and the lower bid limit is multiplied by the boundary adjustment factor to obtain the adjusted dynamic bid lower limit. For products in their peak promotion period, if the positive momentum of traffic characteristics exceeds the pressure of sales trends, the boundary adjustment factor is greater than 1, and both the dynamic bid upper and lower limits are adjusted upwards simultaneously, raising the overall bid range to meet the strategy requirements of capturing high-quality traffic. For products in their clearance period, if the sales momentum coefficient shows severe sales pressure and traffic characteristics continue to decline, the boundary adjustment factor is less than 1, and the dynamic bid boundary shrinks downwards, anchoring to profit rather than scale. The final dynamic value range defined by the dynamic bid upper and lower limits serves as the dynamic bid boundary for this round of campaigning. This boundary will serve as a mandatory constraint range for subsequent dynamic bidding calculations, ensuring that the real-time bid for each keyword not only conforms to the guidance of the macro-operational stage strategy but also maintains adaptive sensitivity to the current sales and traffic situation.

[0055] The budget allocation rules and keyword matching pattern rules in step S302 are combined with the dynamic bidding boundary in step S303 according to the template's structural definition to instantiate a callable advertising strategy engine. The engine encapsulates the core logic and parameter constraints required for all advertising decisions in the current operational phase of the product, providing an interface to generate specific advertising execution instructions. On one hand, by structuring domain knowledge through strategy templates, it enables rapid strategy reuse and consistent decision-making, solving the problems of strategy switching relying on human experience and delayed response. On the other hand, by introducing two real-time dimensions—sales status and traffic trends—to continuously and dynamically scale the bidding range, the originally fixed phase strategy boundary gains adaptive fine-tuning capabilities. While maintaining the macro-strategy direction, this allows the bidding range to respond sensitively to changes in market supply and demand, reducing budget waste or traffic loss due to rigid bidding ranges and improving the coordination accuracy between advertising strategies and the real-time operational status of the product.

[0056] After instantiating the advertising strategy engine—that is, determining the budget allocation rules, keyword matching rules, and dynamic bidding boundaries adjusted by real-time data to suit the current target operational stage of the product—the strategy engine needs to deeply integrate the above rule framework with the product's real-time operational status data to calculate the precise execution instructions for this round of advertising. This process is a crucial step in transforming macro-strategic guidance into micro-execution parameters, ensuring that budget allocation, bidding range, and keyword matching rules not only adapt to the operational stage at the strategic level but also respond to current sales pressure and traffic trends at the numerical level, thereby generating specific and adaptive advertising execution instructions. The process of generating advertising execution instructions containing budget allocation, bidding range, and keyword matching rules through the advertising strategy engine includes the following steps: Step S305: Obtain the baseline budget allocation ratio and budget ceiling defined in the strategy configuration template, use sales trends and traffic characteristics to weight and adjust the baseline budget allocation ratio to obtain the dynamic budget allocation weight, and calculate the target budget configuration based on the dynamic budget allocation weight and budget ceiling.

[0057] Step S306: Construct a bid range that matches the target budget configuration based on the dynamic bid upper limit and dynamic bid lower limit.

[0058] Step S307: Extract the keyword matching mode type from the strategy configuration template, compare it with the preset matching mode switching threshold based on traffic characteristics, and dynamically select the keyword matching mode type according to the comparison result to obtain the target keyword matching rule.

[0059] Step S308: The target budget configuration, bid range, and target keyword matching rules are uniformly encapsulated into a structured ad delivery execution instruction.

[0060] The strategy engine first reads the preset baseline budget allocation ratio and budget ceiling from the matched strategy configuration template. However, this static ratio cannot perceive the real-time inventory turnover and market traffic. If used directly, it may result in a large budget being invested when sales are under pressure, or insufficient budget to seize opportunities when traffic surges. Therefore, a dynamic weighted adjustment mechanism is introduced. The engine extracts sales trend and traffic features from the product operation feature vector and compares them with preset category baseline values ​​to generate sales momentum coefficients and traffic momentum coefficients. The specific calculation method is the same as in steps S303-1 and S303-2. Then, through a configurable weighting function, such as a weight coefficient preset according to the sensitivity of the product category to sales or traffic, these two factors are merged into a budget adjustment coefficient. The engine multiplies this budget adjustment coefficient by the baseline budget allocation ratio to obtain the dynamic budget allocation weight. Finally, constraint calculation is performed based on this dynamic weight and the budget ceiling: the total budget is multiplied by the dynamic weight to obtain the suggested budget value, and then the smaller value between the suggested budget value and the budget ceiling is taken as the final target budget configuration. This way, when a product's sales trend is positive and traffic characteristics show an upward trend, the budget allocation weight will be amplified, automatically tilting more budget towards that product to seize traffic dividends; conversely, when sales pressure increases and traffic declines, the weight will be compressed, decisively reducing the budget to control costs. This achieves adaptive budget configuration.

[0061] The strategy engine directly obtains the modified dynamic bid upper and lower limits, defining them structurally as a closed bid range. This bid range does not exist independently but is strongly matched with the target budget configuration calculated in step S305: the budget size determines the overall cost of bidding, while the bid range defines the bid range for each click within this range. The two work together to ensure that, under budget constraints, the bid for a single keyword can fully participate in the competition without excessively depleting the budget due to high cost per click, achieving a seamless connection between macro budget and micro bid.

[0062] The strategy configuration template defines the default keyword matching mode type for that operational phase. For example, broad match is often used during the new product launch period to explore traffic, while exact match is often used during the best-selling product period to target high-intent users. To enable dynamic switching of matching modes, the engine extracts the default matching mode type from the strategy template as candidates, while also reading traffic features from the product operation feature vector. The optimization module presets a series of matching mode switching thresholds, which are empirical traffic trend boundaries. The engine compares current traffic characteristics with these thresholds: when traffic characteristics exceed the preset surge threshold (indicating the product is on an upward traffic trajectory and strong market interest), even if the template defaults to exact match, the engine automatically and dynamically selects or adjusts the keyword matching rules to broad match or phrase match to maximize the capture of relevant demand and long-tail opportunities in the rising traffic pool, thereby expanding traffic reach; conversely, when traffic characteristics fall below the preset convergence threshold or are even negative (indicating declining traffic or intensified competition), the engine dynamically tightens the matching rules to exact match, filtering out low-relevance general traffic and concentrating budget on core search terms with high intent and high conversion probability, ensuring every penny is spent effectively and stabilizing conversion results. This adaptive selection of matching modes driven by real-time traffic trends ensures that the traffic acquisition strategy always resonates with market trends.

[0063] Finally, the engine assembles these parameters from different decision branches into a structured data object, namely the ad execution instruction, according to the preset data interaction protocol. This instruction is a self-contained operation package that can be unambiguously interpreted by downstream execution modules. It explicitly specifies all the placement constraints and execution standards that should be followed for a specific product or ad unit in the next execution cycle, including how much money to spend (target budget configuration), what cost range to bid in (bid range), and how to match search terms (target keyword matching rules). This unified encapsulation method decouples strategy decision-making from instruction execution, making the complex intelligent calculations at the upper level independent of the high-frequency placement operations at the lower level, ensuring clarity, scalability, and execution reliability. In this way, budget allocation weights, bid boundaries, and keyword matching rules can be fine-tuned in real time and in linkage with the sales performance of the product and the market traffic. This not only allows the placement instruction to be precisely synchronized with the actual operational pulse of the product at this moment, maximizing budget utilization efficiency and traffic capture accuracy, but also completes a creative leap from staged strategy to adaptive execution without human intervention, improving the intelligence level and overall return on investment of ad placement.

[0064] Step S400: Based on the advertising execution instruction, perform dynamic bidding calculation and automatic optimization on the keywords to obtain refined advertising control parameters that meet the placement constraints, and then execute the advertising placement.

[0065] After generating the ad delivery execution instruction, which includes the target budget configuration, bid range, and target keyword matching rules, the instruction is passed to the execution and optimization layer. The process of dynamically calculating and automatically optimizing keyword bids based on the ad delivery execution instruction to obtain refined ad control parameters that meet the delivery constraints, and then executing the ad delivery, includes the following steps: Step S401: Obtain the real-time performance characteristics of each keyword in the ad delivery data view.

[0066] Step S402: Perform dynamic value assessment on the real-time performance features to obtain the real-time bidding valuation for each keyword, and use the bid range to perform boundary constraint processing on the real-time bidding valuation to obtain dynamic bidding parameters.

[0067] Step S403: Perform a matching performance evaluation on the traffic characteristics of each keyword and the matching rules of the target keyword to obtain a matching mode adjustment instruction.

[0068] Step S404: Perform invalid word identification processing on the search term data in the advertising delivery data view to obtain a set of keywords to be filtered.

[0069] Step S405: Integrate and encapsulate the dynamic bidding parameters, matching mode adjustment instructions, and the set of keywords to be filtered to generate refined advertising control parameters, and execute advertising delivery based on the refined advertising control parameters.

[0070] Specifically, using keywords as indexes, real-time performance features for each keyword are extracted directly from ad delivery data views that support multi-dimensional real-time queries. These features include, but are not limited to, time-series indicators such as click-through rate (CTR) in the most recent hour, conversion rate in the most recent day, and conversion cost, providing a high-quality data foundation for subsequent value assessment. Then, dynamic value assessment and bidding constraints are performed on each keyword. An independent probabilistic prediction model based on an online learning algorithm (such as FTRL) is built for each keyword. This model integrates real-time performance features, contextual features representing the current time period and user intent, and competitive features representing the intensity of market competition, performing non-linear calculations to predict the real-time bid valuation that maximizes the return on investment. This process can be expressed as: Optimal Bid ≈ Estimated Conversion Value × Estimated Click-Through Rate × Budget Utilization Factor, where the budget utilization factor dynamically scales with the overall budget consumption progress of the promotion plan. When consumption is insufficient, the valuation is increased to actively acquire traffic; when consumption is too rapid, the valuation is decreased to smooth expenditure. The real-time bidding valuation is then constrained within the dynamic bidding range defined in the ad execution instructions. If the valuation exceeds the upper limit, the upper limit prevails; if it falls below the lower limit, the lower limit prevails, forming dynamic bidding parameters that conform to the macro strategy boundaries and respond to micro market changes. This dynamically optimizes the return on investment within a limited budget, balancing traffic acquisition and cost control.

[0071] Next, the keyword matching mode is adaptively adjusted. This involves evaluating the matching efficiency of each keyword's traffic characteristics against the target keyword's matching rules. Specifically, when the target matching rule in the instruction is broad match, the percentage of irrelevant search term clicks it generates is monitored. If it exceeds a preset threshold, an adjustment instruction to change to phrase match is automatically generated. Conversely, if the impressions of exact match keywords decrease sharply due to competition, an instruction to change to broad match is generated to re-explore traffic opportunities, achieving a dynamic balance between traffic breadth and precision. Then, real-time filtering of invalid keywords is performed. Based on the keyword loss mitigation model, search term data in the ad delivery data view is queried, and the total cost and total conversions of each keyword over the last N clicks are aggregated and calculated. If the total cost of a keyword exceeds a preset multiple of the target cost per conversion and the total number of conversions is zero, it is identified as an invalid keyword, and a set of keywords to be filtered, including precise negative operations, is generated to achieve real-time loss mitigation. Finally, the dynamic bidding parameters, matching mode adjustment instructions, and the set of keywords to be filtered generated in the above steps are integrated and packaged to generate refined advertising control parameters. These parameters are then executed in real time via API, completing the closed-loop control of ad placement. This achieves automated and refined breakdown and execution from macro-strategy instructions to micro-level bidding, matching modes, and start / stop for each keyword, significantly improving budget utilization efficiency, traffic capture accuracy, and real-time loss mitigation capabilities. Ultimately, it achieves continuous optimization of ROI for ad placement without frequent human intervention.

[0072] Preferably, after executing the ad campaign, the process also includes: linking the post-campaign performance data with historical optimization operations for tracking, and incrementally updating the preset classification model and ad campaign strategy engine through online learning algorithms to form a closed-loop iteration for ad campaign optimization.

[0073] After executing ad placements based on refined ad control parameters, a self-evolving closed-loop iterative mechanism is constructed. This mechanism establishes a causal link between each optimization operation automatically executed by the optimization engine and the real feedback from the market, and uses this real-time generated knowledge to incrementally train and optimize the core decision-making model, thus forming a spiral-like upward closed loop of decision-making, execution, feedback, and evolution. First, a complete decision event recording and effect correlation tracking mechanism needs to be established. Every operation successfully executed based on refined ad control parameters generated by the ad placement strategy engine, such as keyword bid fine-tuning, matching mode switching, keyword addition or negation, will be captured and encapsulated as a structured historical optimization operation (decision event). This decision event includes at least: a unique decision identifier, a precise decision timestamp, the decision source (i.e., the specific strategy engine that triggered the operation, such as the "Top Picks" engine), a complete snapshot of the decision content (i.e., the specific parameter changes before and after the operation), and the execution status. Simultaneously, the ad placement data view in step S100 will continuously and objectively record the micro-changes in ad performance after this decision execution. By aligning and associating the timestamps and target objects of decision events with performance metrics (such as click-through rate, conversion rate, and cost per conversion) in the ad delivery data view, a causal-relationship-based dataset is automatically constructed. The changes in performance after each decision event are associated are further calculated and judged: if performance metrics (such as click-through rate and ROI) show a positive increase, the decision event and its corresponding operational context are labeled with a positive feedback tag; conversely, if performance metrics decrease, they are labeled with a negative feedback tag. Thus, labeled feedback samples are generated in real time.

[0074] Once the accumulated feedback samples reach the preset minimum batch size or time window, the incremental update process is initiated. This process does not involve lengthy retraining using all historical data; instead, it employs algorithms supporting online learning (such as FTRL, Follow The Regularized Leader) to incrementally fine-tune the model parameters. For updates to the preset classification model, the update target is the tree structure or leaf node scores of its gradient boosting model (such as XGBoost). Using the product operation feature vector from the labeled feedback samples as input and the correct operational stage as the true label, the model undergoes one or more rounds of iterative learning, fine-tuning the splitting conditions or prediction scores of the decision tree. This corrects the model's bias in judging the product lifecycle, making its perception of market trends and product status more sensitive and accurate. For example, when changes in the market environment lead to a general shortening of the testing period for new products in a certain category, the model can automatically adjust the relevant stage discrimination thresholds through continuous learning. For updates to the advertising strategy engine, the dynamic parameter generation logic in its internal strategy configuration template is optimized. The feedback samples reveal the real effects of specific bidding strategies, budget allocation schemes, and keyword matching rule combinations at different operational stages. These successful strategy parameter combinations (such as the fusion weight of sales momentum coefficient and traffic momentum coefficient during peak promotion periods) are used as positive learning targets. An online learning algorithm is used to adjust the weighting function coefficients of the dynamic bidding boundary and dynamic budget allocation weights generated in the engine. This allows the strategy engine to autonomously learn which fine-tuning range of bids and budgets will bring the best returns in a specific competitive environment, thereby continuously optimizing its strategy generation capabilities.

[0075] Figure 2 This is a block diagram of the advertising optimization system based on an e-commerce platform provided in this application. For example... Figure 2 As shown, the advertising optimization system based on an e-commerce platform includes: a preprocessing module, used to collect and integrate multi-source heterogeneous advertising and product data from the e-commerce platform, and perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising and product data to obtain a unified standard advertising data view; a discrimination module, used to extract product operation feature vectors based on the advertising data view, and use a preset classification model to perform stage discrimination on the product operation feature vectors to obtain the target operation stage of the product; a matching module, used to match the corresponding advertising strategy engine from a preset strategy library according to the target operation stage, and use the advertising strategy engine to generate advertising execution instructions containing budget configuration, bid range, and keyword matching rules; and an optimization module, used to perform dynamic bidding calculation and automatic optimization processing on keywords according to the advertising execution instructions to obtain refined advertising control parameters that meet the placement constraints and execute advertising.

[0076] The other functions performed by the preprocessing module, discrimination module, matching module, and optimization module, as well as the technical details of each function, are the same as or similar to the corresponding features in the advertising optimization method based on an e-commerce platform described above, so they will not be repeated here.

[0077] This application also provides a computer storage medium storing a computer program that, when run on a computer, enables the computer to execute the steps in the previously described method for optimizing advertising placement on an e-commerce platform.

[0078] It should be understood that although the steps in the flowcharts in the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order requirement for the execution of these steps, and they can be performed in other orders.

[0079] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for optimizing advertising placement on an e-commerce platform, characterized in that, The method includes: Collect and integrate multi-source heterogeneous advertising data and product data from e-commerce platforms, and perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view; Based on the advertising data view, product operation feature vectors are extracted, and the target operation stage of the product is obtained by performing stage discrimination on the product operation feature vectors through a preset classification model. According to the target operation stage, the corresponding advertising strategy engine is matched from the preset strategy library, and the advertising strategy engine generates an advertising execution instruction containing budget configuration, bid range and keyword matching rules. Based on the advertising execution instructions, the keywords are dynamically bid and automatically optimized to obtain refined advertising control parameters that meet the placement constraints, and then the advertising is executed.

2. The method according to claim 1, characterized in that, The process of standardizing, cleaning, and aggregating the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view includes: The multi-source heterogeneous advertising data and the multi-source heterogeneous fields in the product data are mapped to a set of standard fields with unified semantics through a preset standardized data dictionary; Abnormal data values ​​in the standard field set are identified and removed using preset anomaly detection rules. The mean of the source data in the preceding time window is used to fill in the removed and the original missing data values ​​to obtain cleaned standard data with high confidence. Account structure data, campaign performance data, and product dimension data are extracted from the cleaned standard data, and the campaign performance data and product dimension data are subjected to rolling aggregation calculations at a preset period to generate aggregated derivative indicators with preset business entities as anchors. The account structure data and the aggregated derived metrics are written into an in-memory database built on an online analytics engine to form an ad delivery data view that supports multi-dimensional real-time queries.

3. The method according to claim 1, characterized in that, The step of determining the target operational stage of a product by using a preset classification model to classify the product's operational feature vector includes: The product operation feature vector is input into a preset classification model constructed by a rule engine and a gradient boosting model; The rule engine performs preset threshold logic judgments on the number of days on the shelf, sales fluctuation coefficient and sales trend in the product operation feature vector to obtain preliminary stage classification results. Meanwhile, the confidence score of each candidate operation stage is obtained by nonlinear cross-calculation of the sales time series of the near preset number of days, conversion rate fluctuation coefficient, sales trend, price competitiveness and traffic characteristics in the product operation feature vector through the gradient boosting model. The preliminary stage classification results and the confidence scores of each candidate operation stage are weighted and fused to obtain a comprehensive confidence score set. The operation stage corresponding to the highest comprehensive confidence score in the comprehensive confidence score set is determined as the target operation stage of the product.

4. The method according to claim 3, characterized in that, The step of matching the corresponding advertising delivery strategy engine from the preset strategy library according to the target operation stage includes: Obtain the stage identifier corresponding to the target operation stage, and retrieve the strategy configuration template that is bound to the stage identifier from the preset strategy library; Extract the budget allocation rules, bid boundary rules, and keyword matching pattern rules defined in the strategy template; The dynamic bidding boundary is obtained by using the sales trend and traffic features in the product operation feature vector to perform interval constraint correction on the bidding boundary rule; The budget allocation rules, the keyword matching mode rules, and the dynamic bidding boundaries are filled into the strategy configuration template to generate an advertising delivery strategy engine that is adapted to the target operational stage.

5. The method according to claim 4, characterized in that, The step of using the sales trend and traffic features in the product operation feature vector to perform interval constraint correction on the bidding boundary rule to obtain the dynamic bidding boundary includes: The sales momentum coefficient is obtained by calculating the ratio between the sales trend and the preset sales trend benchmark value. The flow momentum coefficient is obtained by calculating the ratio between the flow characteristic and a preset flow characteristic benchmark value; The boundary adjustment factor is obtained by weighted fusion calculation of the sales momentum coefficient and the flow momentum coefficient; Obtain the upper and lower bid limits defined in the bid boundary rules, and use the boundary adjustment factor to proportionally scale the upper and lower bid limits to obtain the corrected dynamic upper and lower bid limits. The interval formed by the dynamic upper and lower bid limits is determined as the dynamic bid boundary.

6. The method according to claim 5, characterized in that, The process of generating ad delivery execution instructions, which include budget configuration, bid range, and keyword matching rules, through the ad delivery strategy engine includes: Obtain the baseline budget allocation ratio and budget ceiling defined in the strategy configuration template, use the sales trend and the traffic trend to weight and adjust the baseline budget allocation ratio to obtain the dynamic budget allocation weight, and calculate the target budget configuration based on the dynamic budget allocation weight and the budget ceiling. A bid range matching the target budget configuration is constructed based on the dynamic bid upper limit and the dynamic bid lower limit; The keyword matching mode type is extracted from the strategy configuration template, and compared with the traffic characteristics and the preset matching mode switching threshold. The target keyword matching rule is obtained by dynamically selecting the keyword matching mode type according to the comparison result. The target budget configuration, the bid range, and the target keyword matching rules are uniformly encapsulated into a structured ad delivery execution instruction.

7. The method according to claim 6, characterized in that, The process of dynamically calculating and automatically optimizing keywords based on the advertising execution instructions to obtain refined advertising control parameters that meet the placement constraints, and then executing the advertising placement, includes: Obtain the real-time performance characteristics of each keyword in the ad delivery data view; The real-time performance features are dynamically evaluated to obtain the real-time bid valuation for each keyword, and the bid range is used to perform boundary constraint processing on the real-time bid valuation to obtain dynamic bidding parameters. The matching efficiency of each keyword's traffic characteristics and the target keyword matching rules is evaluated to obtain a matching mode adjustment instruction; The search term data in the advertising delivery data view is processed to identify invalid words, resulting in a set of keywords to be filtered. The dynamic bidding parameters, the matching mode adjustment instructions, and the set of keywords to be filtered are integrated and encapsulated to generate refined advertising control parameters, and advertising is executed according to the refined advertising control parameters.

8. The method according to claim 1, characterized in that, The process of executing the advertisement delivery also includes: The system links and traces the performance data after ad placement with historical optimization operations, and uses online learning algorithms to incrementally update the preset classification model and ad placement strategy engine to form a closed-loop iteration for ad placement optimization.

9. An advertising optimization system based on an e-commerce platform, characterized in that, The system includes: The preprocessing module is used to collect and integrate multi-source heterogeneous advertising data and product data from e-commerce platforms, and to perform standardized cleaning and dimensional aggregation processing on the multi-source heterogeneous advertising data and product data to obtain a unified standard advertising delivery data view. The discrimination module is used to extract product operation feature vectors based on the advertising data view, and to determine the target operation stage of the product by performing stage discrimination on the product operation feature vectors through a preset classification model. The matching module is used to match the corresponding advertising strategy engine from the preset strategy library according to the target operation stage, and generate an advertising execution instruction containing budget configuration, bid range and keyword matching rules through the advertising strategy engine; The optimization module is used to perform dynamic bidding calculation and automatic optimization of keywords based on the advertising execution instructions to obtain refined advertising control parameters that meet the placement constraints and then execute the advertising.

10. A computer-readable storage medium storing a computer program executable on a processor, characterized in that, When the computer program is executed by the processor, it implements an advertising optimization method based on an e-commerce platform as described in any one of claims 1 to 8.

Images