Advertisement putting parameter adaptation method and system fused with scene analysis

Abstract

The invention relates to the field of advertisement putting optimization, in particular to an advertisement putting parameter adaptation method and system fused with scene analysis, and the method comprises the steps: obtaining multi-dimensional features of a plurality of known scenes in historical advertisement putting data; in combination with a preset feature embedding model, mapping the multi-dimensional features of the plurality of known scenes into high-dimensional vectors, and generating a known scene vector set to construct a scene vector space; in response to the received new advertisement putting request, obtaining multi-dimensional features of a new scene; based on the same feature embedding model, mapping the multi-dimensional features of the new scene into a high-dimensional vector, and generating a new scene vector; determining at least one optimal associated scene from the known scenes according to the new scene vector; generating an initial advertisement putting parameter of the new scene in combination with the historical advertisement putting parameter of the at least one optimal associated scene; and obtaining effect feedback data of the initial advertisement putting parameters in real time, and generating optimized advertisement putting parameters according to the initial advertisement effect feedback data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advertisement delivery optimization, in particular to an advertisement delivery parameter adaptation method and system fusing scene analysis. BACKGROUND

[0002] In programmatic advertisement delivery, quickly adapting accurate delivery parameters to new advertisement scenes is the key to improving advertisement effectiveness and return on investment. The existing technology usually uses shallow scene association technology based on context keyword matching and static strategy migration technology based on historical effect data.

[0003] However, in the existing technology, the shallow scene association technology based on context keyword matching cannot understand the deep semantics and emotional tendencies of the scene, leading to serious misdelivery. A typical failure case is that: next to a negative news report about "new energy vehicle battery safety", the system simply classifies it as a "car information" scene because it contains the keyword "new energy vehicle" and blindly delivers new energy vehicle advertisements. Such delivery not only has very low conversion rate, but also causes serious damage to the brand image. The root cause is that the existing technology can only perform "surface association", but cannot understand the essence of "strategic association" between the scene and the advertisement strategy. Secondly, in the static strategy migration technology based on historical effect data, its static perspective cannot adapt to the dynamic changing competitive environment. For example, the system may find that a historical scene "users browse travel blogs in the evening on weekends" has a high click rate and a low delivery cost at that time. Therefore, when a new scene appears with similar features, the system follows the historical low bid strategy. However, it ignores that the historical data may be collected during the tourism off-season and the competition is mild; while the current new scene may be in the tourism peak season and the bidding environment is extremely fierce. This practice of applying the successful strategy in the past when the competition was mild to the current competitive environment leads to ineffective strategy migration.

[0004] Therefore, the existing technology has defects and needs to be improved. SUMMARY

[0005] In order to solve one or more problems in the prior art, the main purpose of the present application is to provide an advertisement delivery parameter adaptation method and system fusing scene analysis.

[0006] In order to achieve the above-mentioned purpose of the application, the present application provides an advertisement delivery parameter adaptation method fusing scene analysis, which comprises: obtaining multi-dimensional features of a plurality of known scenes in historical advertisement delivery data; mapping the multi-dimensional features of the plurality of known scenes into high-dimensional vectors by combining a preset feature embedding model, generating a known scene vector set, and constructing a scene vector space; In response to receiving the new advertisement placement request, multi-dimensional features of the new scene are acquired; Based on the same feature embedding model, the multi-dimensional features of the new scene are mapped into a high-dimensional vector to generate a new scene vector; According to the new scene vector, at least one optimal associated scene is determined from the known scenes; Combined with the historical advertisement placement parameters of the at least one optimal associated scene, initial advertisement placement parameters of the new scene are generated; Real-time effect feedback data of the initial advertisement placement parameters are acquired, and based on the initial advertisement effect feedback data, optimized advertisement placement parameters are generated.

[0007] The application also provides an advertisement placement parameter adaptation system integrating scene analysis, comprising: A first acquisition module is configured to acquire multi-dimensional features of a plurality of known scenes in historical advertisement placement data; A generation module is configured to map the multi-dimensional features of the plurality of known scenes into high-dimensional vectors based on a preset feature embedding model to generate a known scene vector set and construct a scene vector space; A second acquisition module is configured to acquire multi-dimensional features of a new scene in response to receiving a new advertisement placement request; A mapping module is configured to map the multi-dimensional features of the new scene into a high-dimensional vector based on the same feature embedding model to generate a new scene vector; A determination module is configured to determine at least one optimal associated scene from the known scenes according to the new scene vector; A combination module is configured to generate initial advertisement placement parameters of the new scene by combining historical advertisement placement parameters of the at least one optimal associated scene; An optimization module is configured to acquire real-time effect feedback data of the initial advertisement placement parameters, and generate optimized advertisement placement parameters based on the initial advertisement effect feedback data.

[0008] The application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method according to any one of the preceding embodiments.

[0009] The application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method according to any one of the preceding embodiments.

[0010] The advertisement putting parameter adaptation method and system of the fusion scene analysis of the embodiment of the application, through utilizing the preset feature embedding model, map the multi-dimension features of the scene to high-dimension vectors, and construct the scene vector space rich in semantic information. This basic step enables the system to deeply understand the connotation of the scene, fundamentally overcomes the limitation of the prior art based on the matching of the shallow features such as keywords, realizes the leap from "similar in shape" to "similar in spirit", significantly improves the accuracy of the scene analysis, and effectively avoids the risk of damaging the brand image by mistakenly putting the advertisement beside the negative content. On this basis, the system can accurately locate the optimal associated scene of the strategy migration from the historical experience library by calculating the similarity between the new scene vector and the known scene vector, thereby generating high-quality initial putting parameters for the new scene. This mechanism perfectly solves the cold start problem, provides a high starting point for intelligent initialization for the new scene, greatly shortens the trial and error period, and reduces the waste of early budget. Finally, the system forms a strong self-evolution ability by collecting feedback data in real time and dynamically generating optimization parameters. This closed-loop optimization mechanism ensures that the system can quickly adapt to the characteristics of the new scene and the dynamic competitive environment, and realizes the continuous improvement of the putting effect. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 The flowchart of the advertisement putting parameter adaptation method of the fusion scene analysis of an embodiment of the application; Figure 2 The flowchart of the advertisement putting parameter adaptation method of the fusion scene analysis of an embodiment of the application; Figure 3 The structural schematic block diagram of the advertisement putting parameter adaptation system of the fusion scene analysis of an embodiment of the application; Figure 4 The structural schematic block diagram of the computer device of an embodiment of the application.

[0012] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0014] With reference to Figure 1 The advertisement putting parameter adaptation method of the fusion scene analysis of an embodiment of the application comprises: S1, obtaining the multi-dimension features of a plurality of known scenes in the historical advertisement putting data; S2, in combination with a preset feature embedding model, map the multi-dimensional features of the plurality of known scenes into high-dimensional vectors to generate a set of known scene vectors, so as to construct a scene vector space; S3, in response to receiving a new advertisement placement request, obtaining multi-dimensional features of a new scene; S4, based on the same feature embedding model, map the multi-dimensional features of the new scene into a high-dimensional vector to generate a new scene vector; S5, according to the new scene vector, determine at least one optimal associated scene from the known scenes; S6, in combination with the historical advertisement placement parameters of the at least one optimal associated scene, generate initial advertisement placement parameters of the new scene; S7, real-time acquisition of initial advertisement placement parameter effect feedback data, according to the initial advertisement effect feedback data, generate optimized advertisement placement parameters.

[0015] As described in steps S1-S3 above, the "known scene" refers to a scene whose effect (such as click rate, conversion rate) and optimal parameters have been verified in historical delivery. The "multi-dimensional feature" aims to comprehensively and stereoscopically describe a scene, including: user features: user portrait (age, gender, interest label), historical behavior (search, browse, purchase record). Environment features: device type (mobile phone / PC), operating system, network environment (Wi-Fi / 4G), geographic location (GPS coordinates, business district). Content features: text content of the current browsing page (keywords, topics, entities, sentiment orientation extracted by NLP), media type (text / image / video). Time features: time of day (morning / afternoon / night), day of the week, whether it is a holiday. Through the fusion of multi-dimensional features, a rich data foundation is provided for subsequent deep semantic analysis, avoiding the one-sidedness of single-dimensional judgment, and is the data cornerstone of realizing accurate scene understanding. The feature embedding model (such as a deep learning model based on Transformer) can compress unstructured, high-dimensional sparse original features into a low-dimensional, dense vector. In this process, the model learns from massive data, so that scenes with similar semantics or strategies in the original feature space have a closer distance (such as cosine similarity) in the vector space. The "scene vector space" composed of vectors of all known scenes becomes the "strategy experience library" of the system. Vector representation can capture deep semantics that keyword matching cannot express (for example, it can distinguish between "negative news about new energy vehicle battery safety" and "new energy vehicle performance evaluation"). Converting complex scene similarity comparison into efficient vector space distance calculation makes it possible to quickly and accurately find related scenes. "Digitize" and "store" the new scene. Its processing flow is completely consistent with that of the known scene, ensuring that the new scene vector and the known scene vector are in the same vector space and are comparable. This is a prerequisite for any meaningful comparison. It guarantees the fairness and consistency of the system for new scenes, making the vector space-based similarity measurement have practical significance and achieving unified measurement of new and old scenes under the same standard.

[0016] As described in steps S5-S7 above, in the constructed scenario vector space, the K closest "known scenario vectors" to the "new scenario vector" are quickly found through algorithms such as nearest neighbor search. These known scenarios are considered as the "optimal associated scenarios". The underlying assumption is that the more similar the scenarios in the vector space, the more similar the optimal ad placement strategies applicable to them. A brand new scenario can immediately find the most worthy "mentors" from the vast amount of historical experience, thus breaking away from the traditional cold start mode of completely random exploration or relying on rough rules. One or more "optimal associated scenario" historical proven successful placement parameters (such as bid, creative ID, target audience package, etc.) are obtained, and an initial ad placement parameter set for the new scenario is generated through weighted averaging, voting or model fusion, etc. A high-quality, high-starting-point initial solution is provided for the new scenario. This avoids budget waste or opportunity loss caused by improper parameters in the early cold start, so that the ad is in a relatively optimal state from the beginning. The system will collect feedback data (such as exposure, clicks, conversions, etc.) generated after the initial parameters are used in the new scenario, and use these data to fine-tune the parameters. This is usually achieved through online learning algorithms or reinforcement learning, enabling the system to adapt to the uniqueness of the new scenario and the dynamic changing market environment, thereby continuously improving ad effectiveness.

[0017] As described above, by using a pre-set feature embedding model to map the multi-dimensional features of a scenario into a high-dimensional vector, a scenario vector space rich in semantic information is constructed. This fundamental step enables the system to deeply understand the connotation of the scenario, fundamentally overcoming the limitations of existing technologies based on matching of shallow features such as keywords, achieving a leap from "similar in appearance" to "similar in essence", significantly improving the accuracy of scenario analysis, and effectively avoiding the risk of misplacing ads next to negative content and damaging brand image. On this basis, the system can accurately locate the optimal associated scenarios with transferable strategies from the historical experience library by calculating the similarity between the new scenario vector and the known scenario vector, thereby generating high-quality initial placement parameters for the new scenario. This mechanism perfectly solves the cold start problem, providing a high starting point for intelligent initialization for new scenarios, greatly shortening the trial-and-error period and reducing early budget waste. Finally, the system forms a strong self-evolution ability by collecting feedback data in real time and dynamically generating optimized parameters. This closed-loop optimization mechanism ensures that the system can quickly adapt to the characteristics of the new scenario and the dynamic competitive environment, achieving continuous improvement of placement effectiveness.

[0018] Referring to Figure 2 In one embodiment, the step of generating optimized ad placement parameters according to the initial ad effectiveness feedback data comprises: S71, real-time acquisition of multi-element feedback data generated by the initial ad placement parameters; S72, performing credibility analysis on the multi-element feedback data, and generating a credibility weight of each piece of feedback data according to the analysis result; S73, performing multi-dimensional attribution analysis based on the multi-element feedback data with the credibility weight, and extracting an optimization factor causing fluctuation of the advertising effect, the optimization factor including a positive driving factor and a negative inhibiting factor; S74, inputting the optimization factor into a pre-trained parameter optimization prediction model, and predicting an optimization direction and adjustment amplitude of the advertising delivery parameter through the parameter optimization prediction model; S75, generating the optimized advertising delivery parameter based on the prediction result.

[0019] As described in the above steps, the comprehensive effect of the advertisement is captured. The "multi-feedback data" goes beyond a single click or conversion signal, including: exposure data: measures the reach. Click data: measures initial interest. Conversion data: measures the ultimate value (purchase, registration, etc.). Negative feedback: such as cancel, close, hide, etc., directly reflects the user's aversion. Interaction depth: such as video viewing completion rate, page stay time. A three-dimensional effect evaluation system is built, providing rich, multi-dimensional information input for subsequent analysis, avoiding the limitations of a single indicator (such as click rate) leading to one-sided optimization direction. The credibility analysis of the multi-feedback data is the key to data cleaning and purification. It acknowledges that not all feedback data is equally credible. The analysis is based on: user value: the feedback weight of high-value historical users is higher. Behavior authenticity: through an anti-cheating model to identify and reduce the weight of false feedback such as fake traffic and robot traffic. Feedback behavior pattern: the credibility of instantaneous clicks and deep browsing clicks is different. It can effectively filter noise and cheating data, prevent the optimization system from being "polluted", and ensure that subsequent analysis is based on high-quality, reliable data, so as to make more reliable decisions. From "phenomenon" to "reason" reasoning process. Attribution analysis (such as SHAP, attribution model based on game theory) aims to answer: "which factor or factors led to the improvement or decline of the advertising effect" Positive driving factors: identify the factors that contribute most to positive effects (for example: "creative A" is particularly effective for "young users" in the "evening period"). Negative inhibitory factors: identify the bottlenecks that lead to poor results (for example: "bidding strategy B" leads to insufficient exposure in "high-competition environment"). Make the optimization process interpretable. The system is no longer a "black box", but can clearly point out the driving factors and inhibiting factors that affect the effect, providing a clear and direct action guide for the next precise adjustment. The prediction model (such as gradient boosting tree or lightweight neural network) works by learning the complex mapping relationship between "optimization factors -> optimal parameter adjustment" on historical data. It receives the factors output by the attribution analysis and outputs quantitative adjustment suggestions (for example: for the identified "negative inhibitory factor: insufficient bidding", the model may output "increase the bid by 8%"). It realizes automatic, quantitative and intelligent parameter tuning. It replaces the rule adjustment that relies on human experience, can find and apply complex strategies that the human brain cannot summarize, making the optimization process more efficient and scientific. Apply the "direction and amplitude" output by the prediction model to the current advertising parameters to generate a new set of optimized parameters. Complete the closed loop from analysis to action. It finally lands the intelligent analysis results of all the previous steps into executable operation instructions, driving the advertising effect to continuously approach the optimal solution.

[0020] In an embodiment, the determining at least one optimal associated scene from the known scenes according to the new scene vector comprises: obtaining a set of known scene vectors, calculating the spatial distance between the new scene vector and each known scene vector in the set of known scene vectors to obtain a first similarity corresponding to each known scene; obtaining historical advertising effect data of each known scene; and performing matching degree calculation on the current advertising delivery target of the new scene and the historical advertising effect data of each known scene to obtain a second similarity corresponding to each known scene; for each known scene, inputting the first similarity and the second similarity into a preset weighted fusion function for calculation, and outputting a comprehensive correlation degree of the known scene according to the calculation result; sorting all known scenes according to the comprehensive correlation degrees, and selecting one or more known scenes with the highest rankings as the optimal correlation scenes.

[0021] As mentioned above, based on the constructed scene vector space, the similarity degree of the new scene and each known scene in the underlying features such as user, environment and content is quantified by using measurement methods such as cosine similarity or Euclidean distance. The assumption behind it is that scenes with similar features may have similar user intentions. Fast and preliminary scene screening is achieved. It can efficiently retrieve a batch of candidate scenes that are "on the surface" most relevant from a large number of known scenes. This is the basic guarantee for efficient matching. Pay attention to whether the effect achieved by the historical scene is consistent with the goal pursued by the new scene. For example, if the goal of the new scene is to "improve conversion rate (CVR)", then a known scene with historical effect data showing "high CVR" will have a high second similarity, even if its feature vector is not very similar to the new scene. Conversely, a scene with similar features but a historical strategy goal of "brand exposure" (low CVR) will be assigned a lower second similarity. It fundamentally solves the "mis-migration" problem. It ensures that the system is looking for a "mentor" scene that is strategy-learnable and goal-aligned, rather than just a "similar-looking" scene. This perfectly avoids classic mistakes such as "advertising next to negative news" or "using a moderate competitive strategy in a fierce environment", and is the core of improving the accuracy of initial parameters. The two are combined through a weighted fusion function. This function can be fixed or dynamic (for example, when the advertiser's goal is "brand exposure", more emphasis is placed on the first similarity; when the goal is "conversion", more emphasis is placed on the second similarity). The purpose is to generate a final score that reflects both "shape" and "spirit". It realizes accurate quantitative evaluation. It gives each candidate scene a unified and comprehensive evaluation score (comprehensive correlation degree), allowing all scenes to be compared fairly under the same measurement standard, providing a scientific and quantitative basis for the final selection. This is the final survival of the fittest link. The system sorts all known scenes based on the comprehensive correlation degree, the only standard, and automatically selects the top scene as the "optimal correlation scene". It completes the closed loop from "evaluation" to "selection". It outputs the highest-quality strategy source that has been double-checked, providing the most reliable basis for generating initial launch parameters for the new scene, thereby greatly improving the success rate and effect starting point of cold start.

[0022] In an embodiment, the multi-dimensional attribution analysis based on the multi-element feedback data with the credibility weight is performed to extract optimization factors causing fluctuations in advertising effects, and the steps include: determining whether the total amount of currently collected feedback data reaches an effective analysis threshold; if the total amount of currently collected feedback data does not reach the effective analysis threshold, extending the data collection time window, and performing weighted calculation on the historical advertising launch parameters of the optimal correlation scene according to the first similarity of each known scene to generate preliminary optimization parameters; If yes, start the multi-dimensional attribution analysis.

[0023] As mentioned above, based on the law of large numbers and the principle of data effectiveness in statistics. Attribution analysis (such as SHAP, counterfactual reasoning) is a powerful data analysis tool, but it needs enough data samples to guarantee the statistical significance and stability of its results. In the case of extremely small amount of data, the result variance of attribution analysis is extremely large, and it is easy to be biased by individual accidental events (such as an accidental click), resulting in misleading "optimization factors". A scientific decision threshold is established. It enables the system to self-diagnose whether it currently has the conditions to conduct in-depth analysis, thereby avoiding forced and high-risk optimization operations in the "data infancy period", reflecting the rigor and scientific nature of the system. A conservative and reliable optimization strategy is defined for the "data sparse period". It contains two actions: extending the data collection time window: the principle is "time for data", which accumulates more feedback samples by waiting for a longer time to create conditions for subsequent reliable analysis. Retreating to the strategy weighted based on the first similarity: this is the core risk-averse mechanism. Its principle is that when there is a lack of self-data, the most reliable source of knowledge is still the "scene vector space" and "feature semantic similarity (first similarity)" that the system relies on at initialization. At this time, the system no longer attempts to "innovate" and risks, but consolidates and fine-tunes its initial, deep semantic-based "best guess". It again utilizes the verified similarity to weight and fuse the known strategies, generating a more smooth and robust fine-tuning parameter. Ensuring the stability of the optimization direction. When in-depth analysis is not possible, a low-risk optimization path is provided, which still utilizes the most reliable information (feature similarity) to improve the parameters. When the data volume meets the requirements, the system activates its "advanced intelligence" - the complete optimization process, including credibility analysis, attribution analysis, and prediction model. Ensuring that the system can fully exert its powerful, data-driven optimization capabilities when the conditions are mature, making precise and in-depth strategy adjustments to maximize the improvement of advertising effectiveness.

[0024] In an embodiment, before the step of calculating the matching degree between the current advertising delivery target of the new scene and the historical advertising effect data of each known scene, the method further comprises: Obtaining a historical competitive environment intensity index corresponding to the generation of the known scene historical data; Obtaining a real-time competitive environment intensity index of the new scene; According to the difference between the historical competitive environment intensity index and the real-time competitive environment intensity index, the historical advertising effect data is discounted and compensated for correction; Based on the corrected historical advertising effect data, the matching degree is calculated again with the current advertising delivery target of the new scene.

[0025] As mentioned above, the historical competition environment intensity indicator is the metadata quantifying the intensity of market bidding at that time, including: auction participation: the average number of advertisers participating in bidding in the historical bidding request. Market price level: the average or quantile of the cost per thousand impressions (eCPM) or cost per click (CPC) of the historical same period and same type of ad position. Budget consumption rate: the consumption speed of the advertiser's budget at that time. The context of the historical effect is established. It restores a isolated historical effect data (such as "click rate 2%") to the specific market environment in which it was generated, enabling us to understand whether this "2%" was achieved in a mild or intense competition. This is the prerequisite for any cross-time comparison. Through the real-time data interface from the ad exchange platform (ADX), the same competition indicators as in step 1 are obtained, but reflecting the real-time market situation at the time of the new scene. The real and specific competition pressure faced by the new scene is determined, providing a benchmark for comparison with the historical environment. There is a strong correlation between ad effect (such as cost per click, conversion rate) and competition intensity. The more intense the competition, the higher the cost to achieve the same effect, or the worse the effect obtained with the same budget. If the real-time competition intensity > historical competition intensity, it is determined that the historical effect data is "achieved in a more relaxed environment", so it needs to be downwardly adjusted (such as increasing the historical cost data or decreasing the historical conversion rate data) to "devalue" it to an expected level that matches the current intense environment. Conversely, if the real-time competition is more relaxed, the historical data can be adjusted upwards, indicating that the expected effect can be better. The "de-time filter" processing of historical data is achieved. It eliminates data bias caused by market environment changes, so that a "past" successful strategy can truly reflect its expected effect in the "current" market environment after adjustment. This fundamentally solves the "carving a boat to seek a sword" type of strategy migration error. Instead of using the original, "moist" historical effect data, the system uses "adjusted historical ad effect data" that is more comparable to calculate the second similarity. This greatly improves the accuracy and reliability of the "strategy-oriented similarity (second similarity)". The system now looks for those historical scenes that "successfully achieved the goal of the new scene in a similar competition pressure to the current one". This ensures that the "optimal associated scene" and its strategy are truly feasible and can be learned from in the current environment.

[0026] In an embodiment, the step of discounting and compensating the historical ad effect data according to the difference between the historical competition environment intensity indicator and the real-time competition environment intensity indicator comprises: calculating the ratio of the real-time competition environment intensity indicator to the historical competition environment intensity indicator to obtain a competition intensity change rate; The competition intensity change rate is input into a preset nonlinear mapping function, and a mapping output is obtained to obtain an effect data discount factor; wherein the nonlinear mapping function is configured to: when the competition intensity change rate is greater than 1, the output value is less than 1, and the output value decreases with the increase of the input value; The historical advertising effect data of the known scene is multiplied by the effect data discount factor to obtain the modified historical advertising effect data.

[0027] As described above, the ratio (rather than the absolute value difference) is used to calculate the change rate, and the calculation formula is: competition intensity change rate = real-time index / historical index. When the change rate > 1: it means that the real-time competition is more intense than the historical period. When the change rate = 1: it means that the competition environment has not changed. When the change rate < 1: it means that the real-time competition is more moderate than the historical period. The difference between a multi-dimensional and complex competition environment is condensed into a single, dimensionless scalar value. This provides a clear and unified input signal for any subsequent correction function, so that the correction model can be independent of the specific index dimension and has universality and scalability. Based on deep insight into the law of advertising bidding market: reverse correction mechanism (output value < 1): when the competition intensifies (change rate > 1), it means that the effect achieved with the same cost in history will be more difficult to achieve in the current environment. Therefore, the historical effect data (such as high conversion rate and low cost) must be discounted downward, and a discount factor less than 1 is output to reduce its value assessment. Nonlinear decreasing characteristic: it is realized that the erosion effect of competition on effect is not linear, but conforms to the law of diminishing marginal utility. When the competition changes from “moderate” to “very intense”, the decline of the effect may be slow; but when the competition changes from “very intense” to “extremely intense”, the decline of the effect will be very sharp. A S-shaped function or an exponential decay function can well simulate this effect. Nonlinear mapping can more accurately reflect the complexity of the real market than simple linear discounting. Even in the face of extreme changes in competition intensity, the function can output a reasonable and smoothly changing discount factor to prevent the system from producing an over-reaction correction. Multiplication is used to directly act on the key performance indicators (KPIs) in the historical advertising effect data by the effect data discount factor obtained in the previous step, for example: modified click-through rate = historical click-through rate × discount factor. Modified conversion cost = historical conversion cost / discount factor (note: for cost indicators, the operation may be division, but the principle is the same, that is, adjustment based on factors). Multiplication is extremely efficient in calculation and meets the stringent real-time requirements of the advertising bidding system. The “modified historical advertising effect data” finally generated is an environment-calibrated effect that is estimated to be achieved by the historical strategy in the current competition environment. This makes the subsequent “second similarity” calculation based on fair and comparable basis.

[0028] In an embodiment, after the step of generating the optimized advertising delivery parameters according to the initial advertising effect feedback data, the method further comprises: In the advertising delivery process, the negative feedback signal in the advertising effect feedback data is monitored in real time. If the intensity of the negative feedback signal exceeds the emergency threshold within a preset time, the use of the current advertising delivery parameters is immediately suspended, and the initial advertising delivery parameters are returned.

[0029] As described above, the negative behavior signal directly related to user aversion is identified, which is usually more direct and faster than "low click rate" to reflect the serious failure of the strategy. The monitored negative feedback signal includes: actively hiding / closing the advertisement: the user explicitly expresses that he / she does not want to see the advertisement. Reporting the advertisement: the user thinks that the content of the advertisement is inappropriate, fraudulent or offensive. Negative comments / low score: direct negative evaluation left in the advertisement interaction interface. Real-time perception of the quality of user experience and brand safety risk of advertising delivery is realized. It expands the focus of the system from simply "poor effect" to "whether it causes negative impact", realizing the dimensional upgrade from pursuing "positive income" to preventing "negative loss". The decision logic is based on a simple principle: when the system receives an abnormally high intensity of negative feedback in a short time, it indicates that the current optimization strategy may have serious problems, and the expected risk of continuing to execute the strategy is much greater than the potential benefit. "Preset time" and "emergency threshold": these two parameters together define a "risk window", which ensures that the system only triggers the fuse when necessary (i.e. negative feedback is not an accidental individual case, but a trend of outbreak), avoiding false actions due to individual noise. "Immediately suspend": it reflects the high priority and real-time nature of the decision, interrupting the execution of the current harmful strategy. "Return to the initial advertising delivery parameters": this is the most critical design. The principle is that the system considers the initial delivery parameters generated as a "baseline strategy" that has been verified by semantic scene analysis, which is relatively robust and safe. When encountering unknown risks, returning to this known and relatively reliable baseline point is the least loss choice. It can quickly stop further waste of budget and continuous damage to brand image, and control the negative impact of the problem within the minimum range. It provides the fault tolerance capability essential for the fully automatic optimization system. Even if the optimization algorithm "deviates" in some extreme cases, the system can automatically "pull it back on track", ensuring the robustness and commercial reliability of the entire scheme.

[0030] Referring to Figure 3 The embodiment of the present application also provides an advertising delivery parameter adaptation system fusing scene analysis, comprising: A first acquisition module 1 is configured to acquire multi-dimensional features of a plurality of known scenes in historical advertising delivery data. The generating module 2 is configured to map the multi-dimensional features of the plurality of known scenes into high-dimensional vectors based on a preset feature embedding model, to generate a set of known scene vectors, and to construct a scene vector space. The second obtaining module 3 is configured to obtain the multi-dimensional features of the new scene in response to receiving a new advertisement launching request. The mapping module 4 is configured to map the multi-dimensional features of the new scene into a high-dimensional vector based on the same feature embedding model, to generate a new scene vector. The determining module 5 is configured to determine at least one optimal associated scene from the known scenes based on the new scene vector. The combining module 6 is configured to combine the historical advertisement launching parameters of the at least one optimal associated scene, to generate initial advertisement launching parameters of the new scene. The optimizing module 7 is configured to obtain effect feedback data of the real-time initial advertisement launching parameters, and to generate optimized advertisement launching parameters based on the initial advertisement effect feedback data.

[0031] As described above, it can be understood that each component of the advertisement launching parameter adaptation system for fusing scene analysis proposed in the present application can realize the function of any one of the advertisement launching parameter adaptation methods for fusing scene analysis as described above, and the specific structure will not be described again.

[0032] Reference Figure 4 In the embodiments of the present application, a computer device, which can be a server, is also provided. The internal structure of the computer device can be as shown in Figure 4 The computer device includes a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store monitoring data and other data. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement an advertisement launching parameter adaptation method for fusing scene analysis.

[0033] The processor executes the above-mentioned advertisement delivery parameter adaptation method based on fusion scene analysis, comprising: obtaining multi-dimensional features of a plurality of known scenes in historical advertisement delivery data; mapping the multi-dimensional features of the plurality of known scenes into high-dimensional vectors based on a preset feature embedding model to generate a known scene vector set and construct a scene vector space; in response to receiving a new advertisement delivery request, obtaining multi-dimensional features of a new scene; mapping the multi-dimensional features of the new scene into a high-dimensional vector based on the same feature embedding model to generate a new scene vector; determining at least one optimal associated scene from the known scenes according to the new scene vector; generating initial advertisement delivery parameters of the new scene in combination with historical advertisement delivery parameters of the at least one optimal associated scene; obtaining effect feedback data of the initial advertisement delivery parameters in real time, and generating optimized advertisement delivery parameters according to the initial advertisement effect feedback data.

[0034] An embodiment of the present application also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement an advertisement delivery parameter adaptation method based on fusion scene analysis, comprising the steps of: obtaining multi-dimensional features of a plurality of known scenes in historical advertisement delivery data; mapping the multi-dimensional features of the plurality of known scenes into high-dimensional vectors based on a preset feature embedding model to generate a known scene vector set and construct a scene vector space; in response to receiving a new advertisement delivery request, obtaining multi-dimensional features of a new scene; mapping the multi-dimensional features of the new scene into a high-dimensional vector based on the same feature embedding model to generate a new scene vector; determining at least one optimal associated scene from the known scenes according to the new scene vector; generating initial advertisement delivery parameters of the new scene in combination with historical advertisement delivery parameters of the at least one optimal associated scene; obtaining effect feedback data of the initial advertisement delivery parameters in real time, and generating optimized advertisement delivery parameters according to the initial advertisement effect feedback data.

[0035] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, storage, databases, or other media in this application and in examples used herein, unless specifically stated otherwise, can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), or external cache memory. As an illustration but not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0036] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, device, article, or method that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, device, article, or method. Without more limitations, an element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, device, article, or method that includes the element.

[0037] The above description is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for adapting advertising delivery parameters by integrating scenario analysis, characterized in that, The method includes: Obtain multi-dimensional features from multiple known scenarios in historical advertising delivery data; By combining a preset feature embedding model, the multi-dimensional features of the multiple known scenes are mapped into high-dimensional vectors to generate a set of known scene vectors, thereby constructing a scene vector space; In response to receiving a new ad delivery request, obtain multi-dimensional features of the new scenario; Based on the same feature embedding model, the multi-dimensional features of the new scene are mapped into high-dimensional vectors to generate a new scene vector; Based on the new scene vector, determine at least one optimal associated scene from the known scenes; By combining the historical ad delivery parameters of the at least one optimal associated scenario, the initial ad delivery parameters for the new scenario are generated; Get real-time feedback data on the performance of initial ad delivery parameters, and generate optimized ad delivery parameters based on the initial ad performance feedback data.

2. The advertising delivery parameter adaptation method based on integrated scene analysis according to claim 1, characterized in that, The step of generating optimized advertising delivery parameters based on the initial advertising performance feedback data includes: Real-time acquisition of diverse feedback data generated by the initial ad delivery parameters; A credibility analysis is performed on the multivariate feedback data, and a credibility weight is generated for each piece of feedback data based on the analysis results. Based on the multivariate feedback data assigned the credibility weight, a multidimensional attribution analysis is performed to extract the optimization factors that cause fluctuations in advertising effectiveness. The optimization factors include positive driving factors and negative inhibiting factors. The optimization factors are input into a pre-trained parameter optimization prediction model, which then predicts the optimization direction and adjustment range of the advertising placement parameters. Based on the prediction results, the optimized advertising delivery parameters are generated.

3. The advertising delivery parameter adaptation method based on integrated scenario analysis according to claim 2, characterized in that, The step of determining at least one optimal associated scene from the known scenes based on the new scene vector includes: Obtain a set of known scene vectors, and calculate the spatial distance between the new scene vector and each known scene vector in the set of known scene vectors to obtain the first similarity corresponding to each known scene; Obtain historical advertising performance data for each of the known scenarios; calculate the matching degree between the current advertising target of the new scenario and the historical advertising performance data of each of the known scenarios to obtain a second similarity corresponding to each known scenario; For each known scene, its first similarity and second similarity are input into a preset weighted fusion function for calculation, and the comprehensive relevance of the known scene is output based on the calculation result. Based on the overall relevance of all known scenarios, sort them and select one or more known scenarios with the highest ranking to determine the optimal relevance scenario.

4. The advertising delivery parameter adaptation method based on integrated scene analysis according to claim 3, characterized in that, The step of performing multi-dimensional attribution analysis based on the multivariate feedback data assigned with the credibility weights to extract the optimization factors that cause fluctuations in advertising effectiveness includes: Determine whether the total amount of feedback data collected so far has reached the effective analysis threshold; If the total amount of feedback data collected so far does not reach the effective analysis threshold, the data collection time window is extended, and the historical advertising parameters of the optimal associated scenario are weighted and calculated based on the first similarity of each known scenario to generate preliminary optimization parameters. If this has been achieved, then initiate the multidimensional attribution analysis.

5. The advertising delivery parameter adaptation method based on integrated scene analysis according to claim 3, characterized in that, Before the step of calculating the matching degree between the current advertising target of the new scenario and the historical advertising performance data of each known scenario, the method further includes: Obtain the historical competitive environment intensity index corresponding to the time when the historical data of the known scenario was generated; Obtain the current real-time competitive environment intensity index for the new scenario; Based on the difference between the historical competitive environment intensity index and the real-time competitive environment intensity index, the historical advertising performance data is adjusted by discount compensation. Based on the corrected historical advertising performance data, a matching degree is calculated between the data and the current advertising target of the new scenario.

6. The advertising delivery parameter adaptation method based on integrated scene analysis according to claim 5, characterized in that, The step of adjusting the historical advertising performance data based on the difference between the historical competitive environment intensity index and the real-time competitive environment intensity index includes: The ratio of the real-time competitive environment intensity index to the historical competitive environment intensity index is calculated to obtain the rate of change of competitive intensity. The rate of change of competition intensity is input into a preset nonlinear mapping function, and the mapping output is used to obtain the effect data discount factor; wherein, the nonlinear mapping function is configured such that when the rate of change of competition intensity is greater than 1, the output value is less than 1, and the output value decreases as the input value increases; The corrected historical advertising performance data is obtained by multiplying the historical advertising performance data of the known scenario by the performance data discount factor.

7. The advertising delivery parameter adaptation method based on integrated scene analysis according to claim 2, characterized in that, After the step of generating optimized ad delivery parameters based on the initial ad performance feedback data, the method further includes: During the advertising campaign, negative feedback signals in the advertising performance feedback data are monitored in real time. If the strength of the negative feedback signal exceeds the emergency threshold within a preset time, the use of the current advertising parameters will be immediately suspended, and the system will revert to the initial advertising parameters.

8. An advertising delivery parameter adaptation system integrating scenario analysis, characterized in that, include: The first acquisition module is used to acquire multi-dimensional features of multiple known scenarios in historical advertising data; The generation module is used to combine a preset feature embedding model to map the multi-dimensional features of the multiple known scenes into high-dimensional vectors, generate a set of known scene vectors, and construct a scene vector space. The second acquisition module is used to acquire multi-dimensional features of the new scenario in response to receiving a new ad delivery request; The mapping module is used to map the multi-dimensional features of the new scene into a high-dimensional vector based on the same feature embedding model, thereby generating a new scene vector; The determining module is configured to determine at least one optimal associated scene from the known scenes based on the new scene vector; The module is used to combine historical ad delivery parameters of the at least one optimal associated scenario to generate initial ad delivery parameters for the new scenario. The optimization module is used to obtain real-time feedback data on the performance of initial ad delivery parameters and generate optimized ad delivery parameters based on the initial ad performance feedback data.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images