Advertisement creativity dynamic evaluation and intelligent decision-making system based on multi-source data fusion

Abstract

The invention discloses an advertisement creativity dynamic evaluation and intelligent decision-making system based on multi-source data fusion, and relates to the technical field of advertisement design. Comprising a state perception and mapping module used for obtaining real-time interaction behaviors and external environment data, mapping the real-time interaction behaviors and the external environment data to a dynamic creative characteristic spectrum and outputting a spectrum structured state vector; the online decision-making module takes the state vector as input, calls an incremental learning decision-making model to calculate an expected efficiency value and a decision-making uncertainty value of each candidate creative combination, and generates a creative selection instruction through a strategy function according to the decision-making uncertainty value; the execution feedback module is used for outputting an instruction and receiving corresponding actual efficiency data; and the co-evolution updating module is used for synchronously updating weights of related nodes and edges in the dynamic creative characteristic spectrum and internal parameters of a decision model according to the efficiency data, the selection instruction and the state vector. According to the invention, real-time evaluation, intelligent decision making and collaborative self-evolution of the advertisement creativity in a dynamic putting environment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advertisement design, in particular to an advertisement creative dynamic evaluation and intelligent decision-making system based on multi-source data fusion. BACKGROUND

[0002] Intelligent evaluation and decision-making of advertisement creatives is the core development direction in the field of digital marketing, aiming to achieve accurate prediction and optimization of advertisement delivery effect through analysis and modeling of massive data, so as to improve the return on investment of marketing. In recent years, with the progress of big data and machine learning technology, advertisement delivery strategies have gradually developed towards data-driven and dynamically adjusted intelligence.

[0003] The existing technology usually builds a static creative feature knowledge base or graph, and assigns weights or scores to creative elements based on historical performance data (such as click-through rate, conversion rate). When facing new delivery targets, the system calculates the semantic similarity between target features and creative elements, or directly selects the creative combination with the highest historical score for delivery. This method can utilize historical experience to a certain extent, realize the preliminary matching of advertisement creatives and delivery targets, and has the static recommendation ability based on historical data.

[0004] However, advertisement delivery is essentially a sequential decision-making process closely interacting with dynamic market environment and real-time user feedback. The existing methods based on static knowledge base and post-hoc historical scoring have the following interrelated fundamental limitations: first, their decision basis relies on the historical effectiveness data collected and lagged in the past period, which cannot real-time perceive and respond to the rapid changes of user preferences, competitive environment or external hot events at the current moment. Secondly, its decision mechanism is essentially a "one-time" offline matching or retrieval, lacking a closed loop that can continuously learn, actively explore and iteratively optimize strategies during the delivery process. This leads to the system's rigidity when facing new user groups, new market scenarios or new creative elements, and it cannot verify hypotheses, discover potential high-quality creatives or timely avoid ineffective strategies through real-time interaction.

[0005] The existing methods are difficult to realize the adaptive optimization and long-term revenue maximization of advertisement creatives in the long-term and dynamic delivery process, especially in the rapidly changing market environment, where the stability and growth of its delivery effect face bottlenecks. SUMMARY

[0006] In view of the deficiencies of the existing technology, the present application provides an advertisement creative dynamic evaluation and intelligent decision-making system based on multi-source data fusion.

[0007] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0008] The application discloses an advertisement creative dynamic evaluation and intelligent decision system based on multi-source data fusion, comprising:

[0009] A state perception and graphing module is configured to acquire real-time interactive behavior data and external environment event data of a user, and map the data to a pre-stored dynamic creative feature graph for matching and mapping processing, and output a graph structured state vector.

[0010] An online decision module is configured to take the graph structured state vector as input, call a pre-trained incremental learning decision model, calculate expected performance values and decision uncertainty values of a plurality of candidate creative combinations sampled from the dynamic creative feature graph, and generate a creative selection instruction through a preset strategy function according to the decision uncertainty values; when the decision uncertainty value is higher than a first preset threshold, the strategy function preferentially selects a creative combination with low node weight but high associated edge strength.

[0011] An execution feedback module is configured to output the creative selection instruction to an advertisement launching execution end, and receive actual performance data corresponding to the creative selection instruction from the advertisement launching execution end.

[0012] A collaborative evolution updating module is configured to update weights of related nodes and associated edge strengths in the dynamic creative feature graph and update internal parameters of the incremental learning decision model according to the actual performance data, the corresponding creative selection instruction and the graph structured state vector.

[0013] Compared with the prior art, the application has the following beneficial effects:

[0014] 1. The application maps real-time user interactive behavior and external environment events to a dynamic creative feature graph, and generates a graph structured state vector that can reflect the current market environment and user focus of attention. This process enables the system to go beyond traditional methods that rely on single or lagging historical data, and to incorporate more accurate and comprehensive contextual information when making decisions, thereby providing an environment perception basis highly relevant to the current moment for subsequent intelligent decision-making.

[0015] 2. The application calls a pre-trained incremental learning decision model and dynamically switches strategy functions based on decision uncertainty values, achieving intelligent trade-off between exploration and utilization. In particular, when uncertainty is high, the strategy of preferentially selecting a creative combination with low node weight but high associated edge strength enables the system to actively discover and verify novel creative combinations that have insufficient historical performance data but have closely related internal elements and potential synergistic effects. This enables the system to effectively avoid falling into a local optimum and systematically broaden the path of exploring high-quality creatives during the initial launching period or when the market environment changes suddenly.

[0016] 3、The application synchronously drives the weight update of the nodes and edges in the dynamic creative feature graph and the adjustment of the internal parameters of the decision model by using the actual delivery generated performance data, and constructs a closed loop of knowledge and strategy co-evolution. The knowledge base (graph) and decision logic (model) of the system can continuously self-optimize with each delivery feedback, and they enhance each other. The ability to maintain strategy adaptability and effect stability in long-term and dynamic advertising delivery activities is improved, and the performance decay problem of static models due to environmental changes is overcome. BRIEF DESCRIPTION OF DRAWINGS

[0017] The disclosure of the application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the application. In the drawings, the same reference signs are used to refer to the same parts. Among them:

[0018] Figure 1 is a system module connection diagram of the application;

[0019] Figure 2 is a workflow step flowchart of the application;

[0020] Figure 3 is a matching mapping process flowchart of the application. DETAILED DESCRIPTION

[0021] It is easy to understand that according to the technical scheme of the application, those skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical scheme of the application, and should not be considered as the whole or as a limitation or restriction on the technical scheme of the application.

[0022] In the prior art, advertising creative optimization depends on historical data matching or offline content generation, which is difficult to cope with real-time changing market environment and user feedback. The existing system lacks a closed-loop online decision mechanism and cannot dynamically evaluate creative performance and adjust delivery strategy based on real-time interaction data. Especially in the early stage of delivery or market mutation, the existing static recommendation model cannot effectively balance the decision-making contradiction between "exploring new creative" and "using known creative", resulting in difficulty in continuously optimizing long-term delivery effect.

[0023] In order to solve the above problems, the application deeply integrates the creative element knowledge graph and the reinforcement learning decision model, constructs the state vector of the advertising environment by dynamically perceiving real-time data, and adaptively selects the creative delivery strategy based on the decision uncertainty. Further, the actual performance data after delivery is used to synchronously update the node weight and association strength of the knowledge graph, and the internal parameters of the decision model are optimized, forming a continuously evolving intelligent closed loop.

[0024] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Example:

[0026] like Figure 1 As shown, the advertising creative dynamic evaluation and intelligent decision-making system based on multi-source data fusion includes:

[0027] The state awareness and graphing module acquires real-time user interaction data and external environmental event data, maps them to a pre-stored dynamic creative feature graph for matching and mapping processing, and outputs a graph-structured state vector. The dynamic creative feature graph is essentially a graph structure stored in an in-memory database (such as Neo4j), where nodes are vectorized creative genes (e.g., visual style vectors and copywriting theme vectors extracted through deep neural networks), and edges represent the co-occurrence or semantic association strength between creative genes. Matching and mapping processing involves using natural language processing techniques to parse keywords from event data and calculate similarity with the text labels of graph nodes; simultaneously, through a rule engine or simple model, user behavior (e.g., clicking on a certain area in an advertisement) is mapped to specific creative gene nodes representing the visual elements of that area.

[0028] The initial construction of the dynamic creative feature map includes: extracting visual and textual creative elements from a historical ad creative library; extracting visual feature vectors using a pre-trained convolutional neural network (such as ResNet-50); and extracting textual semantic vectors using a Sentence-BERT model, which serve as nodes in the map. Edges between nodes are constructed based on the co-occurrence frequency of creative elements in historical ads, with the initial association strength set to a normalized value of the co-occurrence frequency. Initial node weights can be set to a normalized value of historical click-through rates, or uniformly initialized to a median value (such as 0.5). The map is stored in a graph database (such as Neo4j) and supports real-time querying and updating.

[0029] The online decision-making module takes a graph-structured state vector as input, calls a pre-trained incremental learning decision model, and calculates the expected performance value and decision uncertainty value of multiple candidate creative combinations sampled from the dynamic creative feature graph. Based on the decision uncertainty value, it generates creative selection instructions through a preset policy function. The preset policy function dynamically switches between exploration and utilization strategies based on the comparison between the decision uncertainty value and a preset threshold. When the decision uncertainty value is higher than a first preset threshold, the policy function prioritizes creative combinations with low node weights but high edge strength. The incremental learning decision model includes, but is not limited to, a context-based multi-armed gambling machine model.

[0030] The pre-training model of the incremental learning decision model is derived from: offline training using public advertisement data sets (such as Criteo, Avazu). The model weight is initialized by using the Xavier normal distribution initialization method, and the bias is initialized to zero. During the pre-training process, the input feature is the historical advertisement context vector, the label is the click / conversion signal, the mean square error loss function is used, and the Adam optimizer is used for training until convergence. The trained model parameters are loaded as the initial decision model of the system, supporting subsequent online incremental learning.

[0031] The feedback module is configured to output the creative selection instruction to the advertisement delivery execution end and receive actual performance data corresponding to the creative selection instruction from the advertisement delivery execution end;

[0032] The collaborative evolution update module is configured to update the weight of the related node and the correlation edge strength in the dynamic creative feature graph according to the actual performance data and the corresponding creative selection instruction and graph structured state vector, and update the internal parameters of the incremental learning decision model.

[0033] As shown in Figure 2 The specific working process is as follows: a random strategy or an initialization method based on historical data is used when the system is initialized. The state perception and graphing module first continuously obtains real-time data streams from the data source through stream processing technology, matches each original data with the dynamic creative feature graph pre-stored in the distributed storage module, and temporarily marks the related nodes and their associated edges in the dynamic creative feature graph memory copy according to the matching strength. In this way, the "instantaneous influence" of the current environment on the local structure of the graph is simulated. Then, the structure information (such as key node ID, temporary weight, edge strength) of the weighted and marked local subgraph is encoded into a fixed-dimensional graph structured state vector as a digital snapshot of the current advertisement delivery environment.

[0034] The online decision module receives the graph structured state vector as the model input, generates multiple candidate creative combinations in real time from the dynamic creative feature graph according to a preset sampling strategy (such as probability sampling based on node weight), and calculates two core outputs in parallel: an expected performance value (a scalar, predicting the click rate or conversion rate expectation that the creative combination can bring in the current state) and a decision uncertainty value (such as the predicted variance obtained by the Bayesian neural network, or the divergence obtained by ensemble learning). The preset strategy function makes a decision according to the uncertainty value: if the uncertainty value is higher than the first preset threshold, it means that the model is not enough to recognize the current environment, and needs to be "explored". At this time, the strategy function will filter out those creative combinations with low average node weight (representing mediocre or novel historical performance) but high average strength of internal node correlation (representing good internal element synergy and stable structure of the combination), and preferentially select one of them. Finally, the structured creative selection instruction is output, including the identifier and delivery parameters of the target combination.

[0035] The execution feedback module delivers the creative selection instruction to the advertisement delivery execution end (such as an advertisement transaction platform) in quasi-real time. After a delivery cycle (such as a few hours) ends, the execution feedback module receives the aggregated actual performance data (such as exposure, clicks, conversion cost, etc.) corresponding to this instruction from the advertisement delivery execution end.

[0036] The collaborative evolution update module receives the actual performance data from the feedback module, the corresponding historical selection instruction, and the graph structured state vector used for decision-making. Its update process is synchronous and collaborative, including:

[0037] 1. Update the dynamic creative feature graph: the system identifies the specific creative gene nodes involved in the selection instruction. According to the comparison of actual performance data and preset benchmark, a reward signal is generated. For the dynamic creative feature graph, the reward signal is used to adjust the long-term weight of the relevant nodes (for example, the weight value is updated using a time-decaying weighted average algorithm) and the strength of the correlation edges between nodes (for example, if the current delivery effect is excellent, the strength of all edges within the current combination is enhanced).

[0038] 2. Update the internal parameters of the incremental learning decision model: at the same time, the same reward signal and historical state vector together constitute a training sample, which is used to fine-tune the incremental learning model in the online decision module for one round (for example, update the neural network weight by stochastic gradient descent), so that its future prediction is more accurate.

[0039] This application enables online, autonomous, and continuous optimization of advertising creative delivery strategies. It can dynamically adapt to changes in the market environment and user preferences, intelligently balancing creative exploration and utilization, thereby significantly improving long-term advertising performance (such as overall return on investment (ROI)). This application not only outputs decision results, but its internal knowledge representation and decision logic also continuously evolve with data accumulation, effectively solving the cold start problem and model drift problem, and enhancing the system's robustness and adaptability in complex and ever-changing delivery environments.

[0040] like Figure 3 The diagram shown is a flowchart of the matching mapping process of the present invention. This application further proposes that the specific steps for obtaining real-time user interaction behavior data and external environmental event data, and mapping them to a pre-stored dynamic creative feature map for matching mapping processing, include:

[0041] Obtain the pre-stored dynamic creative feature map, which is constructed with creative gene vectors as nodes. The visual element gene vectors (such as the fc7 layer features extracted by the VGG16 network) are associated with the standardized spatial coordinate region of the advertising canvas.

[0042] When inputted from real-time interactive behavior data (e.g., generated by a front-end eye-tracking SDK or mouse trajectory heatmap), the system first parses and generates a heatmap of the user's gaze dwell time in the advertising canvas area. This heatmap is a two-dimensional matrix matched to the canvas resolution, where the value of each element represents the cumulative gaze duration or interaction event density of the corresponding pixel area within a preset time window (e.g., the most recent 10 seconds).

[0043] The threshold segmentation algorithm in image processing is used to extract the coordinate sequence of regions whose intensity exceeds a preset intensity threshold from the heatmap data of the gaze-attention heatmap. This preset intensity threshold can be set through historical data analysis, for example, by taking the 75th percentile of the intensity values ​​in all users' historical heatmap data, with an empirical range of (0.3, 0.7) (after normalization).

[0044] The system matches and associates the region coordinate sequences with the creative gene vectors representing visual elements stored in the dynamic creative feature map. The matching process is achieved by calculating the intersection-over-union (IoU) ratio between the coordinate sequences and the preset coordinate regions of the creative gene vectors. A successful association is determined when the IoU exceeds a matching threshold (e.g., 0.5). A coordinate mapping table is then used to map the heatmap regions to preset visual gene nodes. This mapping table is constructed based on the relationship between the advertising canvas resolution and the gene vector coordinates.

[0045] Based on the matching and association results, temporary enhancement weight coefficients are generated for the relevant nodes and associated edges, and these enhancement weight coefficients are encoded into the graph structured state vector. The calculation formula is:

[0046]

[0047] in, To match the total area of ​​the high-attention region, This represents the total area of ​​the advertising canvas. Characterizes attention concentration; This represents the average length of stay in the area. The baseline dwell time (e.g., 1 second); This is a modulating factor used to control the overall scale of the enhancement magnitude. ∈[0.1,0.5];log( The function is used to smooth out the effect of duration.

[0048] The system will calculate the temporary enhancement weight coefficients. Together with the corresponding node identifiers and associated edge identifiers, they are encoded into a specific dimension range of the graph structured state vector.

[0049] The construction and incremental updating of the dynamic creative feature map is a continuous learning process. The initial creative feature map can be trained from a historical ad creative library. Textual features are extracted using models such as Word2Vec and Sentence-BERT, and visual features are extracted using pre-trained convolutional neural networks (such as ResNet-50). The feature vectors are used as nodes. The edges between nodes and the initial association strength can be calculated by the co-occurrence frequency of creative elements in historical ads, or obtained by training a graph neural network on ad performance data. The online incremental update of the map is driven by a co-evolutionary update module. It uses actual campaign performance data (such as click-through rate, CTR) as a reward signal, updates node weights through a time-decay-based weighted average algorithm, and adjusts the strength of association edges through an optimization method similar to stochastic gradient descent, ensuring that the map knowledge continues to evolve with the campaign process.

[0050] This application enables online decision-making models to explicitly consider the spatial distribution of real-time user attention when evaluating creative combinations, thus favoring creative combinations that include currently highly attention-grabbing visual elements and improving the accuracy of matching advertising creatives with users' immediate interests. It also ensures a balance between capturing fleeting trends and maintaining long-term stable knowledge.

[0051] like Figure 3 The diagram shown is a flowchart of the matching mapping process of the present invention. This application further proposes that the specific steps for obtaining real-time user interaction behavior data and external environmental event data, and mapping them to a pre-stored dynamic creative feature map for matching mapping processing, include:

[0052] Extract keywords or sentiment labels from external environment event data. Specifically, the system continuously fetches external environment event data streams from preset news aggregation APIs, social media streams, or search engine trends. This data stream usually contains text headlines, summaries, and publication timestamps. The event text is processed by using pre-trained natural language processing models. For example, a BERT-based sequence labeling model can be used to extract key entities and noun phrases as keywords, while a RoBERTa-based sentiment analysis model is used to output the overall sentiment label of the text (such as positive, negative, neutral, and its confidence).

[0053] Match the keywords or sentiment labels with the pre-defined creative gene vectors in the dynamic creative feature graph that represent the text theme or sentiment. The creative gene vector is usually obtained by encoding and clustering a large number of advertising copy and brand promotion text using a sentence encoding model such as Sentence-BERT, and each vector is associated with one or more semantic labels. For keywords, the system encodes them into word vectors (such as using Word2Vec or directly using Sentence-BERT to encode short sentences), and then calculates their cosine similarity with each text creative gene vector. For sentiment labels, they are compared for consistency with the sentiment attributes associated with the graph nodes.

[0054] The system sets a preset matching threshold (for example, cosine similarity greater than 0.7, or sentiment consistency is positive / negative and confidence greater than 0.8) to determine the significance of the association. The range of the preset matching threshold can be set based on the similarity distribution of effective associations in historical data, and the experience value is usually between 0.65 and 0.75.

[0055] Nodes in the dynamic creative feature graph that exceed the preset matching threshold are marked as nodes affected by the current external event, and their influence factors are encoded into the graph structured state vector.

[0056] In order to quantify the degree of influence, the influence factor The calculation can take into account the matching degree score and the propagation heat of the event itself (such as the frequency of mentioning the event on social media). In one embodiment, the calculation formula is as follows:

[0057]

[0058] where, is the normalized semantic matching score (for example, cosine similarity); is the number of mentions or heat index of the event in the recent time window; is the baseline heat value for standardization; is the weight coefficient, ∈ [0, 1], for example, take 0.6, to balance the influence of semantic relevance and event heat.

[0059] Finally, the system encodes the identifiers of these nodes and their corresponding influence factors into another specific dimension interval of the atlas structured state vector.

[0060] The present application dynamically associates the semantic information of external environmental events with the creative feature atlas, and quantifies it as an influence factor encoded into the state vector, so that the online decision module can perceive and respond to changes in the macro social environment. The decision model can therefore tend to select creative combinations that are semantically related to current hot topics or consistent with the mainstream emotional tone, thereby improving the environmental adaptability and timeliness of the advertising content. This mechanism enables the advertising placement strategy to take advantage of social hotspots, avoid placing inappropriate advertising content in a negative public opinion environment, and enhance brand safety and communication effectiveness.

[0061] The present application further proposes that the called incremental learning decision model is preferably a contextual multi-armed bandit model, the technical principle of which is to formalize the advertising creative dynamic decision problem as a sequential decision task: at each decision time slot (for example, every half hour or for each batch of new user traffic), the system needs to select one from a set of candidate creatives (arms) for placement, and evaluate the selection according to the user feedback (reward) obtained immediately or later, the goal is to maximize the long-term cumulative reward (such as total clicks or conversion value). The contextual multi-armed bandit model specifically includes:

[0062] The "arms" of the contextual multi-armed bandit model are not fixed, but are dynamically generated at each decision time slot, and each arm of the contextual multi-armed bandit model strictly corresponds to a candidate creative combination sampled in real time from the dynamic creative feature atlas. The candidate creative combination is sampled from the dynamic creative feature atlas according to node weights (representing historical effectiveness), so that high-weight nodes have a higher probability of being selected; at the same time, the associated edge strength is generated, which tends to combine nodes with high associated edge strength in the graph to take advantage of the synergistic effect between creative elements.

[0063] For each decision time slot when constructing the combination, the context input of the contextual multi-armed bandit model is the atlas structured state vector. This state vector integrates real-time user attention, external event influence, etc., providing the model with a feature representation of the current decision environment.

[0064] When the model is initialized or faces a brand new creative combination (arm), it needs to set the initial expectation of its return. The parameters of the return prior distribution of each arm in the contextual multi-armed bandit model are initialized based on the current weights of the creative gene vector nodes corresponding to the current arm in the dynamic creative feature graph and the strength of the associated edges. Specifically, for any new arm (corresponding to a set of creative gene nodes), the model will calculate the initial return expectation of the arm based on the current weights of the nodes in the graph (representing the historical average performance of the element) and the average strength of the associated edges between the nodes in the combination (representing the internal synergy of the combination). For example, the mean of the return prior distribution is initialized as It can be set as follows:

[0065]

[0066] wherein is the average weight of the nodes included in the arm (normalized to the [0, 1] interval), is the average strength of all related edges between the nodes in the graph (normalized to the [0, 1] interval). is a weighting coefficient, for example, set to 0.7, indicating that more reliance is placed on the individual historical performance of the nodes.

[0067] The variance of the return distribution can be set as a quantity inversely proportional to , reflecting the uncertainty brought about by the sufficiency of historical data.

[0068] The training of the contextual multi-armed bandit model is a continuous incremental learning process. The training conditions include: the learning rate is usually set to a small value (such as 0.01 to 0.1) to adapt to the non-stationarity of the data; the neural network structure is a 3-layer fully connected layer, the number of training rounds is 1000, the optimizer is Adam, the public advertising data set (such as Avazu, Criteo) is used for pre-training, 5-fold cross-validation is used, and strategies such as Thompson Sampling or UCB (Upper Confidence Bound) are used for exploration and utilization balance. When a creative combination is selected and launched, the actual performance data (such as click-through rate) collected is converted into a scalar reward, which, together with the context vector input at the time, forms a training sample. The model then updates the internal parameters (for example, updates the weight vector in the linear context bandit or updates the posterior distribution in the Bayesian linear regression) based on this sample, thereby realizing the rapid absorption of new knowledge and the continuous improvement of future predictions.

[0069] By using the contextual multi-armed bandit model and innovatively using dynamic graph knowledge for arm initialization and sampling, the decision mechanism can effectively cope with the high dynamics and uncertainty of the advertising environment. Not only can it make conditional decisions using rich contextual information, but its incremental learning characteristics also ensure that the strategy can quickly adapt to real-time feedback, enabling the system to establish a probabilistic model-based, quantifiable balance between exploring new ideas and utilizing mature ideas, thereby achieving more stable and optimal cumulative advertising effects in the long run.

[0070] The application further proposes that, according to the decision uncertainty value, the specific steps of generating the creative selection instruction through the preset strategy function include presetting two key thresholds:

[0071] The first preset threshold and the second preset threshold , and satisfying These two thresholds can be set according to the prediction confidence distribution of the model on the historical verification set. For example, after sorting the decision uncertainty values, the 80th percentile is taken as the first preset threshold , the 20th percentile is taken as the second preset threshold , and the value range is (0.3, 0.7).

[0072] When the decision uncertainty value is higher than the first preset threshold, the strategy function adopts an exploration strategy, which includes: calculating the average weight of the creative gene vector nodes contained in each candidate creative combination The weight of the node is a reflection of its long-term effectiveness in the dynamic graph, and a high weight represents stable and excellent historical performance. The goal of exploration is to discover potential high-quality creatives, so the system sets a preset weight threshold (e.g., taking the median or mean of all node weights). The average weight is lower than the preset weight threshold (i.e. ); secondly, the system sets a preset correlation strength threshold (e.g., taking the 75th percentile of all edge strength values), and further calculates the average strength of the associated edges within the combination of the candidate combination whose average weight meets the weight condition is higher than the preset correlation strength threshold (i.e. ).

[0073] When the decision uncertainty value is lower than the second preset threshold , it indicates that the model is highly confident in the current environment. At this time, the candidate creative combination with the highest expected performance value is selected to obtain the most predictable maximum immediate return.

[0074] The application enables the adaptive switching of the decision-making behavior of the system according to the model's own cognitive state (uncertainty) through the design of the strategy function. The efficiency and systematicness of the novel and creative value discovery are improved. The system can continuously maintain the adaptive balance between exploration and utilization in a changing environment, thereby effectively avoiding the loss of short-term benefits due to excessive exploration or the aging of creativity and the rigidification of strategies due to excessive utilization, providing a methodological guarantee for the long-term and stable improvement of advertising effectiveness.

[0075] The application further proposes that the specific steps of synchronously updating the weights of the related nodes and the strengths of the associated edges in the dynamic creative feature graph include:

[0076] After receiving the actual performance data (for example, the click-through rate CTR), the collaborative evolution updating module first compares the actual performance data with the preset benchmark performance value. According to the comparison result of the actual performance data and the preset benchmark performance value, a weight adjustment coefficient is generated. The preset benchmark performance value can be the historical average performance of the advertising position or similar products, the industry benchmark value, or a dynamic moving average value. The weight adjustment coefficient can be defined as the ratio of the actual performance to the benchmark performance minus 1 , or a piecewise function mapping is used so that a positive adjustment is obtained when the performance is excellent (such as more than 20% of the benchmark), and a negative adjustment is obtained when the performance is poor.

[0077] The system then identifies the creative gene vector nodes involved in the creative selection instruction, that is, the graph nodes corresponding to all the creative elements contained in the advertised advertisement.

[0078] The current weight value of the creative gene vector node is calculated using a time-decay-based weighted update algorithm combined with the weight adjustment coefficient to obtain the updated node weight value, which is written back to the dynamic creative feature graph. The system applies a time-decay-based weighted update algorithm to calculate the updated node weight The formula is as follows:

[0079]

[0080] wherein, is the node weight before updating; is the weight adjustment coefficient generated this time; is a time decay function, whose value decreases as the interval between the last weight update time of the node and the current time increases. For example, wherein is a decay rate parameter, for example, set to 0.1, indicating that the influence is halved approximately every 10 time units; It is the learning rate, which controls the size of the step size in a single update. It is set through grid search or experience, and is usually set to a small positive number (such as between 0.05 and 0.2). Is with The adaptive factor of linkage, among which, .

[0081] Using this formula, a long-unupdated ( The weight of nodes with smaller values ​​is affected by the new evidence. The impact is relatively small, mainly maintaining historical values; while a value that has been frequently updated recently ( Nodes with weights close to 1 have weights that more sensitively reflect recent performance trends. Finally, the updated node weights are calculated. It is written back to the dynamic creative feature map, replacing the original weight values.

[0082] Employing a time-decay-based weighted update algorithm, the weight evolution of nodes in the dynamic creative feature map possesses both memory and adaptability. This allows node weights to smoothly track their long-term performance trends, exhibiting robustness to short-term noisy data while gradually incorporating new and valid evidence, adapting to changes in the value of creative elements over time and in the environment. It avoids the update lag or excessive fluctuations that may arise from simple averaging or fixed-window statistics in traditional methods, thus improving overall stability in non-stationary environments.

[0083] This application further proposes that the specific steps for synchronously updating the weights of relevant nodes and the strength of associated edges in the dynamic creative feature graph also include:

[0084] When evaluating the effectiveness of a campaign, the collaborative evolution update module first performs a binary judgment based on actual performance data to determine whether the campaign was effective in the current decision-making time slot. This judgment is based on comparing key performance indicators (such as click-through rate and conversion rate) with a dynamic effectiveness threshold. The effectiveness threshold can be a moving average of recent historical performance data or a fixed benchmark set based on business objectives (e.g., click-through rate > 0.5%). If the actual performance exceeds this effectiveness threshold, the campaign is considered effective.

[0085] If deemed valid, the system performs edge enhancement. This enhances the strength of all associated edges between the creative gene vector nodes involved in the creative selection instruction. The enhancement magnitude is then determined. This can be positively correlated with the relative excellence of this campaign. Where E represents the performance index ( These are actual performance metrics (such as click-through rate and conversion rate). (This is a validity threshold used to determine whether the campaign is effective.) This is the enhancement coefficient, usually set to a small positive number (such as between 0.01 and 0.1) to ensure a smooth increase in edge strength. The formula for updating the strength value of associated edges is: Ensure that the strength value does not exceed the preset upper limit (usually 1).

[0086] If an ad is deemed invalid, and the graph structured state vector contains explicit negative user feedback information (e.g., the cursor position when the ad is closed, or the user's selected "not interested" reason tag), the system performs a more refined edge weakening operation. Based on the negative user feedback information recorded in the graph structured state vector, negative feedback regions are determined. For example, if a user frequently closes an ad near a specific visual element area (such as an image of a spokesperson or a certain background color), that area is marked. Subsequently, the strength of specific associated edges between creative gene vector nodes associated with the negative feedback region is weakened. The degree of weakening is determined accordingly. A more aggressive approach could be taken to accelerate the elimination of unfavorable portfolio patterns and reduce the magnitude of the impact. The formula is: ,in This is a mitigation factor, ranging from 0.1 to 0.3. The updated formula is: , This represents the edge strength value before the weakening. This approach avoids indiscriminately weakening all edges when there is invalid delivery, and instead applies the penalty precisely to the elements directly related to the negative feedback.

[0087] This application enables dynamic creative feature maps to go beyond simple co-occurrence statistics, improving the intrinsic quality of subsequent candidate creative combinations generated by the system. It guides the online decision-making module to be more inclined to select creative combinations with strong internal synergy and avoid elements that are known to be rejected by users, thereby reducing the probability of ineffective delivery at the source and achieving continuous optimization of creative combination strategies.

[0088] This application further proposes to generate creative selection instructions through a preset strategy function; wherein, the creative selection instructions are creative gene vector recombination instructions, and the specific implementation steps of the generation logic are as follows:

[0089] After the online decision-making module completes the evaluation of a batch of candidate creative combinations, the preset strategy function performs a diagnostic round before executing the regular selection. This identifies candidate creative combinations whose current expected performance value is lower than a preset score threshold. (Preset score threshold) Based on the distribution of historical evaluation scores, for example, the median or 40th percentile of the expected performance values ​​of all historical candidate combinations is taken, with a typical normalization range between 0.4 and 0.6.

[0090] The system calculates the contribution (or sensitivity) of each creative gene vector node within the combination to the expected performance value of the combination. The node with the lowest contribution is identified as the node to be replaced. From the dynamic creative feature map, at least one alternative node with the highest correlation edge strength to the node to be replaced and a higher node weight is searched. The search criteria are based on two points: first, the correlation edge strength between the alternative node and the original low-contribution node must be the highest, ensuring that the alternative element is highly related to the original element in creative semantics or visual style, and the replacement will not cause a break in the combination logic; second, the node weight of the alternative node must be higher, ensuring that the replacement evolves in the direction of better historical performance. The system selects at least one alternative node that meets the above conditions and has the highest weight.

[0091] The system generates a reorganization instruction to replace the original low-contribution node with a replacement node. Subsequently, the reorganization instruction and the resulting new candidate idea combination are returned to the online decision-making module for a new round of performance evaluation. Upon receiving this new combination, the online decision-making module immediately performs a new round of performance evaluation based on the current graph structured state vector, calculating its new expected performance value and decision uncertainty value. This newly generated combination will participate in the subsequent decision-making process of the current time slot (e.g., it may be directly selected due to its improved performance, or it may enter the next round of evaluation).

[0092] By employing a reorganization instruction mechanism, this application achieves a leap from passive selection to proactive construction. When faced with initially underperforming creative solutions, the system does not simply abandon them or passively wait for future exploration. Instead, it proactively initiates a precise and directed improvement attempt based on rich graph knowledge (association strength and node weights). This mechanism enhances the flexibility and intelligence of creative combination generation, improves the efficiency of creative resource utilization, and raises the output quality of the decision-making process.

[0093] This application further proposes that the system also includes a policy embedding module, used for:

[0094] The decision data output by the online decision-making module is monitored. This data includes at least the creative selection instruction and its corresponding decision certainty value. When the graph-structured state vector for the same feature pattern is detected, and the same creative selection instruction is output in a predetermined number of consecutive decisions, with the corresponding decision certainty value consistently exceeding a preset confidence threshold, the monitoring process is successful. At this point, it is determined that the decision-making strategy for that feature pattern has stabilized. Here, "same feature pattern" does not refer to completely identical state vectors, but rather to state vectors falling into a specific cluster or region within a feature space. For example, the system can maintain a dynamic clustering model (such as using incremental K-Means or DBSCAN algorithms) to perform online clustering of continuously generated graph-structured state vectors. When a new state vector arrives, it is assigned to a specific cluster ID, which represents a type of feature pattern.

[0095] The module maintains a decision history record for each identified feature pattern (cluster). When it detects that for the same cluster ID, in a preset number of consecutive decisions (e.g., N=20), the online decision module outputs the same creative selection instruction, and the decision certainty value corresponding to each decision is higher than a preset confidence threshold, the module ensures that the decision is correct. (For example, When the threshold (which can be set according to the business's requirements for decision reliability, typically between 0.85 and 0.95) is reached, the module determines that the decision strategy for that feature pattern has reached a stable state.

[0096] Once stability is determined, the module generates and stores fixed decision rules. These fixed decision rules essentially establish a direct mapping from feature patterns (described by cluster center vectors or cluster boundaries) to creative selection instructions. The rules are stored in a high-speed query rule base. When a new graph-structured state vector is generated, the system first quickly matches it with the feature patterns in the rule base (e.g., calculating the cosine similarity with each cluster center; a match is considered successful if the similarity exceeds a threshold such as 0.95). If a match is successful, the system directly calls the corresponding fixed decision rule to generate creative selection instructions, bypassing the complete model calculation process of the online decision module, thus significantly shortening the decision-making process.

[0097] The introduction of the strategy solidification module enables the system to transform well-validated, high-performance online learning results into efficient execution rules. This reduces the computational load and decision latency when dealing with familiar scenarios, improving overall throughput and response speed. Solidified rules, as carriers of stable knowledge, prevent decision-making fluctuations that may occur in online models due to short-term data volatility, enhancing the certainty and consistency of system decisions in mature scenarios. This allows complex computing resources to be focused on handling novel and uncertain decision-making scenarios, thereby optimizing the system's overall performance and energy efficiency.

[0098] This application further proposes that the system also includes a constraint generation module, used for:

[0099] After the online decision-making module generates an initial set of candidate creative combinations, a graph structure homogeneity detection is performed on the sampled initial candidate creative combinations. The core of the detection is to calculate the graph structure similarity between any two creative combinations. The structure of a creative combination can be characterized by the set of nodes it contains and the set of edges connecting the nodes.

[0100] If a homogeneous subset of graph structures has a similarity higher than a preset similarity threshold, then a representative subset is retained based on the average node weight. Here, similarity... It can be calculated by combining the Jaccard coefficient (which measures the overlap between sets of nodes) and the similarity between sets of edges. ,in, These are the Jaccard coefficients, where N and E represent the node set and edge set, respectively. These are the weighting coefficients. ∈[0,1], for example, set to 0.6. The system sets a preset similarity threshold. (e.g., 0.75), preset similarity threshold The value typically ranges from 0.7 to 0.85 and can be adjusted according to the desired level of diversity.

[0101] For each identified homogeneous subset, the module performs a redundancy removal operation: based on the average node weights, a representative subset is retained. Specifically, the average weight of the nodes in each subset is calculated, and the subset with the highest average weight is retained as the representative, while the remaining subsets are temporarily removed. The purpose of this is to prioritize retaining historically best-performing individuals from the homogeneous options.

[0102] Next, based on the dynamic creative feature map, new alternative creative combinations are generated through resampling, ensuring that the structural differences between the new combination and all currently retained combinations meet a preset diversity requirement. The goal of generating new alternative creative combinations is to ensure sufficient structural differences between them and all currently retained combinations. This difference is assessed by calculating the structural similarity between the new combination and each existing retained combination, and taking the maximum similarity. The system sets a preset diversity requirement: the maximum similarity between the new combination and all existing combinations must be below a diversity threshold. (e.g., 0.4). The module generates combinations that meet this requirement through an iterative sampling process: sampling nodes from the graph (nodes that are intentionally far away from existing retained combinations in the graph can be selected), constructing combinations, and checking whether they meet the diversity threshold. The requirement is to continue until a sufficient number of new alternative combinations are generated.

[0103] Finally, the module merges the representative combinations selected after redundancy removal with the new alternative creative combinations generated by resampling to form a set of candidate creative combinations after redundancy removal. This set is then input into the online decision-making module for subsequent performance and uncertainty evaluation.

[0104] By introducing a constraint generation module, this application avoids wasting limited computational and evaluation resources on evaluating highly similar creative combinations. It improves the efficiency of exploring a broad creative space by not only exploring strategies (through uncertainty) but also actively managing diversity at the source of candidate set generation. This makes the exploration more comprehensive and systematic, helping to discover more potential high-value creative types in long-term operation and reducing the risk of getting trapped in local optima due to sampling bias.

[0105] The following is a specific example of an advertising creative dynamic evaluation and intelligent decision-making system based on multi-source data fusion:

[0106] During a major promotional period, an e-commerce platform launched advertising campaigns for a new athletic shoe product, utilizing this system for dynamic creative evaluation and decision-making. The system is deployed on an Alibaba Cloud GPU-accelerated computing cluster, with dynamic creative feature maps stored in the Neo4j graph database. Real-time data is transmitted via a Kafka message queue (10 broker nodes).

[0107] The state awareness and graphing module first acquires real-time user interaction data: It collects mouse click coordinates and page dwell time on the advertising canvas using a front-end tracking SDK (sampling frequency 50Hz), and parses the gaze dwell heatmap data (two-dimensional matrix resolution 1920×1080). The system uses the Otsu threshold segmentation algorithm to extract region coordinate sequences (e.g., (0.3, 0.4), (0.5, 0.6)) with intensity exceeding a preset threshold (0.6), and performs IoU matching (threshold 0.5) with creative gene vectors representing visual elements in the dynamic creative feature graph (e.g., features extracted from the fc7 layer of the VGG16 network). After successful matching, temporary enhancement weight coefficients are generated. in, =0.2 (the proportion of the high-attention area to the total canvas area). (Average length of stay) Seconds (base duration), α=0.3 (adjustment factor).

[0108] Simultaneously, the system extracts keywords "618 promotion" and "sneaker discount" from external environmental event data (news aggregation API), performs cosine similarity matching (threshold 0.7) with the text creative gene vector in the graph, and generates an impact factor: in, =0.85 (match score) =1200 (Event popularity) =1000 (baseline popularity), β=0.6 (weighting coefficient).

[0109] The system will and Encoded as a graph-structured state vector (128 dimensions).

[0110] The online decision-making module calls the contextual multi-armed gambling machine model (3-layer fully connected network, 1000 training epochs, Adam optimizer) to sample candidate creative combinations (arms) from the graph, using the state vector as input. The model calculates the expected performance value and decision uncertainty value of each arm: when the uncertainty value (0.65) is higher than the first preset threshold (0.5), an exploration strategy is adopted to prioritize creative combinations with an average node weight (0.4) lower than the preset weight threshold (0.5) and an average edge strength (0.8) higher than the preset association strength threshold (0.7) (such as "red shoe body + 'limited-time discount' copy").

[0111] The execution feedback module sends creative selection instructions to the ad execution terminal and receives actual performance data (such as click-through rate and conversion cost) after the campaign period ends. The collaborative evolution update module generates weight adjustment coefficients based on the actual performance data (click-through rate 0.8%, baseline click-through rate 0.6%). The node weights are updated using a time-decay weighted update algorithm: ,in, =0.4 (original weight) (Time decay function, λ=0.1) =0.1 (learning rate) =0.09 (adaptive factor). Simultaneously, the strength of the inter-node connections within this creative combination is enhanced (original strength 0.8, enhancement coefficient 0.05): .

[0112] The constraint generation module performs homogenization detection on the initial candidate creative combinations (similarity threshold 0.75), retains the combination with the highest average node weight, and resamples to generate new combinations with a difference that meets the diversity threshold (0.4) (such as "blue shoe body + 'celebrity style' copywriting"). The strategy solidification module detects that the decision for the feature pattern of "young users + promotional scenario" outputs the same instruction 20 times consecutively, and the deterministic value (0.92) is higher than the preset confidence threshold (0.9), and generates a solidified decision rule: when the state vector matches this feature pattern, directly select the combination of "red shoe body + 'limited-time discount' copywriting" to bypass the model calculation.

[0113] Technical Results: During major promotional periods, the system responds in real-time to changes in user attention (e.g., users pay more attention to the "limited-time discount" copy area), proactively exploring novel creative combinations (e.g., "celebrity endorsement" copy), avoiding the pitfall of "only recommending historically high-click-through-rate creatives" as the local optimum. Through collaborative evolution and updates, the weight of the "limited-time discount" copy node in the dynamic creative feature graph is gradually optimized based on campaign feedback, making the decision model more accurate in predicting the creative preferences of younger user groups. After the strategy is solidified, the decision latency is reduced from 500ms to 100ms, system throughput is increased by 30%, and advertising performance remains stable in complex and dynamic environments, with user click-through rate and conversion rate continuously optimizing over the campaign cycle.

[0114] This embodiment demonstrates the complete application process of the system in e-commerce promotion scenarios. Through multi-source data fusion, dynamic decision-making, and collaborative evolution, it solves the technical problems of traditional advertising systems being unable to respond to market changes in real time and struggling to balance exploration and utilization, and achieves adaptive optimization and continuous evolution of advertising creative placement strategies.

[0115] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An advertisement creative dynamic evaluation and intelligent decision system based on multi-source data fusion, characterized in that: Comprise: State perception and mapping module, for obtaining real-time interaction behavior data and external environment event data of the user, and mapping to the pre-stored dynamic creative feature graph for matching mapping processing, outputting the graph structured state vector; Online decision module, for taking the graph structured state vector as input, calling a pre-trained incremental learning decision model, calculating the expected performance value and decision uncertainty value of a plurality of candidate creative combinations sampled from the dynamic creative feature graph; and generating a creative selection instruction through a preset strategy function according to the decision uncertainty value; Execution feedback module, for outputting the creative selection instruction to the advertisement launching execution end, and receiving actual performance data corresponding to the creative selection instruction from the advertisement launching execution end; Co-evolution update module, for synchronously updating the weight of the related node and the correlation edge strength in the dynamic creative feature graph according to the actual performance data, the corresponding creative selection instruction and the graph structured state vector, and updating the internal parameters of the incremental learning decision model; Wherein, the preset strategy function dynamically switches the exploration strategy and the utilization strategy according to the comparison result of the decision uncertainty value and the preset threshold value; when the decision uncertainty value is higher than the first preset threshold value, the exploration strategy is adopted, and the creative combination with low node weight but high correlation edge strength is preferentially selected. 2.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system according to claim 1, characterized in that: The specific steps of obtaining real-time interaction behavior data and external environment event data of the user, and mapping to the pre-stored dynamic creative feature graph for matching mapping processing include: Obtaining the dynamic creative feature graph; wherein the dynamic creative feature graph is constructed with creative gene vectors as nodes; Parsing the user's eye-stopping heat map data in the advertisement canvas area from the real-time interaction behavior data; Extracting the region coordinate sequence with intensity exceeding the preset intensity threshold from the eye-stopping heat map data; Matching and associating the region coordinate sequence with the creative gene vector representing the visual element; According to the matching and association result, generating temporary enhanced weight coefficients of the related nodes and the correlation edges, and encoding the enhanced weight coefficients into the graph structured state vector. 3.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system according to claim 2, characterized in that: The specific steps of obtaining real-time interaction behavior data and external environment event data of the user, and mapping to the pre-stored dynamic creative feature graph for matching mapping processing also include: Extracting keywords or sentiment labels from the external environment event data; Performing semantic similarity matching of the keywords or sentiment labels with the creative gene vectors representing the text theme or emotion; Marking the nodes of the dynamic creative feature graph with matching degree exceeding the preset matching threshold as nodes affected by the current external event, and encoding the node influence factor as the graph structured state vector. 4.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system of claim 1, wherein: The called incremental learning decision model is a context multi-armed bandit model, and the context multi-armed bandit model specifically includes: Each arm of the context multi-armed bandit model corresponds to a candidate creative combination, which is generated from the dynamic creative feature graph according to the node weight and the correlation edge strength; The context input of the context multi-armed bandit model is the graph structured state vector; The return prior distribution parameters of each arm of the context multi-arm bandit model are initialized based on the current weight of the creative gene vector node corresponding to the current arm in the dynamic creative feature map and the associated edge strength.

5. The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system of claim 4, wherein: According to the decision uncertainty value, the specific steps of generating a creative selection instruction by a preset strategy function include: When the decision uncertainty value is higher than the first preset threshold, the exploration strategy is adopted; wherein the exploration strategy includes: calculating the average weight of the creative gene vector nodes in each candidate creative combination, and preferentially selecting a candidate creative combination whose average weight is lower than a preset weight threshold and whose average strength of the associated edges between the nodes in the combination is higher than a preset associated strength threshold; When the decision uncertainty value is lower than the second preset threshold, the utilization strategy is adopted to select the candidate creative combination with the highest expected performance value; wherein the second preset threshold is lower than the first preset threshold. 6.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system according to claim 1, characterized in that: The specific steps of synchronously updating the weights of the related nodes and the associated edge strengths in the dynamic creative feature map include: According to the comparison result of the actual performance data and the preset benchmark performance value, a weight adjustment coefficient is generated; The creative gene vector nodes involved in the creative selection instruction are identified; The current weight value of the creative gene vector node is calculated by using a time-decay-based weighted update algorithm combined with the weight adjustment coefficient to obtain an updated node weight value, which is written back to the dynamic creative feature map.

7. The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system of claim 1, wherein: The specific steps of synchronously updating the weights of the related nodes and the associated edge strengths in the dynamic creative feature map further include: According to the actual performance data, it is judged whether the delivery at the current decision time slot is effective; If it is determined to be effective, the strength values of all associated edges between the creative gene vector nodes involved in the creative selection instruction are enhanced; If it is determined to be ineffective, a negative feedback region is determined according to the user negative feedback information recorded in the graph structure state vector, and then the strength values of specific associated edges between the creative gene vector nodes associated with the negative feedback region are weakened. 8.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system of claim 1, wherein: A creative selection instruction is generated by a preset strategy function; wherein the creative selection instruction is a creative gene vector recombination instruction, and the generation logic includes: A candidate creative combination whose current expected performance value is lower than a preset score threshold is identified; For the candidate creative combination, the creative gene vector node with the lowest contribution degree to the expected performance value is identified; At least one replacement node with the highest associated edge strength and higher node weight than the node with the lowest contribution degree is found from the dynamic creative feature map; A recombination instruction replacing the original low-contribution node with the replacement node is generated; The recombination instruction and the new candidate creative combination formed after recombination are returned to the online decision module for a new round of performance evaluation. 9.The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system according to claim 1, characterized in that: The system further includes a strategy solidification module for: Monitoring the decision data output by the online decision module, the decision data at least including the creative selection instruction and the corresponding decision certainty value; When the graph structured state vector for the same feature mode is monitored, the same creative selection instruction is output in a continuous preset number of decisions, and the corresponding decision certainty value is higher than the preset confidence threshold, it is determined that the decision strategy for the feature mode has stabilized; A solidified decision rule is generated and stored, when the graph structured state vector matches the feature mode, the solidified decision rule is called to generate a creative selection instruction, and the calculation process of the online decision module is bypassed; Wherein, the solidified decision rule establishes a mapping relationship between the feature mode and the creative selection instruction.

10. The multi-source data fusion based advertisement creative dynamic evaluation and intelligent decision system of claim 1, wherein: The system further comprises a constraint generation module, which is used to: Perform graph structure homogeneity detection on the sampled initial candidate creative combinations, if there is a homogenization combination subset with high graph structure similarity, then retain a representative combination from the homogenization combination subset according to the average weight of the nodes; Based on the dynamic creative feature graph, resample to generate new alternative creative combinations with graph structure difference meeting the preset diversity requirement; The representative combination and the new alternative creative combination jointly constitute the candidate creative combination set after redundancy removal, and are input to the online decision module.

Images