An intelligent advertisement delivery decision method and system

Abstract

The application discloses a kind of intelligent advertisement delivery decision method and system, including based on user multivariate behavior data constructs user behavior model feature set and user dynamic knowledge graph;According to the user behavior model feature set and the user dynamic knowledge graph form user digital feature representation set;Based on the advertisement feature set of advertisement to be put, the environmental information feature set of current delivery environment and the user digital feature representation set, utilize the delivery strategy of pre-trained deep learning network generation advertisement to be put;Wherein the delivery strategy includes macroscopic delivery strategy and fine delivery strategy;The application can quickly generate macroscopic and fine delivery strategy, realizes from advertisement delivery user positioning, channel adaptation to specific delivery period, style presentation and full range of decision-making of bid control, improve the efficiency and automation degree of advertisement delivery.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent advertising placement decision-making method and system. Background Technology

[0002] In today's digital age, market competition is fierce, new products emerge constantly, and consumers face a deluge of information. Through precise advertising, businesses can accurately convey the unique selling points and functional advantages of their products to their target audience, stimulating consumer desire. Continuous and widespread advertising can strengthen brand image in consumers' minds, enhancing brand awareness and reputation. Advertising can also influence consumer purchasing behavior in various ways. However, with the rapid development of internet technology, user behavior data is becoming increasingly diverse and massive. Traditional advertising methods often rely on limited data and experience for decision-making, making it difficult to comprehensively and deeply understand user needs. This results in low conversion rates and significant resource waste. Furthermore, traditional methods struggle to make timely and accurate adjustments based on real-time market conditions and dynamic user changes. Summary of the Invention

[0003] To address the aforementioned problems in the existing technology, this invention provides an intelligent advertising delivery decision-making method and system. The technical problem to be solved by this invention is achieved through the following technical solution:

[0004] A smart advertising delivery decision-making method includes:

[0005] A user behavior model feature set and a dynamic user knowledge graph are constructed based on diverse user behavior data; wherein, the diverse user behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data;

[0006] A user digital feature representation set is formed based on the user behavior model feature set and the user dynamic knowledge graph;

[0007] Based on the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, a delivery strategy for the ad to be delivered is generated using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy.

[0008] In one specific implementation, a feature set for constructing a user behavior model is built based on diverse user behavior data, including:

[0009] Standardize and normalize user behavior data during preprocessing;

[0010] The preprocessed user multi-behavioral data is divided into multiple behavioral sequences according to time windows; wherein, the behavioral sequence includes user multi-behavioral data slices, first timestamps, behavioral statistics, behavioral tags, and sequence dependencies;

[0011] Each of the aforementioned behavior sequences is subjected to feature extraction to form multiple subsets of user behavior features;

[0012] Multiple subsets of user behavior features are merged to form a user behavior model feature set.

[0013] In one specific implementation, multiple subsets of user behavior features are fused to form a user behavior model feature set, including:

[0014] Multiple subsets of user behavior features are sequentially aligned and aggregated to form multiple aggregated features;

[0015] Multiple aggregated features are concatenated to form the user behavior model feature set; wherein, the user behavior model feature set includes behavioral statistical features, behavioral label features, sequence dependency features, time window features, and derived features.

[0016] In one specific implementation, a dynamic user knowledge graph is constructed based on diverse user behavior data, including:

[0017] Based on the user's diverse behavioral data, entities and the relationships between entities are extracted.

[0018] A user knowledge graph is constructed based on a graph database according to the entities and the relationships between the entities, wherein the nodes of the user knowledge graph are the entities, and the edges of the user knowledge graph are the relationships between the entities;

[0019] The user knowledge graph is updated at a preset interval, and a second timestamp is marked in the user knowledge graph to form a dynamic user knowledge graph.

[0020] In one specific implementation, a user digital feature representation set is formed based on the user behavior model feature set and the user dynamic knowledge graph, including:

[0021] The relational features and semantic features of the user dynamic knowledge graph are extracted, wherein the relational features are obtained by one-hot encoding the relational strength value and relational type between the entities, and the semantic features are obtained by transforming the semantic information of the entities;

[0022] The features, relational features, and semantic features in the user behavior model feature set are mapped and aligned to a unified feature space to have the same feature dimension, and a high-dimensional dense feature vector is generated based on the Transformer encoder.

[0023] The high-dimensional dense feature vector is combined with the user identifier and the third timestamp to form a user digital feature representation set, wherein the third timestamp represents the time point at which the user digital feature representation set was generated.

[0024] In one specific implementation, the advertising feature set of the advertisement to be delivered includes a subset of advertising content features, a subset of advertising historical performance features, and a subset of advertising targeting features. The environmental information feature set of the current delivery environment includes a subset of platform traffic features, a subset of market competition features, a subset of user activity features, and a subset of external event features.

[0025] In one specific implementation, the deep learning network includes a top-level decision network and a bottom-level execution network;

[0026] The top-level decision network is used to generate the macro-level delivery strategy, which includes an advertising user targeting strategy and an advertising channel adaptation strategy.

[0027] The underlying execution network is used to generate the refined delivery strategy, which includes an ad delivery time optimization strategy, an ad delivery style presentation strategy, and an ad delivery bidding control strategy.

[0028] In one specific implementation, the top-level decision network includes a first input layer, a feature encoding layer, a strategy generation layer, and a first output layer; wherein, the first input layer is used to fuse the received advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set to form a fused feature vector;

[0029] The feature encoding layer incorporates the fused feature vector into a global dependency based on a multi-head attention mechanism to form an attention-weighted feature, which is then input into a feedforward neural network to obtain a macro-policy vector.

[0030] The strategy generation layer includes a first fully connected sublayer and a second fully connected sublayer. The first fully connected sublayer receives the macro strategy vector and maps it to a pre-defined user segmentation label space to generate user segmentation labels and user segmentation probabilities. The user segmentation probabilities are sorted to generate a first priority sort to obtain a user positioning strategy. The second fully connected sublayer receives the macro strategy vector and maps it to a pre-defined channel allocation weight space to generate channel allocation weights. The channel allocation weights are normalized to generate a second priority sort to finally obtain a channel adaptation strategy.

[0031] The output layer receives the user positioning strategy and the channel adaptation strategy and merges them to generate a macro-level delivery strategy.

[0032] In one specific embodiment, the underlying execution network includes a second input layer, an execution decoding layer, and a second output layer;

[0033] The second input layer receives the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, the user digital feature representation set, and the macro delivery strategy, and concatenates them to generate a concatenated feature set. The concatenated feature set is then weighted through an attention mechanism to generate a strategy-feature joint vector.

[0034] The execution decoding layer receives the strategy-feature joint vector input into the Transformer decoder to obtain a three-dimensional strategy vector of the ad content to be delivered, the ad user, and the ad delivery environment; and inputs the three-dimensional strategy vector into the action generation module to obtain style presentation action, time optimization action, and bidding adjustment action;

[0035] The second output layer receives the style presentation action, time optimization action, and bidding adjustment action, and performs ad style selection, ad time period selection, and real-time bidding adjustment to finally obtain a refined delivery strategy.

[0036] In one specific embodiment, an intelligent advertising delivery decision system includes:

[0037] The construction unit is used to build a user behavior model feature set and a user dynamic knowledge graph based on user multi-dimensional behavior data; wherein, the user multi-dimensional behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data;

[0038] The user digital feature representation set forming unit forms a user digital feature representation set based on the user behavior model feature set and the user dynamic knowledge graph.

[0039] The advertising strategy delivery unit is used to generate a delivery strategy for the advertisement to be delivered based on the advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy.

[0040] The beneficial effects of this invention are:

[0041] This invention discloses an intelligent advertising placement decision-making method and system. By utilizing multi-dimensional behavioral data from user search, social interaction, consumption, content interaction, and application usage, it constructs a user behavior model feature set and a dynamic user knowledge graph. This enables a comprehensive and in-depth understanding of users' interests, preferences, needs, and consumption habits, allowing for more precise targeting of advertising users and improving the relevance of ads to users, thereby increasing click-through rates and conversion rates. The system updates the user knowledge graph at preset intervals and marks it with timestamps, forming a dynamic user knowledge graph. This reflects changes in user behavior and interests in real time, allowing advertising strategies to be dynamically adjusted based on the latest user information. Through the application of deep learning networks and the architectural design of a top-level decision network and a bottom-level execution network, it can quickly generate macro and fine-grained placement strategies, achieving comprehensive decision-making from user positioning and channel adaptation to specific placement times, presentation styles, and bidding adjustments, thus improving the efficiency and automation of advertising placement.

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0043] Figure 1 This is a flowchart of an intelligent advertising placement decision-making method provided by an embodiment of the present invention;

[0044] Figure 2 This is a flowchart of a user behavior model feature set construction method provided by an embodiment of the present invention;

[0045] Figure 3 This is a flowchart of a method for forming a user digital feature representation set provided by an embodiment of the present invention;

[0046] Figure 4 This is a block diagram of an intelligent advertising placement decision system provided in an embodiment of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0048] Example 1

[0049] In one specific implementation, please refer to Figure 1 , Figure 1 This is a flowchart of an intelligent advertising placement decision-making method, with the following specific steps:

[0050] S1: Construct a user behavior model feature set and a user dynamic knowledge graph based on diverse user behavior data; wherein, the diverse user behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data;

[0051] For example, user search behavior data may include user search keywords, search time, search results, clicks, etc.; user social behavior data may include user following lists, likes, comments, and reposts on social platforms such as Weibo and WeChat; user consumption behavior data may include user purchase information, purchase time, purchase amount, purchase frequency, etc.; user content interaction behavior data may include user browsing time, likes, comments, and shares of articles and videos on content distribution platforms such as Toutiao and Douyin; and user application usage behavior data may include user application opening time, usage time, and function usage in mobile applications.

[0052] In one embodiment of this application, such as Figure 2 As shown, Figure 2 This is a flowchart of a user behavior model feature set construction method provided by an embodiment of the present invention. Based on diverse user behavior data, it constructs a user behavior model feature set, including:

[0053] S121: Perform standardization and normalization preprocessing on the user's diverse behavioral data;

[0054] For example, user multi-dimensional behavioral data can be standardized and normalized to convert data of different dimensions into a unified scale.

[0055] S122: The preprocessed user multi-behavioral data is divided into multiple behavioral sequences according to time windows; wherein, the behavioral sequence includes the user multi-behavioral data slices, first timestamps, behavioral statistics, behavioral tags, and sequence dependencies;

[0056] For example, a time window can be set according to factors such as business needs, data update frequency, and user behavior change cycle. The preprocessed user multi-behavior data can be divided into multiple behavior sequences. For some fields with more frequent and faster changes in consumption behavior, such as fashion and beauty, and fast-moving consumer goods, the time window can be set to a relatively short one or three days in order to capture the dynamic changes in user behavior in a timely manner. For some fields with longer consumption decision cycles and relatively stable behavior, such as real estate and automobiles, the time window can be set to one month or one quarter.

[0057] The first timestamp indicates the start time of the behavior sequence, clearly defining the time range covered by the behavior sequence. The first timestamp can serve as a reference for the time dimension, helping to understand the changing trend of user behavior over time. For example, assuming the current time window is from October 1, 2024 to October 7, 2024, the first timestamp is 00:00:00 on October 1, 2024. Behavioral data within this time window are all included in this behavior sequence for analysis.

[0058] Behavioral statistics are the statistics of the number of times a user performs various behaviors within a time window. They can intuitively reflect the user's activity level and preferences during a specific time period. For example, statistics can be used to count the number of searches a user makes through a search engine within a certain time window. For instance, a user might have searched for keywords such as "new mobile phone," "travel guide," and "food recommendations" a total of 15 times within a certain time window. Statistics can also be used to count the number of purchases and the amount spent on e-commerce platforms within a certain time window. For instance, a user might have purchased three items within a certain time window: a Bluetooth headset worth 299 yuan, a sports T-shirt worth 199 yuan, and a smartwatch worth 599 yuan. Finally, statistics can be used to count the number of likes, comments, and reposts a user makes on social media platforms within a certain time window. For instance, a user might have liked 20 Weibo posts, commented on 5, and reposted 3.

[0059] Behavioral tags identify the categories to which user behaviors belong, helping to classify and summarize user behaviors and thus better understand user needs and interests. For example, search keyword category tags categorize users' search keywords, such as "new mobile phone" belonging to the "electronic products" category, "travel guide" belonging to the "travel" category, and "food recommendation" belonging to the "food and beverage" category; purchase category tags categorize the products purchased by users, such as Bluetooth headphones and smartwatches belonging to the "digital products" category, and sports T-shirts belonging to the "clothing" category; social content category tags categorize social content that users like, comment on, and forward. If a user likes 10 Weibo posts about entertainment news, 5 about technology news, and 5 about interesting things in life, they can be categorized as "entertainment news," "technology news," and "interesting things in life," respectively.

[0060] Sequence dependency represents the order and relationship between actions, reflecting the logic and causal relationship of user behavior within a certain time window. By analyzing sequence dependencies, we can understand the user's behavioral path and decision-making process. For example, search-purchase behavior dependency: suppose a user first searches for "new mobile phone reviews" on a search engine, and then buys a mobile phone on an e-commerce platform. There is a dependency between the user's search behavior and purchase behavior, and the search behavior may provide the basis for the user's purchase decision. Social-consumption behavior dependency: a user browses a recommendation article about a certain brand of cosmetics on a social platform, likes and comments on it, and then buys the brand of cosmetics on an e-commerce platform. Social behavior influences the user's consumption behavior, and there is a sequence dependency between the two.

[0061] S123: Perform feature extraction on each of the aforementioned behavior sequences to form multiple subsets of user behavior features;

[0062] For example, features can be extracted based on behavioral statistics, directly extracting numerical features from the behavioral sequence, such as search count, purchase count, purchase amount, like count, comment count, and forward count. These values ​​can also be further calculated to generate derived numerical features such as average search frequency (search count / time window length) and average purchase amount (purchase amount / purchase count).

[0063] For example, features can be extracted based on behavioral labels, and the behavioral labels can be encoded using one-hot encoding or multi-label encoding. One-hot encoding is suitable when a behavioral sequence belongs to only one category. The category is converted into a binary vector with a length equal to the total number of categories. Only the elements corresponding to the category are 1, and the rest are 0. Multi-label encoding is suitable for situations where a behavior may belong to multiple categories. For example, if there are three categories for search keywords: "electronic products", "travel", and "food", and for a certain behavioral sequence, the search keyword category features vector for the behavioral sequence is [1,1,0] after multi-label encoding. If there are three categories for purchased goods: "digital products", "clothing", and "home furnishings", and the purchased goods category is "digital products", then it is encoded as [1,0,0].

[0064] For example, features can be extracted based on sequence dependencies, such as the probability of transitioning from one behavior to another, i.e., the statistical behavior transition probability. Specifically, the probability of transitioning from a search behavior to a purchase behavior can be calculated by determining the proportion of all behavior sequences containing search behavior that are immediately followed by a purchase behavior. The time interval between adjacent behaviors, i.e., the behavior interval time, such as the time interval between a search behavior and a purchase behavior, can be calculated to extract features such as average interval time, longest interval time, and shortest interval time. For example, if there are 100 behavior sequences containing search behavior, and 30 of them are immediately followed by a purchase behavior, then the probability of transitioning from search to purchase is 30 ÷ 100 = 0.3. In a behavior sequence, if a user first searches at 10:00 on October 2nd and then makes a purchase at 15:00 on October 3rd, with a time interval of 29 hours between these two behaviors, the time interval between all search-purchase behaviors of this user can be calculated to obtain features such as average interval time, longest interval time, and shortest interval time.

[0065] For example, features can be extracted based on time windows. Specifically, features related to the time window can be extracted, such as the length of the time window, the month to which the time window belongs, and the week of the week. Such features can help the model understand the changing patterns of user behavior in different time periods. For example, if the length of the time window is 7 days (one week), this value can be directly used as a feature. If the time window is from October 1, 2024 to October 7, 2024, and the month is October, the week of the week can be represented by 1 to 7. Assuming that October 1 is a Tuesday, then the week of the week feature is 2.

[0066] Features based on behavioral statistics, behavioral tags, sequence dependencies, and time windows are combined to form a subset of user behavior features for a behavioral sequence. For example, a subset of user behavior features for a behavioral sequence might include: [average search frequency: 2.14 times / day; average purchase amount: 365.67 yuan / time; search keyword category feature vector: [1,1,0]; purchase product category feature vector: [1,0,0]; conversion probability from search to purchase: 0.3; average search-purchase interval: 24 hours; time window length: 7 days; time window month: October; time window week: 2].

[0067] S124: Merging multiple subsets of user behavior features to form a user behavior model feature set, specifically including the following steps:

[0068] S1241: Sequentially perform feature alignment and feature aggregation on multiple subsets of user behavior features to form multiple aggregated features;

[0069] For example, features extracted from all user behavior feature subsets are aligned. Specifically, feature templates can be established, including feature names, feature types (numerical, categorical, etc.), feature descriptions, and feature value ranges. For each user, all user behavior feature subsets are traversed, and features in each user feature subset are matched and integrated according to the feature template and feature names. If a feature is missing, it can be filled in according to business needs and data characteristics, such as using default values. Feature format unification includes converting numerical values ​​in all user behavior feature subsets to the same unit, such as unifying time units to seconds or minutes, and monetary units to yuan. Category labels in all user behavior feature subsets use the same encoding method, such as one-hot encoding or multi-label encoding.

[0070] For example, aligned features can be aggregated to form aggregated features. For numerical features, such as average search frequency and average purchase amount, aggregation methods such as summation, average, maximum, and minimum values ​​can be used to reflect the overall behavioral characteristics of users in different time windows. For categorical features, such as search keyword category features and purchased product category features, aggregation methods such as concatenation and counting can be used. For example, the search keyword category features of users in different time windows can be concatenated, or the number of searches by users in different categories can be counted. For sequence dependency features, such as behavior transition probability and behavior interval time, global statistics such as average transition probability and longest / shortest interval time can be calculated, or the transition path information of each user between different behavior sequences can be retained so that subsequent models can learn more complex sequence dependencies. For time window features, such as time window length, month, and weekday, the feature values ​​of each user in different time windows can be directly retained, or further encoding or transformation can be performed according to business needs.

[0071] S1242: Concatenate multiple aggregated features to form the user behavior model feature set; wherein, the user behavior model feature set includes behavioral statistical features, behavioral label features, sequence dependency features, time window features, and derived features.

[0072] For example, the aggregated features are concatenated to form a user behavior model feature set, including behavioral statistical features, specifically: average search frequency (the average number of searches by the user across all time windows); average purchase amount (the average purchase amount by the user across all time windows); total likes, comments, and shares (the number of social interactions by the user across all time windows); behavioral tag features, specifically: search keyword category (the encoding vector representing the category of keywords searched by the user in different time windows); purchased product category (the encoding vector representing the category of products purchased by the user in different time windows); social content category (the encoding vector representing the category of social content liked, commented on, and shared by the user in different time windows); and sequence dependency features, specifically: average behavior transition probability (the probability of transitions between different behavior sequences). The features include: average conversion probability, longest / shortest behavior interval (the longest or shortest time interval between different behavior sequences); behavior path features (representing the user's behavior sequences and paths within different time windows, which can be used to learn user behavior patterns and decision-making processes); time window features (specifically, time window length distribution, representing the distribution of users across different time window lengths, which can be used to understand the temporal patterns of user behavior); month / week features (representing the user's behavioral characteristics in different months or weeks, which can be used to analyze seasonal or periodic changes in user behavior); and other derived features. Depending on business needs and data characteristics, other types of features can be derived, such as user activity features (comprehensive activity based on search, purchase, social, etc.), user preference index features (preference evaluation based on behavioral tags and statistical features), etc.

[0073] By fusing features, a comprehensive feature set for user behavior models can be constructed, providing strong support for subsequent tasks such as user behavior prediction and personalized recommendation.

[0074] In one embodiment of this application, constructing a dynamic user knowledge graph based on diverse user behavior data includes:

[0075] S131: Extract entities and relationships between entities based on the user's diverse behavioral data;

[0076] For example, first, the entity type is determined, clarifying the entity categories involved in the user behavior data, mainly including user entities, product entities, platform entities, behavior entities, time entities, etc. Second, entity extraction is performed, extracting specific entities from the user's multi-dimensional behavior data. For example, user A's multi-dimensional behavior data on October 2, 2024 is as follows:

[0077] Search "smartwatch recommendations" on Baidu;

[0078] I liked a promotional post for Huawei smartwatches on Weibo;

[0079] I bought an Apple Watch on JD.com.

[0080] The extracted entities include: user entity, User A; product entity, smartwatch, Huawei smartwatch, Apple Watch; platform entity, Baidu, Weibo, JD.com; behavior entity, search, like, purchase; and time entity: October 2, 2024. Finally, the relationships between entities are extracted, the associations between entities are analyzed, the relationship types are determined, and the relationships between entities are extracted. Relationship types can be defined as "user-search-keyword," "user-like-social content," "user-purchase-product," "behavior-occurrence-time," etc. Therefore, the extracted relationships between entities include: User A-search-smartwatch recommendation; User A-like-Huawei smartwatch promotional post; User A-purchase-Apple Watch; Search-occurrence-October 2, 2024; Like-occurrence-October 2, 2024; Purchase-occurrence-October 2, 2024.

[0081] S132: Construct a user knowledge graph based on the entities and the relationships between the entities using a graph database; wherein, the nodes of the user knowledge graph are the entities, and the edges of the user knowledge graph are the relationships between the entities;

[0082] For example, first, a graph database is selected, such as Neo4j. Second, entity and relation modeling is performed, with extracted entities as nodes in the graph and relations as edges between nodes. Attributes are added to the nodes and edges. For instance, the attributes for the User A node could be: User ID = 12345, Age = 30, Gender = Male; the attributes for the Apple Watch node could be: Product ID = P67890, Brand = Apple, Price = 3999 yuan; and the edge attributes could be: User A - Purchase - Apple Watch, Purchase Time = October 2, 2024, Payment Method = WeChat Pay. Finally, the graph is constructed. The node and edge data can be imported using the graph database's API to generate a user knowledge graph.

[0083] S133: Update the user knowledge graph according to a preset interval, and mark the second timestamp in the user knowledge graph to form a dynamic user knowledge graph;

[0084] For example, first, a preset interval is set, which can be set according to business needs and data change frequency, such as 1 day. Second, incremental updates are performed. Within each preset interval, steps S131-S132 are repeated to extract new entities and relationships, and these are added to the existing knowledge graph. For example, if user A watched and saved a Xiaomi smartwatch review video on Douyin on October 3, 2024, then new entities "Xiaomi smartwatch", "Douyin", and "watching and saving behavior" are extracted, along with the relationships "User A - Watching and saving - Xiaomi smartwatch review video" and "Watching and saving - Occurred - October 3, 2024". The new nodes and edges are added to the original knowledge graph, and finally, a timestamp is added for each update operation to record the update time of the knowledge graph, forming a user dynamic knowledge graph. For example, after the update operation on October 3, 2024 is completed, a second timestamp "October 3, 2024, 23:59:59" is added to the newly added nodes and edges and the updated relationships. At the same time, the second timestamp of this update is recorded to form a user dynamic knowledge graph. Through the second timestamp, the evolution process of the knowledge graph can be traced. For example, it is possible to analyze the changes in user A's interest in smartwatch brands at different points in time.

[0085] Through the above steps, a dynamic user knowledge graph is constructed from diverse user behavior data, which can dynamically reflect changes in user behavior and interests, providing rich semantic information support for subsequent advertising decisions.

[0086] S2: Form a user digital feature representation set based on the user behavior model feature set and the user dynamic knowledge graph; such as Figure 3 As shown, Figure 3 This is a flowchart of a method for forming a user digital feature representation set provided by an embodiment of the present invention;

[0087] S21: Extract the relational features and semantic features of the user dynamic knowledge graph, wherein the relational features are obtained by one-hot encoding the relational strength value and relational type between the entities, and the semantic features are obtained by transforming the semantic information of the entities;

[0088] For example, the relationship feature extraction of the user dynamic knowledge graph can extract information such as relationship type and relationship strength from the user dynamic knowledge graph to construct relationship features. First, each node and edge is accessed sequentially. For each edge, its starting node, ending node, and edge attributes are obtained. For example, the Cypher query language can be used to traverse the nodes and edges in the user dynamic knowledge graph, and the traversed relationship data is stored in a list, which can include the starting node ID, ending node ID, edge type, and edge attributes, etc. Second, all relationship types existing in the user dynamic knowledge graph are determined, such as "user-search results-keywords", "user-likes-social content", "user-purchases-products", etc. Each relationship type is counted to determine its occurrence frequency in the user dynamic knowledge graph. This can be achieved by traversing the previously stored relationship data, or by using a dictionary or other data structure to record the occurrence frequency of each relationship type. The relationship strength calculation index is determined based on the edge attributes. For example, for the "user-purchases-products" relationship, the purchase amount can be used as the strength index; for the "user-likes-social content" relationship, the purchase amount can be used as the strength index; for the "user-likes-social content" relationship, the purchase amount can be used as the strength index. The "content" relationship can be defined using the number of likes as a strength indicator. Further, based on the determined strength indicator, the relationship strength of each edge is calculated. For a purchase relationship, the relationship strength can be the purchase amount; for a like relationship, the relationship strength can be the number of likes. Finally, relationship features are constructed using one-hot encoding to encode the relationship type. For example, if there are three relationship types: "user-search result-keyword" and "user-purchase-product", then each relationship type corresponds to a one-hot encoded vector, such as "user-search result-keyword". The relationship strength value corresponds to [1,0,0], "user-like-social content" corresponds to [0,1,0], and "user-purchase-product" corresponds to [0,0,1]. The relationship strength value is combined with the corresponding relationship type encoding vector to form the relationship feature. For example, for a "user-purchase-product" relationship with a purchase amount of 3999 yuan, the relationship feature can be represented as: [relationship type: purchase (one-hot encoding), relationship strength: 3999 yuan], that is, [0,0,1,3999] (assuming that the one-hot encoding dimension corresponding to the first three relationship types is 3).

[0089] For example, the semantic feature extraction of the user dynamic knowledge graph involves extracting entity information, including category and attributes, and using word embedding technology to convert the entity information into semantic features. Specifically, firstly, entity information is extracted by analyzing the entities in the dynamic knowledge graph and extracting information such as the entity's category and attributes. For example, for product entities, attributes such as category (e.g., "digital products," "clothing"), brand, and price can be extracted; for user entities, attributes such as age and gender can be extracted. Secondly, entity semantic featureization is performed by using word embedding technology to convert the semantic information such as the entity's category and attributes into features. For example, word embedding technology can be used... The pre-trained word vector model converts the entity category "digital products" into a fixed-length feature. Then, it analyzes the semantic meaning of the relationships, such as the "user-like-social content" relationship indicating a user's interest in a certain piece of social content, and the "user-search bar keyword" relationship indicating a user's search interest in a certain keyword. Finally, word embedding technology is used to convert the relationship description into features. For example, the relationship "user-like-social content" is converted into a fixed-length feature. User A has the following relationships and semantic information in the user dynamic knowledge graph: User A's relationship with the product Apple Watch is "purchase", and the purchase amount is 3999 yuan. Through one-hot encoding, the encoding vector corresponding to the "purchase" relationship type is [0,0,1]. Therefore, the feature of this relationship is [0,0,1,3999]. In addition, User A liked a promotional blog post about Huawei smartwatches. Through word embedding technology, the text content of this blog post is converted into a 100-dimensional semantic feature vector, such as [0.2,0.5,...,0.7].

[0090] S22: Align the features, relational features, and semantic features in the user behavior model feature set to a unified feature space through feature space mapping so that they have the same feature dimension, and generate a high-dimensional dense feature vector based on the Transformer encoder;

[0091] For example, feature alignment is performed first for feature space mapping. This involves determining the common dimensions of the user behavior model feature set, relational features, and semantic features, and aligning them along the same dimensions. For instance, if the user behavior model feature set has 10 dimensions, the relational features have 5 dimensions, and the semantic features have 100 dimensions, then they need to be mapped to a unified feature space, such as 128 dimensions. Secondly, feature rearrangement and supplementation are performed. Based on the common dimensions, each feature is rearranged and supplemented to align them in the unified feature space. For example, for the user behavior model feature set, its 10-dimensional features can be extended to 128 dimensions through padding or interpolation. Similar processing can be performed for relational and semantic features. Next, feature transformation is performed. For numerical features (such as average search frequency, average purchase amount, etc.), direct transformation can be used. The numerical values ​​need to be normalized to ensure they are consistent with other features. For example, min-max normalization can be used to scale the values ​​to the range of [0,1]. For categorical features (such as relation types, entity categories, etc.), encoding conversion is required. In addition to one-hot encoding, embedding techniques can be used to convert categorical features into vectors. For example, the relation type encoding vector can be converted into a higher-dimensional dense vector through an embedding layer. The aligned user behavior model feature set, relation features, and semantic features are concatenated in a certain order to form a long vector. For example, if the user behavior model feature set has 10 dimensions, the relation features have 5 dimensions, and the semantic features have 100 dimensions, mapping them to a unified 128-dimensional feature space will result in a concatenated vector with 128 dimensions (logically integrating the information of the three vectors into 128 dimensions).

[0092] For example, when a Transformer encoder generates a high-dimensional dense feature vector, the concatenated feature vector is used as the input sequence and fed into the Transformer encoder. If the dimension of the feature vector is inconsistent with the input dimension of the Transformer encoder, it needs to be adjusted accordingly. For example, a linear transformation layer can be used to convert the dimension of the concatenated feature vector to be consistent with the input dimension of the Transformer encoder. In order to capture the positional information in the feature sequence, positional encoding can be added before inputting it into the Transformer encoder. Positional encoding can be a fixed sine or cosine function encoding, or a learnable positional embedding vector. To generate high-dimensional dense feature vectors, a self-attention mechanism can be first implemented. In the Transformer encoder, this mechanism calculates the correlation between features at different positions in the feature sequence, generating attention weights. The self-attention mechanism dynamically adjusts these weights based on the similarity between features, allowing the model to focus on more important features. After the self-attention mechanism, a feedforward neural network is used to perform a non-linear transformation on the features, further extracting high-level representations. This feedforward neural network typically contains two linear transformation layers and an activation function (e.g., ReLU). Finally, the Transformer encoder outputs a high-dimensional dense feature vector that captures the complex relationships between features and possesses high information density.

[0093] For example, it outputs a 512-dimensional dense feature vector that integrates important information from user behavior model feature sets, relational features, and semantic features.

[0094] For example, suppose the feature vector of the user behavior model feature set is [average search frequency:

[0095] The average purchase amount is 365.67 yuan per purchase, with 2.14 purchases per day. After normalization, the values ​​are [0.6, 0.7]. The relational feature vector is [0, 0, 1, 3999], which is [0, 0, 1, 0.8] after normalization and dimensionality adjustment (assuming the maximum purchase amount is 5000 yuan). The semantic feature vector is the word embedding vector of the Huawei smartwatch promotional blog post [0.2, 0.5, ..., 0.7] (100 values ​​in total).

[0096] These features are aligned to a unified 128-dimensional feature space through feature space mapping. Assuming that after feature transformation and alignment, the feature vector of the user behavior model feature set is expanded to 128 dimensions, with the first two non-zero dimensions being 0.6 and 0.7 respectively; the relational feature vector is expanded to 128 dimensions, with the third and fourth non-zero dimensions being 1 and 0.8 respectively; the semantic feature vector is filled into the remaining dimensions of the 128-dimensional feature space after dimensionality reduction. The aligned feature vectors are then concatenated to form a 128-dimensional comprehensive feature vector. This vector is input into the Transformer encoder, where it undergoes computation via a self-attention mechanism and a feedforward neural network to generate a high-dimensional dense feature vector. For example, a 512-dimensional dense feature vector is output, where the value of each dimension ranges from [0,1], such as [0.3, 0.6, ..., 0.8] (a total of 512 values). This dense feature vector captures the complex relationships between the user behavior model feature set, relational features, and semantic features, and can serve as input features for subsequent tasks such as advertising placement decisions.

[0097] S23: Combine the high-dimensional dense feature vector with the user identifier and the third timestamp to form a user digital feature representation set, wherein the third timestamp represents the time point at which the user digital feature representation set was generated.

[0098] In one embodiment of this application, the user identifier (the user's unique ID, such as a string or integer, used to identify the user) and the third timestamp (representing the time point when the user's digital feature representation set was generated, used to record the time of feature generation) are appended as metadata to a high-dimensional dense feature vector to form structured data and finally obtain the user's digital feature representation set; for example, the high-dimensional dense feature vector = [512-dimensional high-dimensional dense feature], and the user's digital feature representation set obtained by appending the user identifier and timestamp as metadata to the feature vector is [user identifier (e.g., user_123), third timestamp (e.g., 2023-10-01 14:30:00), high-dimensional dense feature vector (e.g., 512-dimensional vector)].

[0099] S3: Based on the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, a delivery strategy for the ad to be delivered is generated using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy.

[0100] In one embodiment of this application, the advertising feature set of the advertisement to be delivered includes a subset of advertising content features, a subset of advertising historical performance features, and a subset of advertising targeting features. The environmental information feature set of the current delivery environment includes a subset of platform traffic features, a subset of market competition features, a subset of user activity features, and a subset of external event features.

[0101] For example, the ad feature set of the ad to be delivered includes a subset of ad content features, a subset of ad historical performance features, and a subset of ad targeting features. The subset of ad content features includes text content features of the ad, such as ad title and ad body; image content features, such as product images and promotional posters; and video content features, such as product demonstration videos and ad clips. The subset of ad historical performance features includes metrics such as click-through rate, conversion rate, and number of impressions recorded for the ad. These metrics can reflect the effectiveness of the ad in different user groups and delivery environments. The subset of ad targeting features includes the target user characteristics set for the ad, such as age range, gender, region, and interests.

[0102] For example, the environmental information feature set of the current advertising environment includes a platform traffic feature subset, a market competition feature subset, a user activity feature subset, and an external event feature subset. The platform traffic feature subset includes the daily active users, user online time distribution, page views, and user dwell time for each advertising platform, which helps to understand the platform's traffic scale and user activity. The market competition feature subset includes the number of ads placed by competitors and their bidding range, which helps to develop more competitive advertising strategies. The user activity feature subset includes user login frequency and interaction behavior at different times, reflecting the overall activity level of users on the current platform and helping to select appropriate advertising periods. The external event feature subset includes external factors such as holidays, major events, and social hot topics, considering the impact of external events on advertising; if user interests and behaviors change, the advertising strategy needs to be adjusted accordingly.

[0103] In one embodiment of this application, the deep learning network includes a top-level decision network and a bottom-level execution network; wherein, the top-level decision network is used to generate the macro-level delivery strategy, which includes an ad delivery user positioning strategy and an ad delivery channel adaptation strategy; the bottom-level execution network is used to generate the fine-grained delivery strategy, which includes an ad delivery time period optimization strategy, an ad delivery style presentation strategy, and an ad delivery bidding control strategy.

[0104] In one embodiment of this application, the top-level decision network includes a first input layer, a feature encoding layer, a strategy generation layer, and a first output layer; wherein, the first input layer receives the advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, and fuses them to form a fused feature vector;

[0105] For example, the first input layer receives the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, and then performs feature preprocessing.

[0106] Specifically, for the advertising content feature subset within the advertising feature set, the text content is converted into a fixed-length feature vector using a pre-trained word vector model (Word2Vec, GloVe, or BERT can be selected). For image or video content, a convolutional neural network is used to extract visual features, which are then flattened into one-dimensional vectors. The numerical features such as click-through rate, conversion rate, and number of impressions in the advertising history performance feature subset are normalized, using min-max normalization to scale the values ​​to the [0,1] range. The category features such as age range, gender, and region in the advertising targeting feature subset are one-hot encoded. Alternatively, multi-label encoding can be used to convert categories into numerical vectors. For the platform traffic feature subset in the environmental information feature set, numerical features such as daily active users, user online time distribution, page views, and user dwell time, numerical features such as the number of advertisements placed by competitors and their bidding range in the market competition feature subset, and numerical features such as user login frequency and interaction behavior in different time periods in the user activity feature subset are normalized. For the category features such as holidays and major events in the external event feature subset, holidays can be encoded as 1 and non-holidays as 0. The user digital feature representation set can directly use high-dimensional dense feature vectors.

[0107] For example, the preprocessed advertising feature set, environmental information feature set, and user digital feature representation set are concatenated in a certain order; for example, the various subsets of the advertising feature set can be concatenated first, then the various subsets of the environmental information feature set can be concatenated, and finally they can be concatenated with the user digital feature representation set to form a long vector.

[0108] Calculate the attention weight for each feature subset; a simple fully connected layer can be used, taking each feature subset as input and outputting a weight value. Then, the weights are normalized using the softmax function so that the sum of all weights is 1. The weighted sum of each feature subset is calculated based on the attention weights to obtain the fused feature vector.

[0109] In one embodiment of this application, the feature encoding layer incorporates the fused feature vector into a global dependency based on a multi-head attention mechanism to form an attention-weighted feature, which is then input into a feedforward neural network to obtain a macro-policy vector;

[0110] For example, the fused feature vector is input into a multi-head attention mechanism, and query, key, and value matrices are obtained through multiple different linear transformations. These matrices can be obtained using three different linear transformation matrices. The attention score for each head is calculated by taking the dot product of the query and key matrices and dividing by a scaling factor (usually the square root of the key matrix dimension). The result is then converted to a probability distribution using a softmax function to obtain the attention score. The value matrix is ​​then weighted and summed based on the attention score to obtain the output of each head. The outputs of all heads are concatenated and integrated using a linear transformation matrix to obtain the output of the multi-head attention mechanism. This output is then input into a feedforward neural network, which typically contains two linear transformation layers and an activation function (e.g., ReLU). The input vector is transformed by the first linear transformation layer, then non-linearly transformed by the activation function, and finally passed through the second linear transformation layer to obtain the final output macroscopic policy vector.

[0111] In one embodiment of this application, the strategy generation layer includes a first fully connected sublayer and a second fully connected sublayer. The first fully connected sublayer receives the macro-strategy vector and maps it to a pre-defined user segmentation label space to generate user segmentation labels and user segmentation probabilities. The user segmentation probabilities are sorted to generate a first priority sort, and finally a user positioning strategy is obtained. The second fully connected sublayer receives the macro-strategy vector and maps it to a pre-defined channel allocation weight space to generate channel allocation weights. The channel allocation weights are normalized to generate a second priority sort, and finally a channel adaptation strategy is obtained.

[0112] For example, the macro-strategy vector is received and input into the first fully connected layer. The output dimension of the first fully connected layer is equal to the dimension of the pre-defined user segment label space. For example, if users are divided into three segments: high-value users, potential users, and low-value users, then the output dimension of the first fully connected layer is 3. The output of the fully connected layer is converted into a probability distribution of different user segments using the Sigmoid function. The segment with the highest probability is selected as the segment label for the current user. The user segment probabilities are sorted to generate a first priority ranking, ultimately yielding the user positioning strategy. The macro-strategy vector is then received and input into the second fully connected layer. The output dimension of the second fully connected layer is equal to the dimension of the pre-defined channel allocation weight space. For example, if advertising is conducted on three channels: search engines, social media, and e-commerce platforms, then the output dimension of the fully connected layer is 3. The output of the fully connected layer is converted into allocation weights for different advertising channels using the Sigmoid function. After normalizing the channel allocation weights, a second priority ranking is generated, ultimately yielding the channel adaptation strategy.

[0113] In one embodiment of this application, the output layer receives the user positioning strategy and the channel adaptation strategy and merges them to generate a macro-level delivery strategy.

[0114] For example, the system receives user targeting strategies and channel adaptation strategies, and merges them to generate a macro-level delivery strategy. This can be done by simply concatenating the two strategy vectors together, or by using a weighted summation method, assigning different weights to the two strategies and then performing a weighted summation.

[0115] In one embodiment of this application, the underlying execution network includes a second input layer, an execution decoding layer, and a second output layer;

[0116] The second input layer receives the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, the user digital feature representation set, and the macro delivery strategy, and concatenates them to generate a concatenated feature set. The concatenated feature set is then weighted through an attention mechanism to generate a strategy-feature joint vector.

[0117] For example, the second input layer receives the ad feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, the user digital feature representation set, and the macro delivery strategy; these features are concatenated in a certain order to form a concatenated feature set; for example, the various subsets of the ad feature set can be concatenated first, then the various subsets of the environmental information feature set can be concatenated, then the user digital feature representation set can be concatenated, and finally the macro delivery strategy can be concatenated; the attention weight of each feature in the concatenated feature set can be calculated; a simple fully connected layer can be used to take the concatenated feature set as input and output the weight value of each feature; then the softmax function is used to normalize all weights; the concatenated feature set is weighted and summed according to the attention weight to generate a strategy-feature joint vector.

[0118] In one embodiment of this application, the execution decoding layer receives the policy-feature joint vector input to the Transformer decoder to obtain a three-dimensional policy vector of the ad content to be delivered, the ad user, and the ad delivery environment; and inputs the three-dimensional policy vector into the action generation module to obtain style presentation action, time optimization action, and bidding adjustment action;

[0119] For example, the policy-feature joint vector is input into the Transformer decoder. The Transformer decoder consists of multiple identical layers, each containing a multi-head self-attention mechanism, a multi-head attention mechanism (for focusing on the encoder output), and a feedforward neural network. In the multi-head self-attention mechanism, the joint vector is used as the query, key, and value to calculate the self-attention score and obtain the self-attention output. In the multi-head attention mechanism, the joint vector is used as the query, and the macro-policy vector generated by the top-level decision network is used as the key and value to calculate the attention score and obtain the attention output. The self-attention output and the attention output are added together and processed through layer normalization and residual connections. The processed result is input into the feedforward neural network to obtain the output of the feedforward neural network. The output of the feedforward neural network is then added to the previous output and processed through layer normalization and residual connections. After processing through multiple layers, the Transformer decoder... The codec outputs a 3D strategy vector of the ad content to be delivered, the ad users, and the ad delivery environment. This 3D strategy vector is then input into the action generation module. The action generation module is a fully connected network that generates style presentation actions, time optimization actions, and bid adjustment actions based on different dimensions of the 3D strategy vector. Style presentation actions: Based on the dimensions related to ad display format in the 3D strategy vector, a probability distribution of different display formats is generated through a fully connected layer and a softmax function. The display format with the highest probability is selected as the style presentation action. For example, possible choices include images, videos, and text links. Time optimization actions: Based on the dimensions related to the delivery time period in the 3D strategy vector, a probability distribution of different time periods is generated through a fully connected layer and a softmax function. The time period with the highest probability is selected as the time optimization action. Bid adjustment actions: Based on the dimensions related to bid in the 3D strategy vector, a bid adjustment value is generated through a fully connected layer as the bid adjustment action.

[0120] In one embodiment of this application, the second output layer receives the style presentation action, time optimization action, and bidding adjustment action, and performs ad style selection, ad time period selection, and real-time bidding adjustment to finally obtain a refined delivery strategy.

[0121] For example, the system receives style presentation actions, time optimization actions, and bidding adjustment actions, and performs ad style selection, ad time slot selection, and real-time bidding adjustment; ad style selection: selects a suitable ad display format based on the style presentation action; ad time slot selection: determines the best time to place the ad based on the time optimization action; real-time bidding adjustment: adjusts the ad bid based on the bidding adjustment action to adapt to market competition and ultimately obtain a refined placement strategy.

[0122] In one specific implementation, please refer to Figure 4 , Figure 4 This is a block diagram of a smart advertising delivery decision system provided in an embodiment of the present invention, comprising:

[0123] The construction unit is used to build a user behavior model feature set and a user dynamic knowledge graph based on user multi-dimensional behavior data; wherein, the user multi-dimensional behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data;

[0124] The user digital feature representation set forming unit forms a user digital feature representation set based on the user behavior model feature set and the user dynamic knowledge graph.

[0125] The advertising strategy delivery unit is used to generate a delivery strategy for the advertisement to be delivered based on the advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy.

[0126] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0127] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0128] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A smart advertising placement decision-making method, characterized in that, include: A user behavior model feature set and a dynamic user knowledge graph are constructed based on diverse user behavior data; wherein, the diverse user behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data; A user digital feature representation set is formed based on the user behavior model feature set and the user dynamic knowledge graph; Based on the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, a delivery strategy for the ad to be delivered is generated using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy. The deep learning network includes a top-level decision network and a bottom-level execution network; The top-level decision network is used to generate the macro-level delivery strategy, which includes an advertising user targeting strategy and an advertising channel adaptation strategy. The underlying execution network is used to generate the refined delivery strategy, which includes an ad delivery time optimization strategy, an ad delivery style presentation strategy, and an ad delivery bidding control strategy. The top-level decision network includes a first input layer, a feature encoding layer, a strategy generation layer, and a first output layer; wherein, the first input layer is used to fuse the received advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set to form a fused feature vector; The feature encoding layer incorporates the fused feature vector into a global dependency based on a multi-head attention mechanism to form an attention-weighted feature, which is then input into a feedforward neural network to obtain a macro-policy vector. The strategy generation layer includes a first fully connected sublayer and a second fully connected sublayer. The first fully connected sublayer receives the macro strategy vector and maps it to a pre-defined user segmentation label space to generate user segmentation labels and user segmentation probabilities. The user segmentation probabilities are sorted to generate a first priority sort to obtain a user positioning strategy. The second fully connected sublayer receives the macro strategy vector and maps it to a pre-defined channel allocation weight space to generate channel allocation weights. The channel allocation weights are normalized to generate a second priority sort to finally obtain a channel adaptation strategy. The output layer receives the user positioning strategy and the channel adaptation strategy and merges them to generate a macro-level delivery strategy. The underlying execution network includes a second input layer, an execution decoding layer, and a second output layer; The second input layer receives the ad feature set of the ad to be delivered, the environmental information feature set of the current delivery environment, the user digital feature representation set, and the macro delivery strategy, and concatenates them to generate a concatenated feature set. The concatenated feature set is then weighted through an attention mechanism to generate a strategy-feature joint vector. The execution decoding layer receives the strategy-feature joint vector input into the Transformer decoder to obtain a three-dimensional strategy vector of the ad content to be delivered, the ad user, and the ad delivery environment; and inputs the three-dimensional strategy vector into the action generation module to obtain style presentation action, time optimization action, and bidding adjustment action; The second output layer receives the style presentation action, time optimization action, and bidding adjustment action, and performs ad style selection, ad time period selection, and real-time bidding adjustment to finally obtain a refined delivery strategy.

2. The intelligent advertising placement decision-making method according to claim 1, characterized in that, A feature set for user behavior models is constructed based on diverse user behavior data, including: Standardize and normalize user behavior data during preprocessing; The preprocessed user multi-behavioral data is divided into multiple behavioral sequences according to time windows; wherein, the behavioral sequence includes user multi-behavioral data slices, first timestamps, behavioral statistics, behavioral tags, and sequence dependencies; Each of the aforementioned behavior sequences is subjected to feature extraction to form multiple subsets of user behavior features; Multiple subsets of user behavior features are merged to form a user behavior model feature set.

3. The intelligent advertising placement decision-making method according to claim 2, characterized in that, The user behavior feature set is formed by fusing multiple subsets of the aforementioned user behavior features, including: Multiple subsets of user behavior features are sequentially aligned and aggregated to form multiple aggregated features; Multiple aggregated features are concatenated to form the user behavior model feature set; wherein, the user behavior model feature set includes behavioral statistical features, behavioral label features, sequence dependency features, time window features, and derived features.

4. The intelligent advertising placement decision-making method according to claim 1, characterized in that, A dynamic user knowledge graph is constructed based on diverse user behavior data, including: Based on the user's diverse behavioral data, entities and the relationships between entities are extracted. A user knowledge graph is constructed based on a graph database according to the entities and the relationships between the entities, wherein the nodes of the user knowledge graph are the entities, and the edges of the user knowledge graph are the relationships between the entities; The user knowledge graph is updated at a preset interval, and a second timestamp is marked in the user knowledge graph to form a dynamic user knowledge graph.

5. The intelligent advertising placement decision-making method according to claim 4, characterized in that, A user digital feature representation set is formed based on the user behavior model feature set and the user dynamic knowledge graph, including: The relational features and semantic features of the user dynamic knowledge graph are extracted, wherein the relational features are obtained by one-hot encoding the relational strength value and relational type between the entities, and the semantic features are obtained by transforming the semantic information of the entities; The features, relational features, and semantic features in the user behavior model feature set are mapped and aligned to a unified feature space to have the same feature dimension, and a high-dimensional dense feature vector is generated based on the Transformer encoder. The high-dimensional dense feature vector is combined with the user identifier and the third timestamp to form a user digital feature representation set, wherein the third timestamp represents the time point at which the user digital feature representation set was generated.

6. The intelligent advertising placement decision-making method according to claim 1, characterized in that, The advertising feature set of the advertisement to be placed includes a subset of advertising content features, a subset of advertising historical performance features, and a subset of advertising targeting features. The environmental information feature set of the current placement environment includes a subset of platform traffic features, a subset of market competition features, a subset of user activity features, and a subset of external event features.

7. An intelligent advertising placement decision system, characterized in that, The method for intelligent advertising delivery decision-making as described in any one of claims 1-6 includes: The construction unit is used to build a user behavior model feature set and a user dynamic knowledge graph based on user multi-dimensional behavior data; wherein, the user multi-dimensional behavior data includes user search behavior data, user social behavior data, user consumption behavior data, user content interaction behavior data, and user application usage behavior data; The user digital feature representation set forming unit forms a user digital feature representation set based on the user behavior model feature set and the user dynamic knowledge graph. The advertising strategy delivery unit is used to generate a delivery strategy for the advertisement to be delivered based on the advertising feature set of the advertisement to be delivered, the environmental information feature set of the current delivery environment, and the user digital feature representation set, using a pre-trained deep learning network; wherein the delivery strategy includes a macro delivery strategy and a fine delivery strategy.

Images