A knowledge graph-based digital advertisement intelligent delivery method

Abstract

This invention discloses a knowledge graph-based intelligent digital advertising delivery method, comprising the following steps: constructing an advertising delivery knowledge graph based on delivery relationships, determining target conversion nodes and environment nodes; generating candidate exposure action slots according to delivery tasks, and extracting local delivery cut-out diagrams; tracing borrowed paths, removing borrowed path edges, and obtaining isolated cut-out diagrams; writing to exposure lockhole nodes to form lockhole exposure diagrams; performing KGCN convolution on the tri-state graph to obtain tri-state convolution responses; extracting lockhole incremental responses, and forming lockhole accountability paths through foldback signals; disconnecting lockhole edges and performing tearing reversal convolution to generate delivery strategies and feed back real delivery feedback. This invention relates to the field of intelligent digital advertising delivery technology, achieving accurate digital advertising accountability, improving delivery conversion efficiency and feedback utilization quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent digital advertising delivery technology, and in particular to an intelligent digital advertising delivery method based on knowledge graphs. Background Technology

[0002] Intelligent digital advertising delivery technology has been widely used to match ad creatives, user groups, delivery scenarios, and conversion goals. Existing methods largely rely on user profiles, historical click records, ad tags, contextual features, and conversion data to build recommendation models, obtaining delivery results through similarity calculations, ranking models, graph neural networks, or knowledge graph reasoning. Knowledge graphs can express the relationships between ads, users, scenarios, behaviors, and conversion events; KGCN can aggregate neighborhood information within a graph structure to improve the accuracy of ad delivery matching.

[0003] Existing knowledge graph-based ad delivery methods often directly aggregate neighborhoods using existing edges in the graph when processing historical paths and candidate exposure actions. This easily leads to the mixing of contributions from historical conversion paths, environmentally related paths, and current exposure actions in the same convolutional result. There is a lack of structured verification to determine whether candidate exposure actions truly contribute to the target conversion, and the interference of bypass paths on conversion results is difficult to isolate. When writing back feedback data, global updates or weight adjustments are often used, making it difficult to accurately feed feedback back to verified responsible paths. Subsequent deliveries may still reuse inefficient or interfered candidate exposure actions.

[0004] Therefore, how to provide a knowledge graph-based intelligent digital advertising delivery method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based intelligent digital advertising delivery method. This invention utilizes KGCN convolution, roadside hooking, exposure keyhole node, and tearing counter-evidence convolution to complete the attribution verification of advertising delivery paths, and has the advantages of clear delivery attribution, isolation of interference paths, and accurate strategy generation.

[0006] A method for intelligent digital advertising delivery based on knowledge graphs according to an embodiment of the present invention includes the following steps: Map the delivery relationship to an advertising delivery knowledge graph, and mark the target conversion node and the environment node; Candidate exposure action slots are generated based on the delivery task, and a local delivery cutout diagram containing candidate exposure action slots and environmental nodes is extracted from the advertising delivery knowledge graph. Tracing the bypass path to the target conversion node by bypassing the candidate exposure action slot, detaching the bypass edge to the bypass evidence cache area, and obtaining the isolation cut diagram; Write the exposure lock hole node into the isolation cutout diagram. Connect the candidate exposure action slot to the inlet side and the environmental node to the outlet side. Connect them through the lock hole edge to form the lock hole exposure diagram. The local projection cutout image, the isolated cutout image, and the keyhole exposure image are used as three-state images to perform KGCN convolution to obtain the three-state convolution response; The incremental response of the keyhole is extracted from the three-state convolution response. When the incremental response of the keyhole is written into the target conversion node, a backtracking signal is initiated. When the backtracking signal bypasses the borrowing evidence buffer area and returns to the candidate exposure action slot via the keyhole edge, the keyhole attribution path is obtained. The lock hole edge in the lock hole responsibility path is disconnected and a tear-and-contrast convolution is performed. When the lock hole incremental response is still written to the target conversion node after the tear-and-contrast convolution, the candidate exposure action slot is locked. When the lock hole incremental response stops being written to the target conversion node after the tear-and-contrast convolution, the lock hole responsibility path is marked as a responsibility edge. The digital advertising intelligent delivery strategy is obtained, real delivery feedback is collected, and fed back to the responsibility edge.

[0007] Optionally, the step of generating candidate exposure action slots according to the delivery task and extracting a local delivery cutout diagram containing candidate exposure action slots and environment nodes from the advertising delivery knowledge graph specifically includes: Write the campaign tasks into the advertising knowledge graph to form a campaign entry point, and bind the campaign entry point to the target conversion node; Taking the target conversion node as the starting point of the reverse traction, the reverse traction branch is generated along the delivery relationship to obtain the conversion reverse traction path; Write a slot break mark for the reverse traction branch that does not hit the environmental node in the conversion reverse traction path, and write a slot formation mark for the reverse traction branch that hits the environmental node and points back to the delivery inlet along the delivery relationship. Follow the branch marked by the slot to return from the environment node to the target conversion node, write the delivery entry into the slot entry position, write the environment node into the environment snap position, and write the target conversion node into the conversion return position to form a candidate exposure action slot; Extract the delivery relationship edge enclosed by the slot inlet, environment latch, and conversion back pointer, and write the delivery relationship edge, environment node, and candidate exposure action slot into the local delivery cutout diagram.

[0008] Optionally, obtaining the isolation cut diagram specifically includes: In the local delivery cutout map, write a slot-bypass no-entry marker with the candidate exposure action slot as the cutoff point, and initiate dual-end tracking from the environmental node and the target conversion node along the delivery relationship edge, and combine the dual-end tracking results to form a candidate access path. Candidate access paths that do not pass through candidate exposure action slots marked with bypassing slots and connect environmental nodes and target transformation nodes are written into the borrowed path set. For each borrowed path in the borrowed path set, read the first incoming transformation edge that is directly connected to the target transformation node, and mark the first incoming transformation edge as the borrowed path edge to be removed; Remove the edge to be removed from the edge to be included in the KGCN convolution in the local delivery cut graph, and retain the node positions of the two ends of the edge to be removed in the local delivery cut graph. Write the removed roadside hangings into the roadside evidence cache area and mark them as roadside hangings. Based on the node locations and remaining delivery relationships, the local delivery cut-out diagram is reorganized to obtain the isolation cut-out diagram.

[0009] Optionally, the step of writing exposure lock hole nodes into the isolation cutout diagram, with the candidate exposure action slot connected to the inlet side and the environmental node connected to the outlet side, and connected via the lock hole edge to form a lock hole exposure diagram, specifically includes: Read the candidate exposure action slots, environment nodes, and remaining delivery relationship edges in the isolation cutout diagram; Write an exposure lock hole node between the candidate exposure action slot and the environment node, write the inlet side and the outlet side within the exposure lock hole node, and write the lock hole edge between the inlet side and the outlet side. Write an action interlocking edge between the inlet side and the candidate exposure action slot, and write an environment interlocking edge between the outlet side and the environment node. Connect the action fastening edge, the keyhole edge, and the environmental fastening edge in sequence from the entrance side to the exit side to form a keyhole access path; Write the keyhole access path and the remaining delivery relationship edges together into the isolation cutout graph, and write the connection status between the borrowing edge in the borrowing evidence buffer area and the keyhole access path as a no-reconnection state to obtain the keyhole exposure graph.

[0010] Optionally, obtaining the three-state convolutional response specifically includes: Read the node position corresponding to the side to be removed from the road and write the edge position after the side to be removed from the road into the empty space of the road. The borrowed space is synchronously written into the local delivery cutout map, the isolation cutout map, and the keyhole exposure map. A three-state isotopic convolution sequence is established according to the node position, the borrowed space, the remaining delivery relationship edge, and the keyhole access path. In the local delivery cut diagram, connect the borrowing edge in the borrowing evidence buffer area, disconnect the borrowing space in the isolation cut diagram, disconnect the borrowing space in the keyhole exposure diagram and write the corresponding edge of the keyhole access path into the keyhole release position. When the candidate exposure action slot completes the inlet-side latching and the environment node completes the outlet-side latching, the empty slot code of the lock hole release position is replaced with the relationship code corresponding to the lock hole edge; KGCN convolution is performed along the three-state isotopic convolution sequence, and the original state response, isolated state response, keyhole state response and three-state convolution response are obtained by aligning them according to the target transformation node.

[0011] Optionally, obtaining the keyhole attribution path specifically includes: Read the isolation state response and keyhole state response corresponding to the target transformation node from the three-state convolution response; Using the isolation state response as the subtraction benchmark, differential subtraction is performed on the keyhole state response to obtain the response difference; Negative truncation is performed on the response difference, and response differences less than zero are written as zero to obtain the keyhole incremental response; Write the incremental response of the keyhole into the incremental response bit of the target conversion node; when the incremental response bit is in the write state, initiate a back signal with the target conversion node as the back start point; The signal is blocked from entering the borrowing evidence buffer area. The signal is driven to enter the keyhole access path through the delivery relationship edge between the target conversion node and the environment node, and then return to the candidate exposure action slot through the environment snap-in edge, the keyhole edge and the action snap-in edge. Write the path taken by the return signal into the keyhole responsibility path.

[0012] Optionally, the keyhole edge in the disconnected keyhole attribution path undergoes a tearing rebuttal convolution, specifically including: Read the environmental latching edge, latching edge, action latching edge, candidate exposure action slot and target conversion node in the latching hole accountability path; Disconnect the keyhole edge from the keyhole accountability path, while retaining the node positions of the two ends of the keyhole edge in the keyhole exposure image; After the keyhole edge is broken, the corresponding edge position is written into the tearing space, and the tearing space replaces the keyhole release position in the three-state isotopic convolution sequence. Write edge break codes into the tear gaps to block the relationship codes corresponding to the keyhole edges from entering the KGCN convolution; Perform KGCN convolution along the replaced three-state isotopic convolution sequence, read the tearing counter-evidence response of the target transformation node, and perform isotopic comparison with the incremental response of the keyhole before the keyhole edge is disconnected; The results of the peer comparison show that the incremental response of the keyhole is still written into the target conversion node in the tearing counter-evidence response, which locks the candidate exposure action slot; The peer comparison results show that the keyhole incremental response marks the keyhole attribution path as a responsible edge when it stops writing to the target transformation node in the tearing counter-evidence response.

[0013] Optionally, marking the keyhole accountability path as a responsibility edge to obtain a digital advertising intelligent delivery strategy specifically includes: Read the lock hole attribution path corresponding to the candidate exposure action slot that has not been written to the locked state; The lock hole responsibility path that stops writing to the target transformation node in the tearing counter-evidence response, as shown by the peer comparison results, is marked as a responsibility edge and written into the responsibility edge set. Backtrack along each responsibility edge in the responsibility edge set to the candidate exposure action slot, and extract the environment node and target transformation node corresponding to the candidate exposure action slot; The responsibility edges associated with the same candidate exposure action slot are sorted according to the order in which the keyhole incremental response is written to the target conversion node, and responsibility edges that are not consecutively arranged along the same keyhole responsibility path are deleted to obtain the responsibility edge sequence. The responsibility edge sequence is written sequentially into the delivery confirmation bit of the candidate exposure action slot. When the delivery confirmation bit receives the sequentially written responsibility edge sequence, a delivery action is generated. Intelligent digital advertising delivery strategies are generated based on available actions, environmental nodes, and target conversion nodes.

[0014] Optionally, the collection of actual delivery feedback and its feedback to the responsible side specifically includes: After implementing the intelligent digital advertising delivery strategy, collect real delivery feedback based on the available delivery actions; By binding real delivery feedback with deliverable actions, environmental nodes, and target conversion nodes, feedback binding records are obtained; Tracing back along the feedback binding record to the delivery confirmation position of the candidate exposure action slot, and reading the responsibility edge sequence within the delivery confirmation position; Write the actual delivery feedback into the feedback feedback feed position of the responsible edge according to the sequence order of the responsible edge corresponding to the feedback binding record; When the feedback feedback position receives the actual delivery feedback written to the target conversion node, it marks the responsible edge as a positive feedback edge; When the feedback feed receives real delivery feedback that has not been written to the target conversion node, the responsible edge is marked as a feed-back edge to be reviewed, and a review mapping is established between the feed-back edge to be reviewed and the corresponding feed-back edge in the feed-back evidence cache area. Write the positive backfill edges, backfill edges to be reviewed, and review mappings back into the advertising knowledge graph.

[0015] Optionally, the subsequent process of writing the positive backfeed edges, backfeed edges to be reviewed, and review mappings back to the advertising delivery knowledge graph specifically includes: Read the positive backflow edges, backflow edges to be reviewed, and review mappings from the advertising knowledge graph; Write the positive confirmation bit of the corresponding responsible edge into the positive backfeed edge, and keep the release confirmation bit of the candidate exposure action slot where the corresponding responsible edge of the positive backfeed edge is located. Read the borrowed edge corresponding to the recharge edge to be reviewed along the review mapping, and write the recharge edge to be reviewed into the review lock position of the candidate exposure action slot where the corresponding responsible edge is located; When the lockout position is checked and a write state is formed, the candidate exposure action slot where the corresponding responsible edge is located is written into the no-entry slot; When generating candidate exposure action slots in the next round of deployment tasks, read the no-entry slots, write the connection status between the candidate exposure action slots corresponding to the no-entry slots and the next round of deployment tasks as a no-entry disconnection state, and block the candidate exposure action slots corresponding to the no-entry slots from entering the local deployment cutout map.

[0016] The beneficial effects of this invention are: (1) This invention moves the historical conversion shortcuts that bypass the candidate exposure action slots to reach the target conversion node into the borrowing evidence cache area by borrowing path tracking and borrowing edge decoupling, so that KGCN convolution no longer directly mixes the historical borrowing contribution with the current exposure action contribution. Compared with conventional improvements such as simply adding user profile fields, expanding advertising tags, and increasing the number of model layers, this invention can isolate interference paths at the graph structure level and reduce the misleading influence of historical high-frequency conversion paths on the current delivery judgment.

[0017] (2) This invention uses the exposure keyhole node, keyhole edge, keyhole release position and three-state co-position convolution sequence to perform KGCN convolution on the same side structure to obtain the original state response, isolation state response and keyhole state response that can be aligned. Compared with ordinary single-image convolution prediction, this invention can extract the keyhole incremental response from the three-state convolution response, so that the incremental contribution of the candidate exposure action slot to the target conversion node has a comparable and locatable structural basis.

[0018] (3) This invention uses the return signal, the keyhole accountability path, and the tearing rebuttal convolution to perform reverse accountability and edge break rebuttal on the incremental response of the keyhole; after the keyhole edge is broken, if the incremental response of the keyhole stops writing to the target conversion node, the keyhole accountability path is marked as the accountability edge; if the incremental response of the keyhole is still writing to the target conversion node, the candidate exposure action slot is locked; compared with the method of generating the delivery strategy based only on the click rate, conversion rate or predicted score, this invention can eliminate the false conversion contribution caused by non-exposure actions, so that the real delivery feedback is accurately fed back to the accountability edge, and reduce the repeated use of inefficient candidate exposure actions in subsequent delivery. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based intelligent digital advertising delivery method proposed in this invention; Figure 2 This is a schematic diagram of the process of detaching and attaching the bypass path and the exposure of the keyhole proposed in this invention; Figure 3This is a schematic diagram of the tri-state convolution, foldback attribution, and tearing counter-evidence proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A knowledge graph-based intelligent digital advertising delivery method includes the following steps: Map the delivery relationship to an advertising delivery knowledge graph, and mark the target conversion node and the environment node; In this invention, the delivery relationships are derived from ad delivery logs, exposure logs, click logs, conversion feedback records, delivery configuration records, and environment collection records. Ad objects, delivery tasks, exposure actions, delivery environments, click behaviors, and conversion behaviors are written into graph nodes. The associations between ad objects and exposure actions, exposure actions and delivery environments, delivery environments and click behaviors, and click behaviors and conversion behaviors are written into delivery relationship edges, forming an ad delivery knowledge graph. The graph node corresponding to the conversion behavior specified by the delivery task is marked as the target conversion node, and the graph node corresponding to the delivery environment of the delivery task is marked as the environment node. Candidate exposure action slots are generated based on the delivery task, and a local delivery cutout diagram containing candidate exposure action slots and environmental nodes is extracted from the advertising delivery knowledge graph. In this invention, the delivery task includes the advertising object, the target conversion behavior, and the delivery environment. The advertising object, the target conversion behavior, and the delivery environment are derived from the advertising delivery configuration record. The candidate exposure action slot is the in-graph position in the advertising delivery knowledge graph that carries the exposure action to be verified. It is formed by the connected delivery relationship edges between the advertising object, the environment node, and the target conversion node. When generating the local delivery cut-out graph, the candidate exposure action slot, the environment node, the target conversion node, and the directly connected delivery relationship edges are read from the advertising delivery knowledge graph. The read graph nodes and delivery relationship edges are written into the same local graph structure to obtain the local delivery cut-out graph containing the candidate exposure action slot and the environment node. Tracing the bypass path to the target conversion node by bypassing the candidate exposure action slot, detaching the bypass edge to the bypass evidence cache area, and obtaining the isolation cut diagram; In this invention, the borrowed path is a sequence of delivery relationship edges in the local delivery cut diagram that bypasses the candidate exposure action slot and connects the environment node and the target conversion node; the borrowed edge is the in-conversion edge in the borrowed path that is connected to the target conversion node; the borrowed evidence cache area is an in-graph cache structure that stores the detached borrowed edges; detaching the borrowed edge means removing the borrowed edge from the delivery relationship edges in the local delivery cut diagram, writing it into the borrowed evidence cache area, retaining the corresponding node position, and then reorganizing the remaining delivery relationship edges to generate the isolation cut diagram; Write the exposure lock hole node into the isolation cutout diagram. Connect the candidate exposure action slot to the inlet side and the environmental node to the outlet side. Connect them through the lock hole edge to form the lock hole exposure diagram. In this invention, the exposure lock hole node is an intermediate graph node written into the isolation cut diagram, formed by writing nodes between the candidate exposure action slot and the environment node; the entrance side is the connection end of the exposure lock hole node facing the candidate exposure action slot, and the exit side is the connection end of the exposure lock hole node facing the environment node; the snap-fit ​​is the operation of writing a connection edge between two graph nodes, when the entrance side snaps into the candidate exposure action slot, an action snap-fit ​​edge is written, and when the exit side snaps into the environment node, an environment snap-fit ​​edge is written; the lock hole edge is the graph edge connecting the entrance side and the exit side; the lock hole exposure diagram is the graph structure formed after writing the exposure lock hole node, action snap-fit ​​edge, environment snap-fit ​​edge, and lock hole edge into the isolation cut diagram; The local projection cutout image, the isolated cutout image, and the keyhole exposure image are used as three-state images to perform KGCN convolution to obtain the three-state convolution response; In this invention, the three-state graph consists of three graph structures: a local delivery cutout graph, an isolated cutout graph, and a keyhole exposure graph. KGCN convolution is a neighborhood convolution process based on a knowledge graph, used to aggregate features of adjacent graph nodes along the delivery relationship edges. The three graph structures are aligned using the same node positions and the same target transformation nodes to obtain the original state response, the isolated state response, and the keyhole state response, respectively. The three-state convolution response is the set of responses formed by the original state response, the isolated state response, and the keyhole state response at the target transformation node. The incremental response of the keyhole is extracted from the three-state convolution response. When the incremental response of the keyhole is written into the target conversion node, a backtracking signal is initiated. When the backtracking signal bypasses the borrowing evidence buffer area and returns to the candidate exposure action slot via the keyhole edge, the keyhole attribution path is obtained. In this invention, the keyhole incremental response is the positive response difference formed by the keyhole state response and the isolation state response; the return signal is the reverse tracking signal issued by the target conversion node after the incremental response bit forms a write state, which is used to check the candidate exposure action slot along the keyhole access path; the keyhole accountability path is the path record formed by the return signal starting from the target conversion node, avoiding the borrowed evidence buffer area, passing through the environment node, the environment latching edge, the keyhole edge, and the action latching edge, and returning to the candidate exposure action slot. The lock hole edge in the lock hole responsibility path is disconnected and a tear-and-contrast convolution is performed. When the lock hole incremental response is still written to the target conversion node after the tear-and-contrast convolution, the candidate exposure action slot is locked. When the lock hole incremental response stops being written to the target conversion node after the tear-and-contrast convolution, the lock hole responsibility path is marked as a responsibility edge. The digital advertising intelligent delivery strategy is obtained, real delivery feedback is collected, and fed back to the responsibility edge.

[0022] In this invention, the tear-and-contrast convolution is a KGCN convolution process performed after disconnecting the lockhole edge in the lockhole accountability path. It is used to verify whether the lockhole incremental response received by the target conversion node depends on the lockhole edge. Locking the candidate exposure action slot involves writing the candidate exposure action slot into a locked state. The locked state indicates that the candidate exposure action slot does not generate a deliverable action and does not enter the digital advertising intelligent delivery strategy. The accountability edge is the delivery accountability edge retained after the lockhole accountability path is verified by the tear-and-contrast convolution. The actual delivery feedback is the exposure record, click record, and conversion feedback record collected after the execution of the digital advertising intelligent delivery strategy. Feedback to the accountability edge involves writing the actual delivery feedback into the feedback feedback position of the accountability edge to record the actual delivery result corresponding to the accountability edge.

[0023] In this embodiment, candidate exposure action slots are generated according to the delivery task, and a partial delivery cutout diagram containing candidate exposure action slots and environment nodes is extracted from the advertising delivery knowledge graph, specifically including: Write the campaign tasks into the advertising knowledge graph to form a campaign entry point, and bind the campaign entry point to the target conversion node; In this invention, the entry point is the starting graph node formed after the delivery task is written into the advertising delivery knowledge graph; when the entry point is bound to the target conversion node, a delivery relationship edge is written between the entry point and the target conversion node. Taking the target conversion node as the starting point of the reverse traction, the reverse traction branch is generated along the delivery relationship to obtain the conversion reverse traction path; In this invention, the reverse traction starting point is the target conversion node; the reverse traction branch is a graph path that starts from the target conversion node and reads backward along the delivery relationship edge to the environment node and the delivery entrance; the conversion reverse traction path is formed by the reverse traction branch and is used to determine the path that can be pointed back from the target conversion node to the delivery entrance; Write a slot break mark for the reverse traction branch that does not hit the environmental node in the conversion reverse traction path, and write a slot formation mark for the reverse traction branch that hits the environmental node and points back to the delivery inlet along the delivery relationship. In this invention, the slot break mark is an exclusion mark written on the reverse traction branch that has not hit the environment node, and the reverse traction branch with the slot break mark does not generate candidate exposure action slots; the slot formation mark is a formation mark written on the reverse traction branch that has hit the environment node and can point back to the delivery inlet, and the reverse traction branch with the slot formation mark is used to form candidate exposure action slots. Follow the branch marked by the slot to return from the environment node to the target conversion node, write the delivery entry into the slot entry position, write the environment node into the environment snap position, and write the target conversion node into the conversion return position to form a candidate exposure action slot; In this invention, the slot inlet position is the position where the delivery inlet is written in the candidate exposure action slot, the environment latch position is the position where the environment node is written in the candidate exposure action slot, and the conversion back pointer position is the position where the target conversion node is written in the candidate exposure action slot. Extract the delivery relationship edge enclosed by the slot inlet, environment latch, and conversion back pointer, and write the delivery relationship edge, environment node, and candidate exposure action slot into the local delivery cutout diagram.

[0024] In this embodiment, the isolation cut diagram is obtained, specifically including: In the local delivery cutout map, write a slot-bypass no-entry marker with the candidate exposure action slot as the cutoff point, and initiate dual-end tracking from the environmental node and the target conversion node along the delivery relationship edge, and combine the dual-end tracking results to form a candidate access path. In this invention, the blocking point is a location in the graph that is prohibited from being traversed during the path reading process; the slot-bypass prohibition mark is an exclusion mark written on the candidate exposure action slot to prevent the tracking path from passing through the candidate exposure action slot; dual-end tracking is the operation of reading the path from the environment node and the target conversion node along the delivery relationship edge respectively; when combining the dual-end tracking results, the common graph node tracked by both ends is used as the connection position, and the path from the environment node to the common graph node and the path from the common graph node to the target conversion node are connected to form a candidate accessible path; Candidate access paths that do not pass through candidate exposure action slots marked with bypassing slots and connect environmental nodes and target transformation nodes are written into the borrowed path set. In this invention, the borrowed path set is a set of candidate access paths that bypass the candidate exposure action slot and connect the environment node and the target conversion node. The candidate access paths written into the borrowed path set are denoted as borrowed paths. For each borrowed path in the borrowed path set, read the first incoming transformation edge that is directly connected to the target transformation node, and mark the first incoming transformation edge as the borrowed path edge to be removed; In this invention, the first inbound conversion edge is the delivery relationship edge that enters the target conversion node through the borrowed path; Remove the edge to be removed from the edge to be included in the KGCN convolution in the local delivery cut graph, and retain the node positions of the two ends of the edge to be removed in the local delivery cut graph. Write the removed roadside hangings into the roadside evidence cache area and mark them as roadside hangings. Based on the node locations and remaining delivery relationships, the local delivery cut-out diagram is reorganized to obtain the isolation cut-out diagram.

[0025] In this invention, the remaining placement relationship edges are the placement relationship edges that are retained in the local placement cut diagram after the side to be removed from the hooking / borrowing path.

[0026] In this embodiment, an exposure lock hole node is written into the isolation cutout diagram. The candidate exposure action slot is connected to the inlet side, and the environment node is connected to the outlet side. The lock hole is then connected to form a lock hole exposure diagram. Specifically, this includes: Read the candidate exposure action slots, environment nodes, and remaining delivery relationship edges in the isolation cutout diagram; Write an exposure lock hole node between the candidate exposure action slot and the environment node, write the inlet side and the outlet side within the exposure lock hole node, and write the lock hole edge between the inlet side and the outlet side. Write an action interlocking edge between the inlet side and the candidate exposure action slot, and write an environment interlocking edge between the outlet side and the environment node. In this invention, the action latching edge is the connection edge written between the inlet side and the candidate exposure action slot, and the environment latching edge is the connection edge written between the outlet side and the environment node. Connect the action fastening edge, the keyhole edge, and the environmental fastening edge in sequence from the entrance side to the exit side to form a keyhole access path; Write the keyhole access path and the remaining delivery relationship edges together into the isolation cutout graph, and write the connection status between the borrowing edge in the borrowing evidence buffer area and the keyhole access path as a no-reconnection state to obtain the keyhole exposure graph.

[0027] In this invention, the "no reconnection" state is a disconnection marker between the borrowing edge and the keyhole access path within the borrowing evidence buffer area. After writing the "no reconnection" state, the borrowing edge will not be connected to the keyhole access path.

[0028] In this embodiment, the tri-state convolution response is obtained, specifically including: Read the node position corresponding to the side to be removed from the road and write the edge position after the side to be removed from the road into the empty space of the road. In this invention, the borrowed space is the occupant position left on the original edge after the borrowed edge to be removed, which is used to maintain the edge position correspondence between the local deployment cut diagram, the isolation cut diagram and the keyhole exposure diagram; The borrowed space is synchronously written into the local delivery cutout map, the isolation cutout map, and the keyhole exposure map. A three-state isotopic convolution sequence is established according to the node position, the borrowed space, the remaining delivery relationship edge, and the keyhole access path. In this invention, the tri-state co-position convolution sequence is a convolution input sequence formed by arranging three graph structures according to the same node position and the same edge position order, which is used to align the responses of the three graph structures at the same target transformation node; In the local delivery cut diagram, connect the borrowing edge in the borrowing evidence buffer area, disconnect the borrowing space in the isolation cut diagram, disconnect the borrowing space in the keyhole exposure diagram and write the corresponding edge of the keyhole access path into the keyhole release position. In this invention, the keyhole release position is the occupant position in the keyhole exposure image used to access the corresponding edge of the keyhole access path; When the candidate exposure action slot completes the inlet-side latching and the environment node completes the outlet-side latching, the empty slot code of the lock hole release position is replaced with the relationship code corresponding to the lock hole edge; In this invention, the empty slot code is a code indicating that the edge is not connected, and its value is 0; the edge type field is a field that is recorded synchronously when the placement relationship edge is written into the advertising placement knowledge graph, and is used to distinguish the placement relationship edge types between the advertising object and the exposure action, the exposure action and the placement environment, the placement environment and the click behavior, and the click behavior and the conversion behavior; the relationship code is a code indicating that the corresponding placement relationship edge or keyhole edge is connected. The relationship code of the placement relationship edge is taken from the number of the corresponding edge type field in the advertising placement knowledge graph, and the relationship code of the keyhole edge is taken from the keyhole edge type number generated when the keyhole edge is written. The keyhole edge type number is a positive integer number that is not occupied by the existing edge type field in the advertising placement knowledge graph. KGCN convolution is performed along the three-state isotopic convolution sequence, and the original state response, isolated state response, keyhole state response and three-state convolution response are obtained by aligning them according to the target transformation node.

[0029] In this invention, the original state response is the convolutional output formed at the target conversion node after the local projection cutout image connects the bypass edge; the isolation state response is the convolutional output formed at the target conversion node after the isolation cutout image disconnects the bypass space; and the keyhole state response is the convolutional output formed at the target conversion node after the keyhole edge in the keyhole exposure image enters KGCN convolution. The three-state convolutional response is formed by aligning the original state response, the isolation state response, and the keyhole state response.

[0030] In this embodiment, the path to assign responsibility for the keyhole is obtained, specifically including: Read the isolation state response and keyhole state response corresponding to the target transformation node from the three-state convolution response; Using the isolation state response as the subtraction benchmark, differential subtraction is performed on the keyhole state response to obtain the response difference; In this invention, the subtraction benchmark is the reference response used for differential subtraction, which is the isolation state response corresponding to the target conversion node; the response difference is the difference obtained by subtracting the isolation state response from the keyhole state response at the same target conversion node. Negative truncation is performed on the response difference, and response differences less than zero are written as zero to obtain the keyhole incremental response; In this invention, negative truncation means writing the part of the response difference that is less than zero as zero, and retaining the part that is greater than zero as the keyhole incremental response; Write the incremental response of the keyhole into the incremental response bit of the target conversion node; when the incremental response bit is in the write state, initiate a back signal with the target conversion node as the back start point; In this invention, the incremental response bit is the position in the target conversion node used to write the keyhole incremental response; the write state is the state after the incremental response bit receives a non-zero keyhole incremental response; the return start point is the target conversion node that initiates the return signal. The signal is blocked from entering the borrowing evidence buffer area. The signal is driven to enter the keyhole access path through the delivery relationship edge between the target conversion node and the environment node, and then return to the candidate exposure action slot through the environment snap-in edge, the keyhole edge and the action snap-in edge. In this invention, blocking the return signal from entering the bypass evidence buffer area means setting the bypass evidence buffer area as an object prohibited from entering by the return signal; Write the path taken by the return signal into the keyhole responsibility path.

[0031] In this embodiment, performing a tearing rebuttal convolution on the keyhole edge in the keyhole attribution path specifically includes: Read the environmental latching edge, latching edge, action latching edge, candidate exposure action slot and target conversion node in the latching hole accountability path; Disconnect the keyhole edge from the keyhole accountability path, while retaining the node positions of the two ends of the keyhole edge in the keyhole exposure image; After the keyhole edge is broken, the corresponding edge position is written into the tearing space, and the tearing space replaces the keyhole release position in the three-state isotopic convolution sequence. In this invention, the tear gap is the occupant position left at the original edge after the keyhole edge is broken from the keyhole responsibility path, and is used to replace the keyhole release position in the three-state isotopic convolution sequence; Write edge break codes into the tear gaps to block the relationship codes corresponding to the keyhole edges from entering the KGCN convolution; In this invention, the broken edge code is an unconnected code that indicates that the keyhole edge has been disconnected. It uses the same value as the empty space code, which is 0. It is used to block the relation code corresponding to the keyhole edge from entering the KGCN convolution. Perform KGCN convolution along the replaced three-state isotopic convolution sequence, read the tearing counter-evidence response of the target transformation node, and perform isotopic comparison with the incremental response of the keyhole before the keyhole edge is disconnected; In this invention, the tearing counter-evidence response is the convolution output formed at the target conversion node after the tearing gap replaces the keyhole release position and is convolved by KGCN; the same position comparison is to compare whether there is still a keyhole incremental response before the keyhole edge is disconnected in the tearing counter-evidence response under the same target conversion node and the same side position sequence. The results of the peer comparison show that the incremental response of the keyhole is still written into the target conversion node in the tearing counter-evidence response, which locks the candidate exposure action slot; In this invention, locking the candidate exposure action slot means writing the candidate exposure action slot into a locked state. The locked state indicates that the candidate exposure action slot does not generate a deliverable action and does not enter the digital advertising intelligent delivery strategy. The peer comparison results show that the keyhole incremental response marks the keyhole attribution path as a responsible edge when it stops writing to the target transformation node in the tearing counter-evidence response.

[0032] In this invention, the responsibility edge is the responsibility mark made on the lock hole responsibility path after the lock hole incremental response stops being written to the target conversion node, as shown by the same position comparison result.

[0033] In this embodiment, the keyhole responsibility path is marked as a responsibility edge to obtain a digital advertising intelligent delivery strategy, which specifically includes: Read the lock hole attribution path corresponding to the candidate exposure action slot that has not been written to the locked state; The lock hole responsibility path that stops writing to the target transformation node in the tearing counter-evidence response, as shown by the peer comparison results, is marked as a responsibility edge and written into the responsibility edge set. In this invention, the set of responsible edges is an in-graph set that stores the responsible edges, which is derived from the labeling result of the keyhole responsibility path after tearing and reversing the convolution. Backtrack along each responsibility edge in the responsibility edge set to the candidate exposure action slot, and extract the environment node and target transformation node corresponding to the candidate exposure action slot; In this invention, backtracking is an operation that reads back along the connection direction of the responsibility edge to the candidate exposure action slot, which is used to obtain the candidate exposure action slot, environment node and target conversion node corresponding to the responsibility edge; The responsibility edges associated with the same candidate exposure action slot are sorted according to the order in which the keyhole incremental response is written to the target conversion node, and responsibility edges that are not consecutively arranged along the same keyhole responsibility path are deleted to obtain the responsibility edge sequence. In this invention, a continuous arrangement means that the responsible edges are adjacent to each other in the same keyhole responsibility path, and there are no unmarked placement relationship edges between adjacent responsible edges; a responsible edge that does not meet this condition is a responsible edge that is not continuously arranged along the same keyhole responsibility path. The responsibility edge sequence is the edge sequence obtained by arranging the responsibility edges associated with the same candidate exposure action slot according to the order in which the keyhole incremental response is written into the target conversion node. Responsibility edges that are not consecutively arranged along the same keyhole responsibility path are not written into the responsibility edge sequence. The responsibility edge sequence is written sequentially into the delivery confirmation bit of the candidate exposure action slot. When the delivery confirmation bit receives the sequentially written responsibility edge sequence, a delivery action is generated. In this invention, the delivery confirmation bit is the position in the candidate exposure action slot used to receive the responsibility edge sequence; the deliverable action is the exposure action generated after the delivery confirmation bit receives the responsibility edge sequence written in sequence. Intelligent digital advertising delivery strategies are generated based on available actions, environmental nodes, and target conversion nodes.

[0034] In this invention, the intelligent digital advertising delivery strategy is a delivery strategy record formed by a combination of deliverable actions, environmental nodes, and target conversion nodes.

[0035] In this implementation, collecting actual deployment feedback and feeding it back to the responsible side specifically includes: After implementing the intelligent digital advertising delivery strategy, collect real delivery feedback based on the available delivery actions; By binding real delivery feedback with deliverable actions, environmental nodes, and target conversion nodes, feedback binding records are obtained; In this invention, the feedback binding record is a record formed by establishing a correspondence between the actual delivery feedback and the deliverable actions, environmental nodes and target conversion nodes, and is used to trace back the delivery confirmation position of the candidate exposure action slot; Tracing back along the feedback binding record to the delivery confirmation position of the candidate exposure action slot, and reading the responsibility edge sequence within the delivery confirmation position; Write the actual delivery feedback into the feedback feedback feed position of the responsible edge according to the sequence order of the responsible edge corresponding to the feedback binding record; In this invention, the feedback feedback position is the location in the responsibility edge used to write the actual delivery feedback; When the feedback feedback position receives the actual delivery feedback written to the target conversion node, it marks the responsible edge as a positive feedback edge; When the feedback feed receives real delivery feedback that has not been written to the target conversion node, the responsible edge is marked as a feed-back edge to be reviewed, and a review mapping is established between the feed-back edge to be reviewed and the corresponding feed-back edge in the feed-back evidence cache area. In this invention, positive backfeed edge and backfeed edge to be reviewed are two types of backfeed marks formed after the responsible edge receives the actual delivery feedback; the review mapping is the correspondence between the backfeed edge to be reviewed and the borrowed edge in the borrowed evidence buffer area. The corresponding borrowed edge is the borrowed edge that is associated with the same candidate exposure action slot, the same environment node and the same target conversion node as the backfeed edge to be reviewed. Write the positive backfill edges, backfill edges to be reviewed, and review mappings back into the advertising knowledge graph.

[0036] In this embodiment, the subsequent process of writing the positive backfeeding edge, the backfeeding edge to be reviewed, and the review mapping back to the advertising placement knowledge graph specifically includes: Read the positive backflow edges, backflow edges to be reviewed, and review mappings from the advertising knowledge graph; Write the positive confirmation bit of the corresponding responsible edge into the positive backfeed edge, and keep the release confirmation bit of the candidate exposure action slot where the corresponding responsible edge of the positive backfeed edge is located. In this invention, the positive confirmation position is the position in the responsibility edge used to write the positive backfill edge; after the positive confirmation position is written to the positive backfill edge, the placement confirmation position of the candidate exposure action slot where the responsibility edge is located remains unchanged, and the responsibility edge sequence and the placement action are retained. Read the borrowed edge corresponding to the recharge edge to be reviewed along the review mapping, and write the recharge edge to be reviewed into the review lock position of the candidate exposure action slot where the corresponding responsible edge is located; In this invention, the verification lock position is the position in the candidate exposure action slot used to write the edge to be verified and refilled; when reading the borrowed edge corresponding to the edge to be verified and refilled along the verification map, the correspondence between the edge to be verified and the borrowed edge recorded in the verification map is used as the basis for reading; the verification lock position forms a write state, which is the state after the verification lock position receives the edge to be verified and refilled. When the lockout position is checked and a write state is formed, the candidate exposure action slot where the corresponding responsible edge is located is written into the no-entry slot; When generating candidate exposure action slots in the next round of deployment tasks, read the no-entry slots, write the connection status between the candidate exposure action slots corresponding to the no-entry slots and the next round of deployment tasks as a no-entry disconnection state, and block the candidate exposure action slots corresponding to the no-entry slots from entering the local deployment cutout map.

[0037] In this invention, the "no-entry slot" is the position used to write the corresponding candidate exposure action slot when the verification lock position forms a write state. The "no-entry disconnection state" is a disconnection marker between the candidate exposure action slot corresponding to the no-entry slot and the next round of deployment tasks. After writing the no-entry disconnection state, the candidate exposure action slot corresponding to the no-entry slot will not participate in the capture of the next round of local deployment cutout map.

[0038] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-channel intelligent advertising delivery scenario on a digital advertising platform. This scenario targets the same type of consumer goods advertising, with delivery channels including feed ads, search ads, content recommendation ads, and remarketing ads. The delivery objective is for users to complete a specified conversion behavior. In actual delivery, there are numerous related paths between the advertising target, exposure actions, delivery environment, click behavior, and conversion behavior. Some users experience multiple exposures and clicks before completing a conversion, and are also influenced by factors such as historical browsing, content recommendations, search term triggers, and remarketing outreach. Conventional knowledge graph-based delivery methods tend to attribute conversion results obtained through bypassed paths to the current exposure action, leading to inaccurate identification of the responsible party, resulting in duplicate exposures, inefficient consumption, and incorrect attribution.

[0039] In this scenario, ad delivery logs, impression logs, click logs, conversion feedback records, delivery configuration records, and environment collection records are organized into a unified data source. Ad objects, delivery tasks, impression actions, delivery environments, click behaviors, and conversion behaviors are written into graph nodes, and the relationships between different nodes are written into delivery relationship edges, forming an ad delivery knowledge graph. The conversion behavior specified by the delivery task is marked as a target conversion node, and the delivery environment involved in the delivery task is marked as an environment node. The system generates candidate impression action slots around the delivery task, and extracts candidate impression action slots, environment nodes, target conversion nodes, and connected delivery relationship edges from the ad delivery knowledge graph to form a partial delivery segment graph.

[0040] Within the local delivery cut-off graph, the system uses candidate exposure action slots as occlusion points and tracks bypass paths that still reach the target conversion node. Incoming conversion edges that directly enter the target conversion node from these bypass paths are detached and placed in the bypass evidence buffer, preserving the corresponding node positions. The remaining delivery relationship edges are then reassembled to form an isolated cut-off graph. This process isolates paths that could potentially cause attribution contamination from the convolutional input, preventing subsequent judgments from directly relying on such bypass relationships.

[0041] In the isolation cutout map, the system writes an exposure keyhole node between the candidate exposure action slot and the environment node. The exposure keyhole node is connected to the candidate exposure action slot via the inlet side and to the environment node via the outlet side, with the middle connected by keyhole edges, forming a keyhole exposure map. The local projection cutout map, isolation cutout map, and keyhole exposure map are fed into KGCN convolution as a three-state map. The three-state map uses the same node positions and edge order to obtain the original state response, isolation state response, and keyhole state response, respectively. The positive response difference between the keyhole state response and the isolation state response is taken as the keyhole incremental response.

[0042] After the keyhole incremental response is written to the target transformation node, the system initiates a backtracking signal from the target transformation node. The backtracking signal bypasses the borrowed evidence buffer and returns to the candidate exposure action slot via the environment node, environment latching edge, keyhole edge, and action latching edge, forming a keyhole accountability path. Then, the system disconnects the keyhole edge in the keyhole accountability path and performs a tear-and-reverse-evidence convolution. If the target transformation node still retains the keyhole incremental response after disconnecting the keyhole edge, it indicates that the transformation result does not depend on this keyhole path, and the candidate exposure action slot is locked, preventing the generation of any deployable action. If the keyhole incremental response stops being written to the target transformation node after disconnecting the keyhole edge, it indicates that the current exposure action has a verifiable contribution to the target transformation, and the keyhole accountability path is marked as an accountability edge.

[0043] After the responsibility edge is written to the delivery confirmation position of the candidate exposure action slot, the system generates a deliverable action and generates a digital advertising intelligent delivery strategy based on the deliverable action, environment node, and target conversion node. After the strategy is executed, the system collects real delivery feedback, including exposure records, click records, and conversion feedback records. The real delivery feedback is written to the feedback feedback position of the responsibility edge. If the feedback result is written to the target conversion node, the responsibility edge is marked as a positive feedback edge; if the feedback result is not written to the target conversion node, the responsibility edge is marked as a feedback edge pending review, and a review mapping is established with the corresponding borrowing edge in the borrowing evidence cache. After the feedback edge pending review enters the review lock position, the corresponding candidate exposure action slot is written to the forbidden slot. When the next round of delivery tasks generates candidate exposure action slots, the connection status between the candidate exposure action slot corresponding to the forbidden slot and the next round of delivery tasks is written to the forbidden disconnected state, and the corresponding candidate exposure action slot will no longer enter the local delivery segmentation graph.

[0044] In the validation sample, the conventional KGCN campaign and the campaign of this invention used the same campaign budget, the same ad creative pool, and the same target conversion metric. The conventional approach directly performs neighborhood convolution on the ad campaign knowledge graph, determining the exposure action based on the node response. The campaign of this invention, based on the same data, adds bypass path removal, exposure keyhole nodes, tri-state convolution, backtracking attribution, tear-off verification, and feedback backfeeding. The validation sample covers 1.08 million to 1.2 million exposures, retaining records of exposure, clicks, conversions, feedback backfeeding, and verification. The implementation results show that although the present invention slightly increases the strategy generation time, it significantly improves conversion rate, attribution accuracy, cost per conversion, and return on investment. This result indicates that the present invention does not simply increase model complexity, but rather reduces invalid campaigns caused by incorrect attribution by isolating bypass paths and verification keyhole attribution paths.

[0045] Table 1: Comparison of the Implementation Effects of Intelligent Digital Advertising Delivery

[0046] The table above presents structured data obtained from validating the conventional KGCN campaign and the campaign of this invention under the same campaign budget, ad creative pool, and target conversion metric. The data shows that this invention achieved higher click-through rate and conversion rate despite reduced exposure. Effective exposure decreased from 1,200,000 to 1,080,000, a reduction of 10.00%, but the effective click-through rate increased from 2.84% to 3.61%, and the target conversion rate increased from 0.42% to 0.67%. This indicates that this invention does not rely on increasing exposure scale to achieve conversion, but rather improves the correspondence quality between exposure actions and target conversion nodes through candidate exposure action slot selection and responsibility edge confirmation.

[0047] The interference rate of the borrowed path attribution decreased from 18.60% to 6.20%, and the false attribution rate decreased from 13.40% to 4.10%. This data reflects the effects of borrowed path edge removal, borrowed path evidence buffer, keyhole attribution path, and tear-off rebuttal convolution. Conventional KGCN delivery schemes easily misjudge transformation paths that bypass candidate exposure action slots as current exposure contributions. This invention separates the original state, isolated state, and keyhole state in the three-state diagram, and verifies this through keyhole incremental response and tear-off rebuttal response. It can lock out candidate exposure action slots that do not depend on the keyhole edge, reducing the interference of borrowed paths on the attribution results.

[0048] The cost per conversion decreased from RMB 86.40 to RMB 58.70, and the proportion of inefficient duplicate exposures decreased from 21.80% to 9.70%. This change indicates that after feedback feedback from actual campaigns is fed back, the edge to be reviewed can enter the exclusion control through the review lockout, preventing similar erroneous candidate exposure action slots from continuing to participate in the local campaign segmentation in the next round. Although the strategy generation time increased from 184ms to 211ms, the time is still within the acceptable range of the campaign system, and it has resulted in improved feedback accuracy and return on investment. The return on investment increased from 1.72 to 2.64, indicating that the present invention has achieved a better balance between feasibility and campaign revenue.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent digital advertising delivery based on knowledge graphs, characterized in that, Includes the following steps: Map the delivery relationship to an advertising delivery knowledge graph, and mark the target conversion node and the environment node; Candidate exposure action slots are generated based on the delivery task, and a local delivery cutout diagram containing candidate exposure action slots and environmental nodes is extracted from the advertising delivery knowledge graph. Tracing the bypass path to the target conversion node by bypassing the candidate exposure action slot, detaching the bypass edge to the bypass evidence cache area, and obtaining the isolation cut diagram; Write the exposure lock hole node into the isolation cutout diagram. Connect the candidate exposure action slot to the inlet side and the environmental node to the outlet side. Connect them through the lock hole edge to form the lock hole exposure diagram. The local projection cutout image, the isolated cutout image, and the keyhole exposure image are used as three-state images to perform KGCN convolution to obtain the three-state convolution response; The incremental response of the keyhole is extracted from the three-state convolution response. When the incremental response of the keyhole is written into the target conversion node, a backtracking signal is initiated. When the backtracking signal bypasses the borrowing evidence buffer area and returns to the candidate exposure action slot via the keyhole edge, the keyhole attribution path is obtained. The lock hole edge in the lock hole responsibility path is disconnected and a tear-and-contrast convolution is performed. When the lock hole incremental response is still written to the target conversion node after the tear-and-contrast convolution, the candidate exposure action slot is locked. When the lock hole incremental response stops being written to the target conversion node after the tear-and-contrast convolution, the lock hole responsibility path is marked as a responsibility edge. The digital advertising intelligent delivery strategy is obtained, real delivery feedback is collected, and fed back to the responsibility edge.

2. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 1, characterized in that, The process of generating candidate exposure action slots based on the delivery task, and extracting a local delivery cutout diagram containing candidate exposure action slots and environment nodes from the advertising delivery knowledge graph, specifically includes: Write the campaign tasks into the advertising knowledge graph to form a campaign entry point, and bind the campaign entry point to the target conversion node; Taking the target conversion node as the starting point of the reverse traction, the reverse traction branch is generated along the delivery relationship to obtain the conversion reverse traction path; Write a slot break mark for the reverse traction branch that does not hit the environmental node in the conversion reverse traction path, and write a slot formation mark for the reverse traction branch that hits the environmental node and points back to the delivery inlet along the delivery relationship. Follow the branch marked by the slot to return from the environment node to the target conversion node, write the delivery entry into the slot entry position, write the environment node into the environment snap position, and write the target conversion node into the conversion return position to form a candidate exposure action slot; Extract the delivery relationship edge enclosed by the slot inlet, environment latch, and conversion back pointer, and write the delivery relationship edge, environment node, and candidate exposure action slot into the local delivery cutout diagram.

3. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 1, characterized in that, The obtained isolation incision diagram specifically includes: In the local delivery cutout map, write a slot-bypass no-entry marker with the candidate exposure action slot as the cutoff point, and initiate dual-end tracking from the environmental node and the target conversion node along the delivery relationship edge, and combine the dual-end tracking results to form a candidate access path. Candidate access paths that do not pass through candidate exposure action slots marked with bypassing slots and connect environmental nodes and target transformation nodes are written into the borrowed path set. For each borrowed path in the borrowed path set, read the first incoming transformation edge that is directly connected to the target transformation node, and mark the first incoming transformation edge as the borrowed path edge to be removed; Remove the edge to be removed from the edge to be included in the KGCN convolution in the local delivery cut graph, and retain the node positions of the two ends of the edge to be removed in the local delivery cut graph. Write the removed roadside hangings into the roadside evidence cache area and mark them as roadside hangings. Based on the node locations and remaining delivery relationships, the local delivery cut-out diagram is reorganized to obtain the isolation cut-out diagram.

4. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 3, characterized in that, The process of writing exposure lock hole nodes into the isolation cutout diagram, with the inlet side connected to the candidate exposure action slot and the outlet side connected to the environment node, and connected via the lock hole edge to form a lock hole exposure diagram, specifically includes: Read the candidate exposure action slots, environment nodes, and remaining delivery relationship edges in the isolation cutout diagram; Write an exposure lock hole node between the candidate exposure action slot and the environment node, write the inlet side and the outlet side within the exposure lock hole node, and write the lock hole edge between the inlet side and the outlet side. Write an action interlocking edge between the inlet side and the candidate exposure action slot, and write an environment interlocking edge between the outlet side and the environment node. Connect the action fastening edge, the keyhole edge, and the environmental fastening edge in sequence from the entrance side to the exit side to form a keyhole access path; Write the keyhole access path and the remaining delivery relationship edges together into the isolation cutout graph, and write the connection status between the borrowing edge in the borrowing evidence buffer area and the keyhole access path as a no-reconnection state to obtain the keyhole exposure graph.

5. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 4, characterized in that, The obtained tri-state convolutional response specifically includes: Read the node position corresponding to the side to be removed from the road and write the edge position after the side to be removed from the road into the empty space of the road. The borrowed space is synchronously written into the local delivery cutout map, the isolation cutout map, and the keyhole exposure map. A three-state isotopic convolution sequence is established according to the node position, the borrowed space, the remaining delivery relationship edge, and the keyhole access path. In the local delivery cut diagram, connect the borrowing edge in the borrowing evidence buffer area, disconnect the borrowing space in the isolation cut diagram, disconnect the borrowing space in the keyhole exposure diagram and write the corresponding edge of the keyhole access path into the keyhole release position. When the candidate exposure action slot completes the inlet-side latching and the environment node completes the outlet-side latching, the empty slot code of the lock hole release position is replaced with the relationship code corresponding to the lock hole edge; KGCN convolution is performed along the three-state isotopic convolution sequence, and the original state response, isolated state response, keyhole state response and three-state convolution response are obtained by aligning them according to the target transformation node.

6. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 5, characterized in that, The obtained keyhole accountability path specifically includes: Read the isolation state response and keyhole state response corresponding to the target transformation node from the three-state convolution response; Using the isolation state response as the subtraction benchmark, differential subtraction is performed on the keyhole state response to obtain the response difference; Negative truncation is performed on the response difference, and response differences less than zero are written as zero to obtain the keyhole incremental response; Write the incremental response of the keyhole into the incremental response bit of the target conversion node; when the incremental response bit is in the write state, initiate a back signal with the target conversion node as the back start point; The signal is blocked from entering the borrowing evidence buffer area. The signal is driven to enter the keyhole access path through the delivery relationship edge between the target conversion node and the environment node, and then return to the candidate exposure action slot through the environment snap-in edge, the keyhole edge and the action snap-in edge. Write the path taken by the return signal into the keyhole responsibility path.

7. The method for intelligent digital advertising delivery based on knowledge graphs according to claim 6, characterized in that, The keyhole edge in the disconnected keyhole attribution path undergoes a tearing rebuttal convolution, specifically including: Read the environmental latching edge, latching edge, action latching edge, candidate exposure action slot and target conversion node in the latching hole accountability path; Disconnect the keyhole edge from the keyhole accountability path, while retaining the node positions of the two ends of the keyhole edge in the keyhole exposure image; After the keyhole edge is broken, the corresponding edge position is written into the tearing space, and the tearing space replaces the keyhole release position in the three-state isotopic convolution sequence. Write edge break codes into the tear gaps to block the relationship codes corresponding to the keyhole edges from entering the KGCN convolution; Perform KGCN convolution along the replaced three-state isotopic convolution sequence, read the tearing counter-evidence response of the target transformation node, and perform isotopic comparison with the incremental response of the keyhole before the keyhole edge is disconnected; The results of the peer comparison show that the incremental response of the keyhole is still written into the target conversion node in the tearing counter-evidence response, which locks the candidate exposure action slot; The peer comparison results show that the keyhole incremental response marks the keyhole attribution path as a responsible edge when it stops writing to the target transformation node in the tearing counter-evidence response.

8. The intelligent digital advertising delivery method based on knowledge graphs according to claim 7, characterized in that, The step of marking the keyhole accountability path as a responsibility edge to obtain the intelligent digital advertising delivery strategy specifically includes: Read the lock hole attribution path corresponding to the candidate exposure action slot that has not been written to the locked state; The lock hole responsibility path that stops writing to the target transformation node in the tearing counter-evidence response, as shown by the peer comparison results, is marked as a responsibility edge and written into the responsibility edge set. Backtrack along each responsibility edge in the responsibility edge set to the candidate exposure action slot, and extract the environment node and target transformation node corresponding to the candidate exposure action slot; The responsibility edges associated with the same candidate exposure action slot are sorted according to the order in which the keyhole incremental response is written to the target conversion node, and responsibility edges that are not consecutively arranged along the same keyhole responsibility path are deleted to obtain the responsibility edge sequence. The responsibility edge sequence is written sequentially into the delivery confirmation bit of the candidate exposure action slot. When the delivery confirmation bit receives the sequentially written responsibility edge sequence, a delivery action is generated. Intelligent digital advertising delivery strategies are generated based on available actions, environmental nodes, and target conversion nodes.

9. The intelligent digital advertising delivery method based on knowledge graphs according to claim 8, characterized in that, The collection of real-world delivery feedback, fed back to the responsible side, specifically includes: After implementing the intelligent digital advertising delivery strategy, collect real delivery feedback based on the available delivery actions; By binding real delivery feedback with deliverable actions, environmental nodes, and target conversion nodes, feedback binding records are obtained; Tracing back along the feedback binding record to the delivery confirmation position of the candidate exposure action slot, and reading the responsibility edge sequence within the delivery confirmation position; Write the actual delivery feedback into the feedback feedback feed position of the responsible edge according to the sequence order of the responsible edge corresponding to the feedback binding record; When the feedback feedback position receives the actual delivery feedback written to the target conversion node, it marks the responsible edge as a positive feedback edge; When the feedback feed receives real delivery feedback that has not been written to the target conversion node, the responsible edge is marked as a feed-back edge to be reviewed, and a review mapping is established between the feed-back edge to be reviewed and the corresponding feed-back edge in the feed-back evidence cache area. Write the positive backfill edges, backfill edges to be reviewed, and review mappings back into the advertising knowledge graph.

10. The intelligent digital advertising delivery method based on knowledge graphs according to claim 9, characterized in that, The subsequent process of writing the positive backfeed edges, backfeed edges to be reviewed, and review mappings back to the advertising placement knowledge graph specifically includes: Read the positive backflow edges, backflow edges to be reviewed, and review mappings from the advertising knowledge graph; Write the positive confirmation bit of the corresponding responsible edge into the positive backfeed edge, and keep the release confirmation bit of the candidate exposure action slot where the corresponding responsible edge of the positive backfeed edge is located. Read the borrowed edge corresponding to the recharge edge to be reviewed along the review mapping, and write the recharge edge to be reviewed into the review lock position of the candidate exposure action slot where the corresponding responsible edge is located; When the lockout position is checked and a write state is formed, the candidate exposure action slot where the corresponding responsible edge is located is written into the no-entry slot; When generating candidate exposure action slots in the next round of deployment tasks, read the no-entry slots, write the connection status between the candidate exposure action slots corresponding to the no-entry slots and the next round of deployment tasks as a no-entry disconnection state, and block the candidate exposure action slots corresponding to the no-entry slots from entering the local deployment cutout map.

Images