An e-commerce short video advertisement pushing system based on a dynamic graph convolution network

Abstract

The application relates to the technical field of advertisement service, and discloses an e-commerce short video advertisement pushing system based on a dynamic graph convolution network, which comprises a write request data packet, a double-phase heterogeneous graph construction, a sorting quantity calculation, a candidate set generation, an ordered set generation, a pushing load generation and an e-commerce short video advertisement pushing sequence generation. The double-phase heterogeneous graph containing edge nodes and resident edges is constructed, the underlying cache information can be converted into a calculable graph structure feature, the neighborhood bias caused by the cache advantage can be inhibited through phase-weighted dynamic graph convolution calculation, the matching degree of the candidate advertisement and the user intention can be improved through candidate set screening and correlation filtering, and the diversity of the pushing sequence can be improved by reducing repeated exposure through fidelity sorting processing. The application is suitable for the pushing scene of the e-commerce short video advertisement, and can improve the conversion effect of advertisement putting and the user experience.

Description

Technical Field

[0001] This invention relates to the field of advertising service technology, and more specifically, to an e-commerce short video advertising push system based on dynamic graph convolutional networks. Background Technology

[0002] With the development of the e-commerce industry, short video advertising has gradually become mainstream. This type of content can intuitively showcase product features and attract user attention. To improve the relevance of ad delivery, the industry has gradually introduced dynamic graph convolutional network technology. This technology can model users, products, and ad content as a dynamic graph structure, capturing changes in user interests through graph convolutional propagation and providing a basis for ad ranking. Currently, this technology has been applied in multiple recommendation scenarios, effectively uncovering users' potential interests and improving ad matching efficiency.

[0003] In current ad delivery systems, dynamic graph convolution modeling primarily revolves around the correlation between user behavior and product semantics, using the user's historical interaction behavior as the edge weights of the graph to update node representations. Simultaneously, to improve ad loading speed, the industry caches popular ad content on edge computing nodes, and terminals preload candidate content to reduce user waiting time. These two technologies have become standard practice in the industry and are adopted by most platforms.

[0004] These two technical components typically operate independently, making it difficult for the recommendation system to perceive the underlying caching and preloading status. This results in ads with better cache status gaining more exposure opportunities. There is a temporal misalignment between the calculation of product intent and the accessibility of ad content. Some ads that match user intent may be ignored due to poor caching, while some irrelevant content with better cache status may occupy push positions, affecting the diversity of push notifications and conversion rates. Summary of the Invention

[0005] This invention provides an e-commerce short video advertising push system based on dynamic graph convolutional networks, which solves the technical problems mentioned in the background.

[0006] This invention provides an e-commerce short video advertising push system based on dynamic graph convolutional networks, comprising: The request data packet writing module is used to collect interaction events and behavior times, extract advertising short video identifiers and product identifiers, read the first segment dwell value, preloading progress and first frame arrival delay, and combine them to write the request data packet. The two-phase heterogeneous graph construction module is used to construct a two-phase heterogeneous graph using request data packets, construct interactive edges using interactive events and behavior moments, construct binding edges using product identifiers, and construct dwell edges using the first dwell value. The sorting quantity calculation module is used to calculate the phase coefficient based on the first segment dwell value, preloading progress and first frame arrival delay. It combines the phase coefficient to calculate the message strength of the edges in the two-phase heterogeneous graph, normalizes the message strength to obtain the connection weight, performs activation operation to obtain the node matrix, and concatenates the node matrix with the phase coefficient to calculate the sorting quantity. The candidate set generation module is used to eliminate invalid candidate short video advertisements based on the binding edges of the two-phase heterogeneous graph and the integrity of the first segment reachable data, and to generate a candidate set by using the sorting quantity to sort the candidate short video advertisements with the same product identifier in descending order and retaining the first one. The ordered set generation module is used to perform descending reordering of the candidate set using the sorting quantity, and remove duplicate candidate short video advertisements with the same product identifier to generate an ordered set; The push payload generation module is used to allocate broadcast time slots to ordered sets and calculate the anchoring time of product cards, and generate push payloads by combining advertising short video identifiers and product identifiers; The ad push sequence output module is used to extract the ad short video identifier, product identifier and broadcast time slot from the push payload, and use the broadcast time slot to perform ascending sorting to generate an e-commerce short video ad push sequence.

[0007] The beneficial effects of this invention are as follows: By uniformly collecting user behavior product information and physical reachability data at the same request time, this invention can ensure the temporal consistency of subsequent modeling; by constructing a biphase heterogeneous graph containing edge nodes and dwell edges, it can transform the underlying cached information into computable graph structure features; through phase-weighted dynamic graph convolution calculation, it can suppress the neighborhood bias caused by caching advantages; through candidate set screening and relevance filtering, it can improve the matching degree between candidate advertisements and user intent; through fidelity ranking processing, it can reduce duplicate exposure and improve the diversity of push sequences; through standardized payload encapsulation and sequence output, it can simplify the terminal processing flow. This invention is suitable for the push scenario of e-commerce short video advertisements, which helps to improve the conversion effect of advertising and the user experience. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of a module of an e-commerce short video advertising push system based on a dynamic graph convolutional network according to the present invention.

[0009] In the diagram: Request data packet writing module 101, two-phase heterogeneous graph construction module 102, sorting quantity calculation module 103, candidate set generation module 104, ordered set generation module 105, push payload generation module 106, and advertisement push sequence output module 107. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0012] like Figure 1 As shown, an e-commerce short video advertising push system based on dynamic graph convolutional networks includes: The request data packet writing module 101 is used to collect interactive events and behavior times, extract advertising short video identifiers and product identifiers, read the first segment dwell value, preloading progress and first frame arrival delay, and combine them to write the request data packet. The two-phase heterogeneous graph construction module 102 is used to construct a two-phase heterogeneous graph using request data packets, construct interactive edges using interactive events and behavior moments, construct binding edges using product identifiers, and construct dwell edges using the first dwell value. The sorting quantity calculation module 103 is used to calculate the phase coefficient for the first segment dwell value, preloading progress and first frame arrival delay, combine the phase coefficient to calculate the message strength of the edge in the dual-phase heterogeneous graph, normalize the message strength to obtain the connection weight, perform activation operation to obtain the node matrix, and concatenate the node matrix with the phase coefficient to calculate the sorting quantity. The candidate set generation module 104 is used to eliminate invalid candidate short video advertisements based on the binding edges of the two-phase heterogeneous graph and the integrity of the first segment reachable data, and to generate a candidate set by using the sorting quantity to sort the candidate short video advertisements with the same product identifier in descending order and retaining the first one. The ordered set generation module 105 is used to perform descending reordering of the candidate set using the sorting quantity, and to remove duplicate candidate short video advertisements with the same product identifier to generate an ordered set. The push payload generation module 106 is used to allocate broadcast time slots to the ordered set and calculate the anchoring time of the product card, and generate push payloads by combining the advertising short video identifier and the product identifier. The advertising push sequence output module 107 is used to extract the advertising short video identifier, product identifier and broadcast time slot from the push payload, and use the broadcast time slot to perform ascending sorting to generate an e-commerce short video advertising push sequence.

[0013] In one embodiment of the present invention, the request packet writing module 101 is configured to perform: At the moment when a user initiates a request for an e-commerce short video ad, a request data packet is generated, and the request time, user identifier, candidate ad short video, product identifier, service edge node, first segment dwell value, preloading progress and first frame arrival delay are written into the request data packet. Among them, all candidate short video ads are bound to a unique product identifier, and all candidate short video ad materials have a locatable first media file; Twenty short video interaction events are selected in reverse chronological order within the current session. If the number of short video interaction events is less than twenty, all short video interaction events are selected. The advertising short video identifier, product identifier, exposure time, viewing time percentage, swipe time, product card clicks, add-to-cart count, and transaction volume of each short video interaction event are written into the request data packet. Then read the category, price range, preferential benefits and merchant identifier of the product corresponding to each candidate short video advertisement, and read the first segment dwell value, first frame arrival latency and terminal preloading progress of the first media file at the service edge node of the candidate short video advertisement. The request data packet only reads user behavior events and existing cached data generated within the same request cycle, and does not generate cache replacement instructions.

[0014] It should be noted that the request-time data packet is generated at the same moment the user initiates the e-commerce short video ad request, and contains all the basic data required for this push. The service edge node refers to the computing node located at the network edge that currently carries the user's request. The first segment dwell value refers to the storage status identifier of the first segment of the ad short video media file in the service edge node. The preloading progress refers to the percentage of the terminal's download completion of the first segment of the ad short video media file. The first frame arrival latency refers to the estimated time for the first frame data to arrive at the terminal when transmitting the first segment of the ad short video media file from the service edge node to the terminal. Short video interaction events refer to all interactive behaviors that occur between the user and the ad short video, including exposure, viewing, swiping, product card clicks, adding to cart, and transactions.

[0015] Short video interaction events are collected from the application's user behavior logs, product information is read from the e-commerce platform's product database, the first segment dwell value and the first frame arrival latency are obtained from the edge node's cache management system, and the preloading progress is obtained from the terminal's application playback queue.

[0016] The maximum number of short video interaction events that can be selected in the current session is 20. If there are fewer than 20 events, all events will be selected, with priority given to the most recent event.

[0017] It should be noted that the current session refers to the continuous time period from when a user opens the e-commerce short video app to when they close it. For example, if a user opens the e-commerce short video app at 8:00 AM and closes it at 8:30 AM, this entire continuous time period constitutes a current session. All user actions within this session are considered actions within this session, and actions across sessions will not be included in the scope of this data collection.

[0018] The preferred length of the first media file is the first 3 seconds of the short video ad. For example, for a 15-second short video ad, the first media file is the audio and video content of the first 3 seconds of the ad, which is about 1.2MB in size. This length ensures that users can quickly see the core content of the ad while controlling caching and transmission costs, and does not consume too much storage and bandwidth resources.

[0019] The scope of existing cached data only includes the first segment of the short video ad already stored in the service edge node and the first segment of the short video ad already preloaded locally on the terminal, excluding the complete video ad data stored in the cloud. For example, if the service edge node stores the first 3 seconds of a short video ad, and the terminal also preloads the first 3 seconds of another short video ad, these both fall within the scope of existing cached data. However, the complete 15-second audio and video data of the ad stored in the cloud does not fall within the scope of existing cached data and will not be included in this physical reachability data collection.

[0020] In one embodiment of the present invention, the two-phase heterogeneous graph construction module 102 is configured to perform: A two-phase heterogeneous graph is constructed using request data packets. The nodes of the two-phase heterogeneous graph consist of user nodes, advertising short video nodes, product nodes, and edge nodes. The edges of the two-phase heterogeneous graph consist of user interaction with short video ads, user e-commerce behavior with products, short video ads binding with products, user access to edge nodes, and edge nodes staying in the first segment of short video ads. Record historical exposure, views, swipes, product card clicks, add-to-cart and transactions into the user's interaction with the short video ad or the user's e-commerce behavior with the product; Write each short video ad and its corresponding product into the short video ad and product binding edge; write the user and service edge node into the user and edge node access edge; Write the dwell relationship between the service edge node and the first media file of the short video ad into the dwell edge between the edge node and the first media file of the short video ad; Each edge carries the occurrence time or reading time, and the edges binding the short video advertisement and the product are fixed within the same request cycle.

[0021] It should be noted that the dual-phase heterogeneous graph is a dynamic graph structure that simultaneously contains product intent phase nodes and video reachable phase nodes, as well as edges corresponding to different phase relationships.

[0022] User nodes represent the logged-in user or anonymous device user who is currently initiating the advertising request. For example, when a user logs in using their mobile phone number, the user's unique identifier is used as the identifier of the user node. When the user is in a guest state who is not logged in, the unique identifier of the terminal device is used as the identifier of the user node.

[0023] The short video ad node represents the short video ad material that is being pushed to the candidate. Each short video ad node corresponds to a unique ad identifier.

[0024] Product nodes represent tradable goods linked to the short video advertisement, and each product node corresponds to a unique product identifier.

[0025] The user-ad interaction edge represents the behavioral relationship between the user and the ad short video, such as exposure, viewing, and swiping.

[0026] The "User and Product E-commerce Behavior" side represents the transaction-related behaviors between users and products, such as clicking on product cards, adding to cart, and completing transactions.

[0027] The "advertisement video and product binding" side represents the fixed association between the advertisement video and the product it promotes.

[0028] The user-edge node access edge represents the network connection relationship between the user and the current service edge node.

[0029] The edge node and the first segment of the short video ad reside on the caching storage relationship between the service edge node and the media file of the first segment of the short video ad.

[0030] The edge's time attribute refers to the specific time when the user's behavior occurs, and the read time refers to the specific time when the system reads physical data such as cached state.

[0031] It should be noted that the preferred storage method for two-phase heterogeneous graphs is adjacency list storage. In adjacency list storage, each node corresponds to an independent list containing information about all its neighboring nodes and their corresponding edges. For example, user node U1's neighbor list contains advertising node A12345, whose corresponding edge represents the user's interaction with the short video advertisement, occurring at 8:10 AM; product node P67890, whose corresponding edge represents the user's e-commerce interaction with the product, occurring at 8:15 AM; and edge node E1, whose corresponding edge represents the user's connection to the edge node, accessed at 8:20 AM. This storage method adapts to the sparsity characteristics of dynamic graphs, improving the efficiency of subsequent graph convolution calculations.

[0032] The update rules for the binding edge between the short video ad and the product are as follows: this binding relationship is fixed only within the same request cycle, and can be updated in different request cycles according to the product promotion strategy. For example, in the request cycle at 8:20, the product bound to ad A12345 is P67890. This binding relationship will not change during this request cycle. In the next request cycle at 9:00, the merchant adjusts the product promotion strategy and updates the product bound to ad A12345 to P67891. At this time, the binding edge between the short video ad and the product is updated to adapt to the new promotion requirements.

[0033] The weight representation of the edge between edge nodes and the first segment of the short video ad is directly based on the first segment's residency value. When the first segment residency value is 1, the weight of the corresponding edge is 1, indicating that the service edge node stores the first segment of the short video ad's media file. When the first segment residency value is 0, the weight of the corresponding edge is 0, indicating that the service edge node does not store the first segment of the short video ad's media file. For example, if the first segment of ad A12345 is already stored in edge node E1, then the weight of the edge between E1 and A12345 is 1. If the first segment of ad A67890 is not stored in edge node E1, then the weight of the edge between E1 and A67890 is 0.

[0034] It should be noted that the above embodiments transform the originally invisible cached structure at the system's bottom layer into an explicit neighborhood of the advertising push graph. This makes the propagation object of the dynamic graph convolution no longer limited to the semantic relationship between users, videos, and products, but also includes the relationship of whether or not the first segment of the short advertising video can be seen in time. This achieves the fusion modeling of semantic and physical information. The original cached state is now transformed into an explicit edge in the dynamic graph, enabling the recommendation system to directly use this physical state information when performing graph convolution calculations. This allows the system to consider both semantic matching and physical playback experience at the same time.

[0035] In one embodiment of the present invention, the sorting quantity calculation module 103 is configured to perform: First, add the product of the first dwell value and the first non-negative coefficient, and the product of the preloading progress and the second non-negative coefficient, and subtract the product of the number of natural pairs after the arrival delay of the first frame plus one and the third non-negative coefficient. Input the result into the Sigmoid function to obtain the phase coefficient. Specifically, the formula for calculating the phase coefficient is as follows: Parameter explanation: For advertising short videos At the request time The physically achievable phase coefficient, For the Sigmoid function, For the request time Service edge nodes Chinese advertising short videos The first segment's dwell value, 1 for dwell, 0 for no dwell. For the request time ,terminal For advertising short videos The preloading progress of the first segment has a value ranging from 0 to 1. For the request time Service edge nodes To the terminal Transmit advertising short videos The normalized value of the expected arrival delay of the first frame corresponding to the first media file. , , It is a non-negative coefficient and remains fixed during this push notification. Among them, the first segment dwell value is the dwell value of the first segment media file of the short video advertisement in the service edge node, the preloading progress is the preloading progress of the first segment media file of the short video advertisement on the terminal, and the first frame arrival delay is the normalized value of the expected first frame arrival delay from the service edge node to the terminal. Specifically, the formula for calculating the strength of an unnormalized message is as follows: Parameter explanation: For the request time Relationship types Next, the The node sends to the first The strength of unnormalized messages transmitted by each node. For relation type Next The node and the first The event intensity corresponding to the edges between nodes is determined by the percentage of viewing time, and the intensity of swiping is written with negative intensity. The intensity of product card clicks, adding to cart, and transactions is written according to the e-commerce behavior intensity. For the first The node and the first The formation or reading time of the edges between nodes For relation type Time decay coefficient, For the request time No. The reachable phase of each node is determined by the phase coefficient formula for advertising short video nodes, while non-advertising short video nodes are assigned a value of 1. For relation type Sensitivity coefficient to reachable phase; Multiply the event intensity, time decay, and phase coefficient exponentiation to obtain the unnormalized message intensity; then sum the unnormalized message intensities corresponding to each neighbor node in the same relation type, add a minimum positive number as the denominator, and use the unnormalized message intensity corresponding to a single neighbor node as the numerator. Calculate the ratio of the numerator to the denominator to obtain the connection weight. Specifically, the formula for calculating connection weights is as follows: Parameter explanation: For the request time Relationship types Next, the The node to the first The connection weight of each node For the request time Relationship types The Middle The set of neighbors of a node. For the request time Relationship types Next, the The node sends to the first The strength of unnormalized messages transmitted by each node. To avoid extremely small positive numbers with a denominator of 0; Next, multiply the node's own mapping matrix with the node matrix to obtain the size of the node's own matrix. Then, sum the products of the connection weights, relation mapping matrices, and neighbor node matrices under each relation type to obtain the size of the neighbor matrix. Finally, input the sum of the size of the node's own matrix and the size of the neighbor matrix into the activation function to obtain the updated node matrix. Specifically, the formula for calculating the node matrix is ​​as follows: Parameter explanation: For the request time The first node in the node matrix The node at the th Layer representation vector, For the request time The first node in the node matrix The node at the th Layer representation vector, For the first The layer's node self-mapping matrix, For the first Layer Relationship Types The message mapping matrix, It is a set of five types of relations. For the request time Relationship types The Middle The set of neighbors of a node. For the request time The first node in the node matrix The node at the th Layer representation vector, It is a non-linear activation function; Then concatenate the user node matrix, the short video ad node matrix, the product node matrix, and the phase coefficient, and calculate the product of the fixed sorting vector and the concatenation result to obtain the sorting quantity; Specifically, the formula for calculating the sorting quantity is as follows: Parameter explanation: For the request time ,user For advertising short videos The number of push notifications ranked. For a fixed sorting vector, This represents vector concatenation. For the request time Users in the node matrix In the The representation vector after convolution of a layer dynamic graph. For the request time Short video advertisements in the node matrix In the The representation vector after convolution of a layer dynamic graph. For the request time Short video advertisements in the node matrix Binding products In the The representation vector after convolution of a layer dynamic graph. For advertising short videos At the request time The physical reachable phase coefficient.

[0036] It should be noted that the phase coefficient is a parameter used to quantify the physical reachability of the short video ad at the time of request.

[0037] The first non-negative coefficient, the second non-negative coefficient, and the third non-negative coefficient are calibration coefficients preset by this invention, which are used to adjust the influence of the first segment dwell value, preloading progress, and first frame arrival delay on the phase coefficient, respectively, and preferably all of them are in the range of 0.5 to 2.0.

[0038] Event intensity is a parameter used to quantify the impact of different user behaviors on message propagation. It represents the contribution of different user behaviors to the propagation of node representations. Viewing behavior is valued according to the proportion of viewing time, swiping behavior is valued according to negative intensity, and product card click, add to cart and transaction behaviors are valued according to the increasing e-commerce behavior intensity.

[0039] The time decay is a parameter used to quantify the degree to which the influence of user behavior decays over time. It is calculated from the time decay coefficient corresponding to the relationship type and the time difference between the time when the behavior occurs and the time when the request is made.

[0040] The phase coefficient power is the power of the relationship type sensitivity coefficient of the phase coefficient, and is used to adjust the sensitivity of different relationship types to the physically reachable phase.

[0041] The connection weight is the normalized message propagation strength between nodes, representing the degree of influence of neighboring nodes on the current node.

[0042] The node's own mapping matrix is ​​a parameter matrix used to map the node's current layer representation vector to the next layer, and it is a learnable parameter.

[0043] The relation mapping matrix is ​​a parameter matrix used to map the representation vectors of neighbor nodes under different relation types to a unified space, and it is a learnable parameter.

[0044] The fixed sorting vector is a parameter vector obtained through offline training in this invention for calculating the final push sorting quantity, and it remains fixed during a single push process.

[0045] The smallest positive number is a preset constant used to avoid the denominator being 0 during the calculation process, and it is preferred to take the value of 10 to the power of negative 8.

[0046] The number of convolutional layers in the dynamic graph is the preset number of graph convolutional calculation layers in this invention, preferably 2 to 3 layers.

[0047] It should be noted that the specific rules for assigning event intensity include: when the viewing time percentage is 0 to 1, the event intensity corresponds to 0 to 1; the event intensity for swiping is fixed at -0.5; the event intensity for clicking a product card is 0.8; the event intensity for adding to cart is 1.2; and the event intensity for completing a transaction is 2.0. For example, if a user watches a short advertisement video in its entirety, with a viewing time percentage of 1.0, then the event intensity for this behavior is 1.0. If the user swipes past a short advertisement video without watching it, then the event intensity for this behavior is -0.5. If the user clicks on a product card, the event intensity for this behavior is 0.8. If the user adds the product to their cart, the event intensity is 1.2. If the user completes the purchase of the product, the event intensity is 2.0. Different behaviors correspond to different intensities, accurately reflecting the importance of the behavior.

[0048] The preferred values ​​for the time decay coefficient are: 0.01 per minute for user interaction with short video ads and 0.005 per minute for user interaction with e-commerce products. For example, if a user interacts with a short video ad 10 minutes ago, the time decay is 0.01 multiplied by 10, meaning the influence of that interaction has decreased by 10%. If a user interacts with an e-commerce product 20 minutes ago, the time decay is 0.005 multiplied by 20, again resulting in a 10% decrease in influence. Product-related interactions have a more lasting impact, hence the smaller decay coefficient is set for them.

[0049] The preferred values ​​for the sensitivity coefficient of the relation type to the reachable phase are 1.0 for the user-ad short video interaction edge and 0 for other relation types. In other words, only the interaction edge between the user and the ad short video is affected by the physical reachable phase coefficient. Other types of edges, such as the e-commerce behavior edge between the user and the product, or the binding edge between the ad short video and the product, are not affected by the phase coefficient because only the ad short video node has a physical reachable phase; the phase coefficient of other nodes is fixed at 1. Therefore, other relation types do not need to be sensitively adjusted for phase.

[0050] The dimension of the node representation vector is preferably 64 or 128. The vector of the user node records the user's interest and preference features, the vector of the product node records the product's attribute features, and the vector of the advertisement / short video node records the advertisement's content features. This dimension is sufficient to express the feature information of the node without bringing excessive computational complexity, ensuring the response speed of online push.

[0051] The fixed ranking vector is trained using offline supervised learning, with training labels consisting of users' actual clicks and conversion behaviors. During training, 10 million historical user push data points from the past three months are used, with users' actual clicks, add-to-cart actions, and transactions serving as training labels to train the ranking vector. After training, this ranking vector remains fixed during subsequent online push notifications, ensuring the stability of the push results.

[0052] It should be noted that the above embodiments modulate the message propagation intensity of dynamic graph convolution by introducing a physically reachable phase coefficient. This ensures that cache residency, terminal preloading, and first-frame latency only affect the message propagation intensity, rather than directly replacing the product intent score. This approach guarantees a good user experience while preventing the abnormal amplification of cached content. The message propagation intensity of short advertisements with good cache status will be reasonably adjusted, while the propagation intensity of short advertisements with poor cache status but high product intent matching will not be excessively suppressed. This ensures that short advertisements highly relevant to the user's product intent receive sufficient propagation weight while reasonably suppressing short advertisements that gain excessive exposure solely due to caching advantages.

[0053] In one embodiment of the present invention, the candidate set generation module 104 is configured to perform: Iterate through the candidate ad videos and read the product identifier, phase coefficient, and sorting index for each candidate ad video; When a candidate short video lacks an edge binding it to a product in a two-phase heterogeneous graph, the candidate short video is removed. If a candidate ad video lacks any of the following data: first segment dwell time, preloading progress, or first frame arrival delay, the candidate ad video will be removed. When the product node corresponding to the candidate short video ad does not enter the user node's neighborhood within three hops through the user's e-commerce behavior or the binding of the short video ad ad to the product, the candidate short video ad ad will be removed. When a consistent product identifier is bound to two or more candidate short video ads, the two or more candidate short video ads are sorted in descending order of ranking quantity, and the candidate short video ad with the highest ranking quantity is retained to generate a phase correction candidate set.

[0054] It should be noted that the phase correction candidate set is a set of candidate short video ads obtained after removing failed candidates and redundant ads for the same product.

[0055] The neighborhood within three hops is the set of product nodes that can be reached from the user node via no more than three edges.

[0056] It's important to clarify that the specific calculation rules for neighborhoods within three hops are as follows: only direct paths are counted, i.e., user to product (1 hop), user to ad / video and product (2 hop), and user to ad / video, product, ad / video and product (3 hop). Other indirect paths are not included. For example, if user U directly clicks on product P1, then P1 belongs to user's 1-hop neighborhood product. If user U has viewed ad A1, and ad A1 is linked to product P2, then P2 belongs to user's 2-hop neighborhood product. If user U has viewed ad A2, and ad A2 is linked to product P3, and user U has also viewed ad A3, and ad A3 is also linked to product P3, then P3 belongs to user's 3-hop neighborhood product. Other indirect paths, such as user to edge node to ad to product, are not included in the neighborhood calculation because these paths do not belong to the semantic path of product intent.

[0057] The complete criteria for determining invalid candidate short videos are as follows: in addition to lacking the binding edge between the short video and the product and the first segment of reachable data, it also includes cases where the product node is not within the user's three-hop neighborhood. For example, if a short video is not bound to a corresponding product, then the ad is a candidate for invalidation. If a short video lacks corresponding first segment reachable data, making it impossible to determine its physical reachability, it is also a candidate for invalidation. If the product bound to a short video is not within the user's three-hop neighborhood, it indicates that the product has extremely low relevance to the user's intent, and it is also a candidate for invalidation. These invalid candidates will be directly eliminated and will not enter the subsequent ranking stage.

[0058] When a product is associated with multiple short advertising videos, the processing rule is as follows: regardless of the phase coefficients of each short advertising video, only the one with the highest ranking is retained as a candidate representative for that product. For example, if product P1 is associated with three short advertising videos A1, A2, and A3, with rankings of 0.8, 0.7, and 0.6 respectively, then only A1 with the highest ranking will be retained as a candidate representative for that product. A2 and A3 will be removed, even if A2 has a high phase coefficient and a good cache state. This ensures that each product has at most one representative advertisement in the candidate set, preventing multiple advertisements for the same product from occupying positions in the candidate set.

[0059] It should be noted that the above embodiments ensure that each candidate ad video entering the ranking stage possesses both a valid product intent path and complete physical reachability information. This avoids the monopolization of the candidate set by cached trending content and eliminates ad videos with extremely low relevance to user intent, improving the efficiency and accuracy of subsequent ranking. While cached trending ads may have a good cache state, if they have low relevance to user intent, they will be directly eliminated and will not enter the candidate set. This ensures that all ads in the candidate set are relevant to user intent and possess complete physical reachability information. This guarantees both the relevance of candidate ad videos to user product intent and that each candidate contains complete physical reachability information, providing a high-quality candidate foundation for subsequent e-commerce intent-based ranking.

[0060] In one embodiment of the present invention, the ordered set generation module 105 is configured to perform: The phase correction candidate set is used as the sorting object, and the phase correction candidate set is reordered in descending order using the sorting quantity; Among them, the node matrix corresponding to the product node formed by the transaction volume, the number of items added to the cart and the number of product card clicks participates in the ranking calculation, while swipe events and short-view events are not used as separate criteria for elimination. When two or more candidate ad videos have the same product identifier, only the candidate ad video ranked first will be retained. When a candidate ad video corresponding to a consistent merchant identifier appears more than once consecutively in the descending order reordering result, the candidate ad video appearing for the second time and thereafter is removed during the sequence truncation stage, generating an ordered set without duplicate product identifiers.

[0061] It should be noted that the ordered set is a collection of short advertising videos that do not contain duplicate product identifiers, obtained after being sorted according to e-commerce intent.

[0062] It's important to note that the sequence truncation phase occurs after generating the ordered set but before encapsulating the push payload. The system first sorts the data to generate the ordered set, then determines the required number of push slots based on the terminal's display needs. For example, if a terminal needs to display 10 ads in a single feed, the first 10 ads are truncated from the ordered set. These truncated ads are then encapsulated into the push payload. This timing ensures that all ads encapsulated in the payload are the ones ultimately intended for push, preventing the removal of ads after encapsulation and avoiding resource waste.

[0063] The rule for handling consecutive appearances of ads from the same merchant is as follows: ads from the same merchant will only be removed if they appear consecutively more than once in the descending order. Ads from the same merchant that do not appear consecutively will be retained. For example, if the descending order of ads is A1, A2, A3, A4, where A1 and A2 are ads from merchant M, A3 is an ad from merchant N, and A4 is an ad from merchant M, then the consecutively appearing A2 will be removed, and A1, A3, and A4 will be retained. The non-consecutive A4 will be retained because users may be interested in different products from the same merchant, but simply do not want consecutive appearances of ads from the same merchant to negatively impact their experience.

[0064] A user's swiping or short-viewing behavior will reduce the ranking of the corresponding ad through negative event intensity, but will not directly remove the ad from the candidate set. This is because the user may not have been interested in the previous session but has developed new interest in the current session, or the product in the ad may indeed be something the user needs, but there was an accidental touch in the previous operation. This method can avoid filtering out potential high-conversion ads directly due to a single negative behavior.

[0065] It should be noted that the above embodiments take product conversion as the core objective of ad push. By eliminating ad short videos with duplicate product identifiers, the overexposure of the same product or merchant due to the caching advantage of video materials is avoided, ensuring the diversity of the push sequence and the efficiency of e-commerce conversion. Even if multiple ads for the same product are in good caching status and ranked highly, only one will be retained, avoiding users seeing a large number of duplicate product ads and improving the diversity of the push sequence. At the same time, it also improves the efficiency of product conversion. In this way, the push sequence is always centered on product conversion, reducing the duplicate exposure of the same product and the same merchant, improving user experience and the ROI of ad placement.

[0066] In one embodiment of the present invention, the push payload generation module 106 is configured to perform: Continuous broadcast slots are allocated according to the order of the candidate short video ads in an ordered centralized manner, and each candidate short video ad is written with an ad short video identifier, a product identifier, a broadcast slot, and a product card anchoring time. Among them, the product card anchoring time is located within the actual playback timeline of the candidate ad short video; the accessible location of the first media file of the ad short video is written into the playback field in the push payload, and the playback field is only used by the terminal to retrieve the material corresponding to the current playback time slot; Perform integrity checks on the push payload. The check items include the existence of the short video ad identifier, the existence of the product identifier, the existence of the broadcast time slot, and the product card anchoring time falling into the material timeline. After the integrity check passes, a push payload is generated. The push payload does not contain cache replacement instructions or training sample return commands.

[0067] It should be noted that the push payload is an advertising push data packet that the terminal can directly parse and execute.

[0068] The product card anchoring time is the specific point in time when the product purchase card pops up during the playback of the short advertising video.

[0069] The Playback field is a field in the push payload used to store the accessible location of the first media file of the short video ad.

[0070] It should be noted that the rule for determining the anchoring time of the product card is that it is preset by the merchant when uploading the short video advertisement, and it must fall within the actual playback timeline of the short video advertisement, preferably set between the 2nd and 5th second after the advertisement starts playing.

[0071] The specific method for verifying the integrity of the push payload is to check the existence and correct format of each field. The verification of the product card anchoring time is completed by comparing it with the duration information of the advertisement short video. The system checks whether the fields such as advertisement identifier, product identifier, broadcast time slot, and product card anchoring time in the push payload all exist and whether the format meets the requirements. Then, it checks whether the product card anchoring time is less than the duration of the advertisement short video. For example, if the advertisement duration is 15 seconds and the anchoring time is 3 seconds, it meets the requirements; if the anchoring time is 20 seconds, it does not meet the requirements, the verification fails, and the advertisement will not be included in the push payload, ensuring the integrity and validity of the payload.

[0072] The format for the accessible location of the first media file is a standard URL, pointing to the storage address of the first media file on the service edge node or the local terminal.

[0073] It should be noted that the above embodiments convert the calculation results of dynamic graph convolution into an advertising push payload that can be directly executed by the terminal. At the same time, the payload content is strictly limited and does not include any cache replacement or training sample return instructions, ensuring that the present invention always focuses on the core objective of e-commerce short video advertising push.

[0074] The push payload of this invention contains only information related to advertising push. All cache adjustments are handled independently by the original cache system. This invention only uses cached state information to optimize advertising push, which enables the terminal to accurately and timely display product purchase cards and improves the user's purchase guidance effect.

[0075] In one embodiment of the present invention, the advertisement push sequence output module 107 is configured to perform: Read the push payload and extract the ad short video identifier and product identifier of each candidate ad short video in the order of broadcast time slot; Generate a sequence of items consisting of an ad short video identifier, a product identifier, and a broadcast time slot, and arrange all sequence items in ascending order according to the broadcast time slot to form an e-commerce short video ad push sequence; Among them, each sequence item of the e-commerce short video ad push sequence only includes three dimensions: ad short video identifier, product identifier, and broadcast time slot; The e-commerce short video ad push sequence does not include model scores, rule sets, cache update instructions, or training parameters; after the e-commerce short video ad push sequence is output, the ad push process for the current request cycle ends.

[0076] It should be noted that the e-commerce short video ad push sequence is the final output of the system, which sorts the ad push results by broadcast time slot.

[0077] A sequence item is a single element in the e-commerce short video ad push sequence, containing only three dimensions: ad short video identifier, product identifier, and broadcast time slot.

[0078] The broadcast slot refers to the playback order of the short advertising video in the terminal's information stream.

[0079] It should be noted that the numbering rule for the broadcast slots is that they are numbered consecutively starting from 1, corresponding to the playback order of the advertisement in the information stream. The smaller the number, the earlier the advertisement is played.

[0080] The current request cycle ad push process has ended, meaning that the result of this push will not be used to update model parameters, adjust caching strategies, or perform any subsequent closed-loop control operations. After this push is completed, the push result and the user's subsequent behavior will not be used to update the model parameters used in this push, nor will they be used to adjust the caching strategy of edge nodes. This push is a single-path process; it ends after output is completed, without any subsequent closed-loop feedback operations. All model updates are completed offline.

[0081] The output format of the push sequence is a structured data format, such as JSON, which is easy for the terminal to parse and process. Each sequence item in the push sequence is stored in the form of JSON fields, including three fields: the advertisement short video identifier, the product identifier, and the playback time slot. After receiving the data, the terminal can directly parse these fields without performing any additional conversion processing, and quickly complete the preparation for advertisement playback.

[0082] It should be noted that this invention can be deployed in the online recommendation service cluster of an e-commerce platform, and interface with the platform's edge computing nodes and user terminal applications. Edge computing nodes periodically synchronize their cached state to the recommendation service, and user terminals, when initiating a request, report their local pre-loaded state and the user's current session behavior to the recommendation service. When a user opens an e-commerce short video application and scrolls down the feed to trigger an ad request, the recommendation service processes the request according to the procedure of this invention, generating a corresponding push sequence.

[0083] The final output is a standardized e-commerce short video ad push sequence. Each element in the sequence includes an ad short video identifier, a product identifier, and a playback slot. For example, when a user browses clothing content, the system-generated push sequence might have the following sequence items: the first item is ad identifier A12345, product identifier P67890, and playback slot 1; the second item is ad identifier A67890, product identifier P12345, and playback slot 2; and so on. After receiving the sequence, the terminal plays the ads sequentially according to their playback slots. When the first ad reaches its 3rd second, the terminal will display a corresponding product purchase card. The user can directly click the card to enter the product details page and complete the purchase. Throughout the process, the ads load smoothly without any lag. Furthermore, the pushed ads cover different products and merchants, avoiding duplicate content, allowing users to quickly find products of interest.

[0084] It should be further noted that this invention will display clear operation prompts to users before collecting user operation logs and related information. Data collection will only commence after the user actively consents to authorization. The information collected by this invention is solely for implementing the e-commerce short video advertising push function and does not collect users' sensitive personal information. Users can withdraw their authorization at any time, and the corresponding data collection and use will immediately cease upon withdrawal. All data storage and processing strictly comply with relevant laws and regulations and will not be used for other unrelated purposes or disclosed to any third party.

[0085] The content of this embodiment has been described above, but this embodiment is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this embodiment, all of which are within the protection scope of this embodiment.

Claims

1. An e-commerce short video advertising push system based on dynamic graph convolutional networks, characterized in that, include: The request data packet writing module collects interaction events and behavior times, extracts advertising short video identifiers and product identifiers, reads the first segment dwell value, preloading progress and first frame arrival delay, and combines them to write the request data packet. The two-phase heterogeneous graph construction module uses request data packets to construct a two-phase heterogeneous graph, uses interactive events and behavior moments to construct interactive edges, uses product identifiers to construct binding edges, and uses the first segment dwell value to construct dwell edges. The sorting quantity calculation module calculates the phase coefficient based on the first segment dwell value, preloading progress, and first frame arrival delay. It then calculates the message strength of the edges in the biphase heterogeneous graph using the phase coefficient, normalizes the message strength to obtain the connection weight, performs activation operations to obtain the node matrix, and concatenates the node matrix with the phase coefficient to calculate the sorting quantity. The candidate set generation module removes invalid candidate short video ads based on the binding edges of the two-phase heterogeneous graph and the integrity of the first segment of reachable data. It uses the sorting quantity to sort the candidate short video ads with the same product identifier in descending order and retains the first one to generate a candidate set. The ordered set generation module uses a sorting quantity to perform a descending reordering of the candidate set, and removes duplicate candidate short video advertisements with the same product identifier to generate an ordered set. The push payload generation module allocates broadcast time slots to ordered sets and calculates product card anchoring times, and generates push payloads by combining advertising short video identifiers and product identifiers. The ad push sequence output module extracts the ad short video identifier, product identifier, and broadcast time slot from the push payload, and uses the broadcast time slot to perform ascending sorting to generate an e-commerce short video ad push sequence.

2. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The request packet writing module is configured to execute: At the moment when a user initiates a request for an e-commerce short video ad, a request data packet is generated, and the request time, user identifier, candidate ad short video, product identifier, service edge node, first segment dwell value, preloading progress and first frame arrival delay are written into the request data packet. Among them, all candidate short video ads are bound to a unique product identifier, and all candidate short video ad materials have a locatable first media file; Twenty short video interaction events are selected in reverse chronological order within the current session. If the number of short video interaction events is less than twenty, all short video interaction events are selected. The advertising short video identifier, product identifier, exposure time, viewing time percentage, swipe time, product card clicks, add-to-cart count, and transaction volume of each short video interaction event are written into the request data packet. Then read the category, price range, preferential benefits and merchant identifier of the product corresponding to each candidate short video advertisement, and read the first segment dwell value, first frame arrival latency and terminal preloading progress of the first media file at the service edge node of the candidate short video advertisement. The request data packet only reads user behavior events and existing cached data generated within the same request cycle, and does not generate cache replacement instructions.

3. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The two-phase heterogeneous graph construction module is configured to execute: A two-phase heterogeneous graph is constructed using request data packets. The nodes of the two-phase heterogeneous graph consist of user nodes, advertising short video nodes, product nodes, and edge nodes. The edges of the two-phase heterogeneous graph consist of user interaction with short video ads, user e-commerce behavior with products, short video ads binding with products, user access to edge nodes, and edge nodes staying in the first segment of short video ads. Record historical exposure, views, swipes, product card clicks, add-to-cart and transactions into the user's interaction with the short video ad or the user's e-commerce behavior with the product; Write each short video ad and its corresponding product into the short video ad and product binding edge; write the user and service edge node into the user and edge node access edge; Write the dwell relationship between the service edge node and the first media file of the short video ad into the dwell edge between the edge node and the first media file of the short video ad; Each edge carries the occurrence time or reading time, and the edges binding the short video advertisement and the product are fixed within the same request cycle.

4. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The sorting quantity calculation module is configured to execute: First, add the product of the first dwell value and the first non-negative coefficient, and the product of the preloading progress and the second non-negative coefficient, and subtract the product of the number of natural pairs after the arrival delay of the first frame plus one and the third non-negative coefficient. Input the result into the Sigmoid function to obtain the phase coefficient. Among them, the first segment dwell value is the dwell value of the first segment media file of the short video advertisement in the service edge node, the preloading progress is the preloading progress of the first segment media file of the short video advertisement on the terminal, and the first frame arrival delay is the normalized value of the expected first frame arrival delay from the service edge node to the terminal. Multiply the event intensity, time decay, and phase coefficient exponentiation to obtain the unnormalized message intensity; then sum the unnormalized message intensities corresponding to each neighbor node in the same relation type, add a minimum positive number as the denominator, and use the unnormalized message intensity corresponding to a single neighbor node as the numerator. Calculate the ratio of the numerator to the denominator to obtain the connection weight. Next, multiply the node's own mapping matrix with the node matrix to obtain the size of the node's own matrix. Then, sum the products of the connection weights, relation mapping matrices, and neighbor node matrices under each relation type to obtain the size of the neighbor matrix. Finally, input the sum of the size of the node's own matrix and the size of the neighbor matrix into the activation function to obtain the updated node matrix. Then, the user node matrix, the short video ad node matrix, the product node matrix, and the phase coefficient are concatenated, and the product of the fixed sorting vector and the concatenation result is calculated to obtain the sorting quantity.

5. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The candidate set generation module is configured to execute: Iterate through the candidate ad videos and read the product identifier, phase coefficient, and sorting index for each candidate ad video; When a candidate short video lacks an edge binding it to a product in a two-phase heterogeneous graph, the candidate short video is removed. If a candidate ad video lacks any of the following data: first segment dwell time, preloading progress, or first frame arrival delay, the candidate ad video will be removed. When the product node corresponding to the candidate short video ad does not enter the user node's neighborhood within three hops through the user's e-commerce behavior or the binding of the short video ad ad to the product, the candidate short video ad ad will be removed. When a consistent product identifier is bound to two or more candidate short video ads, the two or more candidate short video ads are sorted in descending order of ranking quantity, and the candidate short video ad with the highest ranking quantity is retained to generate a phase correction candidate set.

6. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The ordered set generation module is configured to execute: The phase correction candidate set is used as the sorting object, and the phase correction candidate set is reordered in descending order using the sorting quantity; Among them, the node matrix corresponding to the product node formed by the transaction volume, the number of items added to the cart and the number of product card clicks participates in the ranking calculation, while swipe events and short-view events are not used as separate criteria for elimination. When two or more candidate ad videos have the same product identifier, only the candidate ad video ranked first will be retained. When a candidate ad video corresponding to a consistent merchant identifier appears more than once consecutively in the descending order reordering result, the candidate ad video appearing for the second time and thereafter is removed during the sequence truncation stage, generating an ordered set without duplicate product identifiers.

7. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The push payload generation module is configured to execute: Continuous broadcast slots are allocated according to the order of the candidate short video ads in an ordered centralized manner, and each candidate short video ad is written with an ad short video identifier, a product identifier, a broadcast slot, and a product card anchoring time. Among them, the product card anchoring time is located within the actual playback timeline of the candidate ad short video; the accessible location of the first media file of the ad short video is written into the playback field in the push payload, and the playback field is only used by the terminal to retrieve the material corresponding to the current playback time slot; Perform integrity checks on the push payload. The check items include the existence of the short video ad identifier, the existence of the product identifier, the existence of the broadcast time slot, and the product card anchoring time falling into the material timeline. After the integrity check passes, a push payload is generated. The push payload does not contain cache replacement instructions or training sample return commands.

8. The e-commerce short video advertising push system based on dynamic graph convolutional networks according to claim 1, characterized in that, The ad push sequence output module is configured to execute: Read the push payload and extract the ad short video identifier and product identifier of each candidate ad short video in the order of broadcast time slot; Generate a sequence of items consisting of an ad short video identifier, a product identifier, and a broadcast time slot, and arrange all sequence items in ascending order according to the broadcast time slot to form an e-commerce short video ad push sequence; Among them, each sequence item of the e-commerce short video ad push sequence only includes three dimensions: ad short video identifier, product identifier, and broadcast time slot; The e-commerce short video ad push sequence does not include model scores, rule sets, cache update instructions, or training parameters; after the e-commerce short video ad push sequence is output, the ad push process for the current request cycle ends.

Images