Propagation evaluation of recommended object and propagation method of recommended object

Abstract

The embodiment of the invention provides a recommendation object propagation evaluation method and a recommendation object propagation method, and the recommendation object propagation evaluation method comprises the steps: obtaining the propagation behavior data of a sample recommendation object, the propagation behavior data comprises the release behavior data, the aftereffect behavior data and the aftereffect behavior value, and the release behavior data comprises the release behavior data, the aftereffect behavior data and the aftereffect behavior value; the release behavior data represents that the content of a sample recommendation object is released through a distribution behavior under the condition that the initiating user receives the sample recommendation object, the aftereffect behavior data represents that the content of the sample recommendation object affects the initiating user to initiate an aftereffect behavior, and the aftereffect behavior value represents a value corresponding to the aftereffect behavior; training a propagation value evaluation model based on the propagation behavior data; and through the propagation value evaluation model, evaluating to obtain the distribution sorting priority of each target recommendation object. The dynamic influence of the whole chain-type propagation behavior mode on the propagation value is learned, and the black-box implicit modeling global propagation feature output distribution sorting priority is adopted to realize dynamic propagation value evaluation.

Description

Technical Field

[0001] The embodiments in this specification relate to the technical field of machine learning, and in particular to a method for evaluating the propagation of recommendation objects and a method for propagating recommendation objects. Background Technology

[0002] With the rapid development of internet technology, especially social media platforms, information dissemination chains have become a key carrier for information dissemination, product promotion, and public opinion management. These networks typically consist of multiple user nodes, each representing a user, interconnected through dissemination behaviors (such as reading, clicking, sharing, and forwarding), forming a dynamic dissemination chain. Assessing the value of dissemination chains is crucial for optimizing resource allocation, improving dissemination efficiency, and enhancing network resilience.

[0003] Currently, the value assessment of propagation links mainly adopts static or simplified models. For example, the initial network value is directly calculated based on the user's basic attributes (such as the number of followers and activity level), or a fixed propagation coefficient is applied to estimate the spread of influence.

[0004] However, because propagation behavior is inherently dynamic, the probability of propagation between different users is influenced by various factors (such as the strength of user relationships and content relevance), and the propagation process exhibits chain reaction characteristics. Ignoring this dynamic propagation chain can lead to distorted network value assessment, thereby limiting the optimization of propagation strategies. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a method for evaluating the propagation of recommendation objects. One or more embodiments of this specification also relate to a method for propagating recommendation objects, a recommendation object platform, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a method for evaluating the propagation of a recommended object is provided, comprising:

[0007] Acquire the propagation behavior data of the sample recommended object. The propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value. Publishing behavior data represents the content of the sample recommended object published by the initiating user through distribution behavior after receiving the sample recommended object. Follow-up behavior data represents the influence of the content of the sample recommended object on the reaching user to initiate follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior.

[0008] Train a communication value assessment model based on communication behavior data;

[0009] By using a propagation value assessment model, the distribution ranking priority of each target recommendation object is evaluated.

[0010] According to a second aspect of the embodiments of this specification, a method for propagating a recommendation object is provided, comprising:

[0011] Obtain the distribution ranking priority of each target recommendation object. The distribution ranking priority of each target recommendation object is determined by the priority ranking result based on the dissemination value evaluation model trained on the dissemination behavior data. The dissemination behavior data includes release behavior data, follow-up behavior data, and follow-up behavior value.

[0012] Based on the distribution priority of each candidate recommendation object, the content of each candidate recommendation object is disseminated to users in a specific order.

[0013] According to a third aspect of the embodiments of this specification, a recommendation object platform is provided, including an offline module and an online module;

[0014] Offline module, used for:

[0015] Acquire the propagation behavior data of the sample recommended object. The propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value. Publishing behavior data represents the content of the sample recommended object published by the initiating user through distribution behavior after receiving the sample recommended object. Follow-up behavior data represents the influence of the content of the sample recommended object on the reaching user to initiate follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior.

[0016] Train a communication value assessment model based on communication behavior data;

[0017] By using a propagation value assessment model, the distribution ranking priority of each target recommendation object is evaluated and obtained.

[0018] Online module, used for:

[0019] Based on the distribution ranking priority of each target recommendation object, the content of each target recommendation object is disseminated to users in a specific order.

[0020] According to a fourth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0021] Memory and processor;

[0022] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the above-mentioned propagation evaluation method for recommended objects or the propagation method for recommended objects.

[0023] According to a fifth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions, which, when executed by a processor, implement the steps of the above-described method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0024] According to a sixth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0025] In one embodiment of this specification, propagation behavior data of sample recommended objects is obtained. This propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value. Publishing behavior data represents the content of the sample recommended object published by the initiating user through distribution behavior after receiving it. Follow-up behavior data represents the impact of the sample recommended object's content on the reaching user initiating follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior. Based on the propagation behavior data, a propagation value evaluation model is trained. The distribution ranking priority of each target recommended object is evaluated using the propagation value evaluation model.

[0026] By acquiring propagation behavior data of sample recommended objects, including publishing behavior data, follow-up behavior data, and follow-up behavior value, a complete representation of the dynamic propagation chain is constructed: publishing behavior data reflects the initiating user's proactive distribution behavior, follow-up behavior data records the feedback behavior of the reached users, and follow-up behavior value quantifies the propagation effect. Through two-hop reach behavior, the chain reaction characteristics in the propagation chain are accurately defined. Based on this data, a propagation value assessment model is trained, learning the dynamic impact of the entire chain propagation behavior pattern on propagation value. Breaking through the limitations of static coefficients, a black box of end-to-end propagation effect is constructed. The black box implicitly models global propagation characteristics, and the dynamic propagation value assessment is achieved through the distribution ranking priority output by the model. This overcomes the assessment distortion problem caused by ignoring the dynamic nature of propagation and provides a precise basis for optimizing the propagation strategy of recommended objects. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the architecture of a recommendation platform;

[0028] Figure 2 This is a diagram illustrating a form of word-of-mouth marketing.

[0029] Figure 3 This is a diagram illustrating the propagation of recommendations.

[0030] Figure 4 This is a flowchart illustrating a method for evaluating the propagation of a recommended object, as provided in one embodiment of this specification.

[0031] Figure 5 This is a flowchart illustrating a method for evaluating the propagation of a recommended object, provided in one embodiment of this specification.

[0032] Figure 6 This is a schematic diagram of the structure of a communication value assessment model in a communication assessment method for a recommended object provided in one embodiment of this specification;

[0033] Figure 7 This is a schematic diagram of the core structure of a deep neural network in a propagation value assessment model of a method for assessing the propagation of a recommended object, provided in one embodiment of this specification.

[0034] Figure 8 This is a flowchart illustrating a method for propagating a recommended object according to one embodiment of this specification;

[0035] Figure 9 This is a flowchart illustrating a method for evaluating the propagation of recommended objects applied to advertising distribution on a content community platform, as provided in one embodiment of this specification.

[0036] Figure 10 This is a schematic diagram of the structure of a recommendation object platform provided in one embodiment of this specification;

[0037] Figure 11 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0038] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0039] The terminology used in one or more embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the invention are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of the invention refers to and includes any or all possible combinations of one or more associated listed items.

[0040] It should be understood that although various information may be described using terms such as first, second, etc., in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of the present invention, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0041] Furthermore, it should be noted that the data involved in one or more embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the statistics, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0042] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0043] Recommended Target: Refers to the specific entity in a transaction or analysis. In different scenarios, it can be represented as: Advertisement: Content carriers such as images / videos that serve as the target of advertising; SPU Collection: A standardized product cluster that serves as a product management unit (e.g., all color versions of the iPhone 15 are collectively referred to as one SPU); SKU: The smallest inventory management unit in retail, e-commerce, and warehousing management; Brand: A company's intellectual property assets that serve as the target of value assessment.

[0044] Standard Product Unit (SPU): A collection of marketable goods that share the same key attributes (such as name, function, model).

[0045] Stock Keeping Unit (SKU): The smallest unit of inventory management in retail, e-commerce and warehousing management. It is used to uniquely identify the specific specifications of a product and is usually composed of letters, numbers or a combination of codes.

[0046] Human-to-Human (H2H) marketing is a sharing marketing method based on community content platforms. It aims to enhance the influence and reach of recommended objects and the long-term effect of recommended content through users’ spontaneous and large-scale content creation and dissemination.

[0047] User-generated content (UGC) is community content created spontaneously by users. It is one of the core elements of word-of-mouth marketing and aims to inspire users to produce user-generated content materials through high-quality marketing content, thereby creating a wider range of influence within the community content platform than a single recommended object.

[0048] Black-box person-to-person marketing recommendation distribution: A recommendation distribution technology that utilizes person-to-person marketing algorithms. This method treats the user content creation and sharing process in a content community as a whole. Instead of explicitly analyzing the specific content and participating users during the dissemination process, it models the overall performance and uses this to influence the distribution of recommendation targets.

[0049] Expected Cost Per Mille (ECPM): This metric measures the efficiency of ad delivery revenue, representing the estimated average revenue per thousand ad impressions for the ad delivery advertiser. The formula is (Total Revenue / Total Impressions) × 1000, used to compare the ad delivery value across different channels.

[0050] Click-Through Rate (CTR): This measures the percentage of times a recommended product or content is clicked. It is calculated by dividing the number of clicks by the number of impressions and is one of the core metrics for evaluating the effectiveness of digital marketing. For example, if a recommended product is displayed 1000 times and generates 50 clicks, then the CTR is 5%.

[0051] Predicted Click-Through Rate (pCtr): This is the probability of content being clicked estimated by a machine learning model, used for optimizing recommendation ranking and delivery. This metric makes real-time predictions based on user characteristics, content characteristics, and contextual features. For example, the estimated click probability for a user on a particular recommended item might be pCtr = 3.2%.

[0052] Currently, the essence of recommendation platforms is to maximize immediate value, and word-of-mouth marketing is one important path: authentic content and sharing are the characteristics of community content platforms. Recommendation platforms can use word-of-mouth and viral spread among people to allow excellent content to be repeatedly disseminated within the community content platform's dissemination chain, reaching high-value users and optimizing the value of recommended objects. Figure 1 This diagram illustrates the architecture of a recommendation platform:

[0053] On traditional traffic-driven recommendation platforms, recommenders expose their products to users, thereby achieving conversions and generating direct / immediate value.

[0054] On a recommendation platform based on a word-of-mouth marketing model, the recommender exposes the recommended object to users. After receiving the recommended object, users can spread it through social networks, user-generated content, and direct communication, reaching multiple users such as user 1, user 2, user 3, and user n, generating dissemination value / viral value. Based on this, the value of the recommended object can be optimized by combining direct value and immediate value.

[0055] The value assessment of the propagation chain mainly adopts static or simplified models. For example, the initial network value is directly calculated based on the user's basic attributes (such as the number of followers and activity level), or the influence diffusion is estimated by applying a fixed propagation coefficient.

[0056] However, because propagation is inherently dynamic, the probability of propagation between different users is influenced by various factors (such as the strength of user relationships and content relevance), and the propagation process exhibits chain reaction characteristics. Ignoring this dynamic propagation chain can lead to distorted network value assessment, thereby limiting the optimization of propagation strategies.

[0057] This specification provides a method for evaluating the propagation of recommended content, such as a user-generated content (UGC) behavior marketing method. This method attributes and models user-generated content behavior and the resulting content interactions on a community content platform to the reading of recommended content. Then, using a trained propagation value evaluation model, it predicts the perturbations that can be applied during the delivery of recommended content, thereby introducing word-of-mouth marketing capabilities into the recommended content platform. This modeling approach treats the user-generated content and content sharing process on the community content platform as a holistic propagation chain. Instead of explicitly analyzing the content and participating users during the propagation process, it models the overall performance, learning the dynamic impact of the entire chain-like propagation behavior pattern on propagation value. This breaks through the limitations of static coefficients and constructs a black box of end-to-end propagation effects.

[0058] Figure 2 This diagram illustrates a person-to-person marketing approach.

[0059] Using the aforementioned black-box strategy, a single, highly atomic person-to-person marketing communication can be defined as:

[0060] The recommended audience delivery system reaches user K through clicks or impressions by delivering recommended audiences;

[0061] User K interacted with the recommended brand within a certain period of time, including at least one of the following: commenting, sharing, saving, or user-generated content;

[0062] User U, after being interacted with by K within a certain period of time, generates various behaviors.

[0063] For example, K's interactive behavior can reach U in the following ways: K's act of saving a content reaches U in the collection page; K's sharing reaches U; K's writing of user-generated content of the same brand reaches U; K's commenting behavior in user-generated content generates interaction with U.

[0064] Figure 3 A schematic diagram of recommendation object propagation is shown:

[0065] Taking user-generated content (UGC) as an example, and the clicks, interactions, and other content exchange behaviors generated by UGC, the entire attribution process includes two steps (two-hop attribution):

[0066] User-generated content behavior that occurs after user K finishes viewing a recommended item is attributed to the most recent recommended item reading touchpoint using the last click attribution method, based on the dimension of user identifier × brand identifier.

[0067] Clicks made by user U1 after viewing content are attributed to the most recent user-generated content exposure touchpoint using the last click attribution method, based on the dimension of user identifier × content identifier.

[0068] The interactive behavior that occurs after user U2 finishes viewing the content is attributed to the most recent user-generated content exposure touchpoint using the last click attribution method, based on the dimension of user identifier × content identifier.

[0069] And so on.

[0070] This specification also relates to a method for disseminating recommendations, a platform for recommending recommendations, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0071] See Figure 4 , Figure 4 A flowchart illustrating a method for evaluating the propagation of a recommended object according to an embodiment of this specification is shown, including the following specific steps:

[0072] Step 402: Obtain the propagation behavior data of the sample recommended object. The propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value. Publishing behavior data represents the content of the sample recommended object published by the initiating user through distribution behavior after receiving the sample recommended object. Follow-up behavior data represents the influence of the content of the sample recommended object on the reaching user to initiate follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior.

[0073] The sample recommendation objects are quantifiable entities used to train the communication value assessment model and are the core analysis objects in the communication chain. Their value is increased through the dynamic communication behavior of user nodes. The sample recommendation objects are the triggering medium of the communication chain, and their value assessment depends on the chain-like communication effect among users. The sample recommendation objects include, but are not limited to, advertising content (image / text / video), standardized product units (SPU), minimum stock units (SKU), or brand intellectual property assets.

[0074] A propagation chain is a dynamic topology structure composed of user nodes and propagation behaviors between nodes, used to simulate the diffusion path of information or recommended objects. For example, an atomized propagation chain is: user K reaches user U1 through user-generated content, and user U1 further reaches user U2 through a click behavior.

[0075] The propagation behavior data of the sample recommended objects is a dynamic dataset that records the interaction behavior between user nodes in the propagation chain and its value impact, used to train the propagation value assessment model. This data is used to quantify the propagation effect to train the model. Propagation behavior data can come from user behavior logs (such as ad exposure, user-generated content creation, and interaction records), user network relationship graphs, or open-source sample datasets; no specific limitations are imposed here. The propagation behavior data of the sample recommended objects typically exhibits temporal sequence (behaviors are sorted by timestamps), attribution (behaviors can be traced back to the initial recommended object touchpoint), and dynamism (the propagation path changes with user interaction).

[0076] Publishing behavior data refers to records of initiating users publishing content related to a sample recommendation object after receiving it, through distribution behavior. Publishing behavior data characterizes the content published by an initiating user upon receiving a sample recommendation object, and includes, but is not limited to: user identifier, brand identifier, distribution behavior type (such as user-generated content creation or sharing), and timestamp. For example, user K created a piece of user-generated content from the same brand within 24 hours of viewing an advertisement.

[0077] Subsequent behavioral data records the interactions or dissemination behaviors of users after the content of the sample recommended object reaches them. Subsequent behavioral data characterizes the influence of the content of the sample recommended object on the subsequent behaviors initiated by the reaching users. Subsequent behavioral data includes, but is not limited to: reaching user identifier, content identifier, behavior type (such as click, comment), and timestamp. For example, user U1 likes user K's user-generated content.

[0078] Subsequent behavior data represents the quantified value of subsequent behaviors, used to measure the effectiveness of dissemination. The value of subsequent behavior can be calculated using metrics such as click-through rate, number of interactions, or conversion rate. For example, user U2's comment behavior is assigned a dissemination value of 0.5 points.

[0079] For example, on a recommendation platform of a certain community content platform, it is necessary to extract 10,000 pieces of dissemination behavior data for a certain brand advertisement from the user behavior logs of the community content platform:

[0080] One of the metrics is: User K's creation of user-generated content after viewing the "Brand A" advertisement, and the clicks and comments generated by users U1 and U2 through this note. Based on the user identifier × brand identifier dimension, the creation of user-generated content is attributed to the advertisement touchpoint using the most recent click attribution method, and subsequent actions are attributed to the exposure touchpoint of the user-generated content, forming a complete chain of two-hop attribution dissemination behavior data.

[0081] In step 402, by acquiring the propagation behavior data of the sample recommended objects, including publishing behavior data, follow-up behavior data and follow-up behavior value, a complete representation of the dynamic propagation link is constructed. Publishing behavior data reflects the initiating user's active distribution behavior, follow-up behavior data records the feedback behavior of the reached users, and follow-up behavior value quantifies the propagation effect. Through two-hop reach behavior, the chain reaction characteristics in the propagation link are accurately defined.

[0082] Step 404: Train a communication value assessment model based on communication behavior data.

[0083] A propagation value assessment model is a machine learning model used to predict and evaluate the potential value of a recommended object in a propagation network. The model's input includes the characteristics of the recommended object, user attributes, content attributes, and propagation behavior, and its output is a propagation probability value and expected value ranging from 0 to 1. For example, a multi-task neural network model can simultaneously predict the probability of user-generated content publication (PUGC), the probability of subsequent behavior (PCTR), and the expected propagation value (PUGV).

[0084] Based on dissemination behavior data, training a dissemination value assessment model can be approached in several ways. One option is self-supervised training using this data. For example, the model predicts the probability of content publication, the probability of subsequent behavior, and the expected value of that behavior based on the data characteristics of the dissemination behavior data. The self-supervised loss value is then determined based on the differences between these data and the dissemination behavior data. The model parameters are then adjusted based on this loss value. Another option is training the model using a multi-task learning framework. This involves predicting the probability of content publication, the probability of subsequent behavior, and the expected value of that behavior based on the data characteristics of the dissemination behavior data, determining the joint loss function, and adjusting the model parameters based on the joint loss value. A third option is training the model using reinforcement learning. This involves modeling the dissemination chain as a Markov decision process, using the dissemination behavior data as state transition samples, and optimizing the model using the policy gradient method. No further limitations are specified here.

[0085] For example, dissemination behavior data is extracted from user behavior logs, including user K's ad exposure records, user-generated content publishing behavior, and subsequent user U's click / interaction behavior, and attributed by user × brand dimension.

[0086] Inputs for model training: user characteristics (such as historical interaction frequency), advertising characteristics (such as brand identity), and communication behavior characteristics (such as exposure time).

[0087] The output of model training is: Posting Behavior Probability PUGC (the probability that user K will post user-generated content), Follow-up Behavior Probability PCTR (the probability that user U will click on user-generated content), and Follow-up Behavior Expected Value PUGV (the expected interaction value of user-generated content).

[0088] Loss Calculation: Cross-entropy loss is used to calculate the error between the probability of published behavior (PUGC) and the probability of subsequent behavior (PCTR), and mean squared error loss is used to calculate the prediction error of the expected value of subsequent behavior (PUGV). The total loss value is obtained by weighted summation.

[0089] Parameter update: Optimize the model parameters through backpropagation until convergence.

[0090] In step 404, a propagation value assessment model is trained based on the data. The dynamic impact of the entire chain propagation behavior pattern on the propagation value is learned, breaking through the static coefficient limitation, constructing a black box of end-to-end propagation effect, and using black box implicit modeling of global propagation characteristics.

[0091] Step 406: Evaluate the distribution ranking priority of each target recommendation object by using the propagation value assessment model.

[0092] The target recommendation object is a quantifiable entity that serves as the core analysis object in the propagation network, and its value is increased through the dynamic propagation behavior of user nodes. The target recommendation object is the triggering medium of the propagation chain, and its value assessment depends on the chain propagation effect among users. The target recommendation object includes, but is not limited to, advertising content (image / text / video), standardized product units (SPU), minimum stock units (SKU), or brand intellectual property assets.

[0093] The distribution ranking priority of each target recommendation object is based on a dynamic score output by the propagation value assessment model, which determines the display order of the target recommendation object in the distribution queue. The distribution ranking priority takes into account both immediate value (such as click-through rate) and long-term propagation value (such as the viral potential of user-generated content), and the distribution ranking priority can be obtained through weighted calculation.

[0094] The distribution ranking priority of each target recommendation object is evaluated and obtained through the propagation value assessment model. One possible approach is to evaluate and obtain the distribution ranking priority of each target recommendation object based on the propagation behavior data of each target recommendation object through the propagation value assessment model. The propagation behavior data includes, but is not limited to, the following:

[0095] For example, in an advertising distribution scenario, candidate ads are evaluated using a dissemination value assessment model:

[0096] The model inputs are: user K's features (historical interaction frequency), candidate ad j's features (brand identity, content type), and contextual features (exposure time, device type).

[0097] Model outputs: PUGC (probability of user K publishing user-generated content of the same brand) = 0.02; PCTR (probability of user U clicking on user-generated content) = 0.05; PUGV (expected interaction value of user-generated content) = 1.5.

[0098] Priority Calculation: The formula for calculating the priority of an ad is hc = X_ij × PCTR_ij × PUGC_i × PUGV_ij, where X_ij represents the baseline weight coefficient for ad j in the i-th exposure scenario, i represents the i-th exposure, and j represents the j-th ad. If ad j has PCTR = 0.05, PUGC = 0.02, and PUGV = 1.5, then its score is 0.041. All ads are sorted in descending order according to their priority scores. Ad j ranks 3rd in the current exposure position and receives the corresponding exposure resources.

[0099] In step 406, dynamic propagation value assessment is achieved through the distribution ranking priority output by the model, which overcomes the assessment distortion problem caused by ignoring the dynamic nature of propagation and provides an accurate basis for optimizing the propagation strategy of the recommended object.

[0100] Corresponding to steps 402 to 406 above, Figure 5 This specification illustrates a flowchart of a method for evaluating the propagation of recommendations, provided in one embodiment:

[0101] The existing logic for the propagation of recommended objects is as follows: the advertiser of recommended objects reaches users through exposure of recommended objects, and converts this exposure into direct / immediate value for users.

[0102] Analyze user posting behaviors, including user-generated content, sharing, and comments. Based on these, transform subsequent behaviors, such as clicks, interactions, and orders, into dissemination / viral value. Determine long-term value based on direct / immediate value and dissemination / viral value.

[0103] The dissemination value / viral value of the release behavior and the subsequent behavior are obtained through the dissemination value assessment model.

[0104] In one optional embodiment of this specification, step 404 includes the following specific steps:

[0105] By using a value assessment model, based on the characteristics of the recommended objects, the user characteristics of the initiating user, the content characteristics of the recommended objects, the behavioral characteristics of the publishing behavior, and the behavioral characteristics of the follow-up behavior, we can obtain the probability of the initiating user publishing content for the recommended objects, the probability of the reach user's follow-up behavior for the recommended objects, and the expected value of the follow-up behavior.

[0106] The loss value is determined based on the differences between the content release probability, the probability of subsequent behavior, the expected value of subsequent behavior, and the dissemination behavior data.

[0107] Adjust the model parameters of the communication value assessment model based on the loss value.

[0108] The recommendation object features of the sample recommendation objects are the quantifiable attribute features of the sample recommendation objects. These features include, but are not limited to: static attributes such as brand identity, content type (text / image / video), and SPU / SKU encoding, as well as historical dissemination effects (such as click-through rate and interaction rate). For example, the recommendation object features of the advertisement "Brand A" include Brand ID = 123 and content type as video.

[0109] The user characteristics of the initiating user are the quantifiable attributes of the user who triggered the dissemination behavior. These characteristics include, but are not limited to, user profiles (such as age and gender), behavioral history (such as historical interaction frequency and content preferences), and the strength of social relationships. For example, user K's characteristics might include high activity level, 5 user-generated content posts in the past 30 days, and 1000 followers.

[0110] The content features of the recommended objects in the sample are the quantifiable attribute features of the user-generated content related to the recommended objects. These content features include content format (text / image / video), sentiment polarity, keyword distribution, and visual features. For example, the content features of user-generated content posted by user K include the keyword "sunscreen" appearing 8 times and a positive sentiment score of 0.7.

[0111] The behavioral characteristics of publishing behavior are the contextual attributes of the user when initiating the dissemination behavior. These characteristics include behavior type (share / comment), timestamp, device type, and geolocation. For example, the publishing behavior characteristics of user-generated content by user K include behavior type = text and image creation, device type = a specific mobile device, and timestamp.

[0112] The behavioral characteristics of follow-up behaviors are the contextual attributes of the actions taken after reaching the user. These characteristics include behavior type (share / comment), timestamp, device type, and geographic location. For example, user U1's click behavior characteristics include source = discovery page and dwell time = 45 seconds.

[0113] The probability of an initiating user publishing content related to a sample recommended object is the probability value predicted by the model for that user generating content related to the recommended object. This probability can be calculated through the interaction of user features and recommended object features, ranging from 0 to 1. For example, user K's PUGC for the advertisement "Brand A" is 0.02, indicating a 2% probability of publishing relevant user-generated content.

[0114] The probability of a user's subsequent behavior after being reached by the recommended content is the model's predicted probability that the user will exhibit a subsequent behavior towards the recommended content. This probability can be calculated by combining content and user characteristics, reflecting the continued potential of the propagation chain. For example, a PCTR of 0.05 for user U1's view of user K's user-generated content indicates a 5% probability of clicking.

[0115] The expected value of subsequent behavior is the quantified value of a single subsequent behavior predicted by the model. The expected value of subsequent behavior can be derived from regression of historical data and can be normalized to a dimensionless score. For example, user U2's comment behavior has a PUGV of 1.5, meaning the expected value is 1.5 times the baseline value.

[0116] The loss value measures the difference between the model's predictions and the actual propagation behavior data. The loss value can be calculated jointly using cross-entropy (for classification tasks) and mean squared error (for regression tasks). For example, the cross-entropy loss for PUGC prediction error is 0.15, and the mean squared error for PUGV is 0.08.

[0117] The loss value is determined based on the differences between the content publication probability, the probability of subsequent behavior, the expected value of subsequent behavior, and the dissemination behavior data. One possible approach is to determine the loss value by weighting these differences. Another possible approach is to determine the loss value through a joint loss function based on these differences. Yet another possible approach is to determine the loss value by weighting these differences through a joint loss function. No further limitations are specified here.

[0118] For example, a data behavior record is: extracting the record of user K exposing the advertisement "Brand A" from the community platform logs, and the user-generated content (content ID = Note_001) published in the following 24 hours;

[0119] Extract click and comment behaviors (behavior type = click / like) generated by users U1 and U2 through Note_001. Attribute user-generated content publishing behavior to user K × brand A, and subsequent behavior to user U × Note_001. Input features include: user K's activity level = high, CTR of ad "brand A" = 3.5%, and keyword of Note_001 = "sunscreen". Model outputs: PUGC = 0.02 (true value = 1), PCTR = 0.05 (true value = 0.03), PUGV = 1.5 (true value = 1.8). Calculate cross-entropy loss (PUGC: 0.15, PCTR: 0.12) and mean squared error (PUGV: 0.09), with a total loss of 0.36; update model parameters using the optimizer with a learning rate of 0.001.

[0120] In the embodiments of this specification, based on the characteristics of the recommended object, the user characteristics of the initiating user, the content characteristics of the recommended object, the behavioral characteristics of the publishing behavior, and the behavioral characteristics of the subsequent behavior, the probability of the initiating user publishing content for the recommended object, the probability of the reaching user's subsequent behavior for the recommended object, and the expected value of the subsequent behavior are predicted from multiple dimensions. Based on the loss value, a propagation value assessment model is trained, and the dynamic impact of the entire chain propagation behavior pattern on the propagation value is learned. This breaks through the limitations of static coefficients, constructs a black box of end-to-end propagation effect, and uses black box implicit modeling of global propagation characteristics.

[0121] In one optional embodiment of this specification, the loss value is determined based on the difference between the content publication probability, the probability of subsequent behavior, the expected value of subsequent behavior, and the dissemination behavior data, including the following specific steps:

[0122] Calculate the first loss value of the posting tag in the content posting probability and posting behavior data;

[0123] Calculate the second loss value of the interaction label between the probability of subsequent behavior and the subsequent behavior data;

[0124] Calculate the third loss value between the expected value of the subsequent behavior and the value label of the subsequent behavior;

[0125] The loss value is determined based on the first loss value, the second loss value, and the third loss value.

[0126] The posting tags for posting behavior data are binary labels used to mark whether the initiating user actually performed the content posting action. These tags can be generated based on whether there are content records in the user's behavior log related to the sample recommended object. For example, if user K created user-generated content of the same brand within 24 hours of viewing an advertisement, the posting tag = 1; otherwise, it is 0.

[0127] The interaction tags for post-event behavioral data are binary tags used to mark whether the reached users have interacted with the disseminated content. These interaction tags are generated based on user behavior logs, including actions such as clicks, likes, and comments. For example, if user U1 likes user-generated content after reading it, the interaction tag is 1; if they only browse without interaction, the tag is 0.

[0128] The follow-up behavior value tag quantifies the actual value score of the follow-up behavior's contribution to the dissemination network. The follow-up behavior value tag maps different behavior types to numerical values ​​based on a pre-defined conversion rate weighting table. For example, a click = 0.5 points, a comment = 1.2 points, and a save = 1.5 points.

[0129] The first loss value is the cross-entropy error between the predicted content posting probability and the actual posting tag. The first loss value measures the accuracy of the model's prediction of user-generated content posting behavior. For example, when PUGC = 0.02 and actual tag = 1, the cross-entropy loss = 3.91.

[0130] The second loss value is the cross-entropy error between the predicted probability of subsequent behavior and the actual interaction label. The second loss value assesses the model's ability to judge the continuity of chain propagation. For example, when PCTR = 0.05 corresponds to an actual label of 0.03, the cross-entropy loss is 0.15.

[0131] The third loss value is the mean squared error between the predicted expected value of subsequent behavior and the true value label. The third loss value reflects the accuracy of the model in quantifying the value of dissemination. For example, when PUGV = 1.5 corresponds to a true value of 1.8, the mean squared error is 0.09.

[0132] The loss value is determined based on the first loss value, the second loss value, and the third loss value. One possible method is to determine the loss value by weighting the first loss value, the second loss value, and the third loss value. Another possible method is to determine the loss value by using a joint loss function based on the first loss value, the second loss value, and the third loss value. No further restrictions are imposed here.

[0133] Optionally, the loss value is determined based on the first loss value, the second loss value, and the third loss value, and the calculation formula is as follows:

[0134] hc = α × PCTR × PUGC × PUGV

[0135] Where α is a fixed parameter.

[0136] For example, a data behavior record is as follows: The record of user K's exposure to the advertisement "Sunscreen B" is extracted from the logs and associated with Note_002 posted by K within the following 48 hours. User U1 clicked Note_002 (staying for 120 seconds) and user U2 commented "Works very well". The following tags are generated: Posting tag = 1 (UGC exists), Interaction tag = 1 (clicked), Value tag = 1.8 (click 0.5 + comment 1.2 + duration coefficient 0.1). The model outputs PUGC = 0.15 / PCTR = 0.08 / PUGV = 1.2, and the calculated first loss = 2.12 (cross-entropy), second loss = 0.33, and third loss = 0.36. Using a weighting coefficient α = 0.6, the total loss is calculated to be 1.55. The model parameters are updated through the optimizer with a learning rate of 0.001.

[0137] In the embodiments described in this specification, the multi-task joint training framework unifies the modeling of the probability of user-generated content, the probability of triggering subsequent behaviors, and the expected propagation value in chain propagation, overcoming the simplistic assumptions of traditional static attribution methods regarding dynamic propagation paths. By learning the implicit correlation between publishing behavior and subsequent behaviors end-to-end, the model can automatically capture the nonlinear impact of complex factors such as user relationship strength and content relevance on propagation value, providing a dynamic evaluation basis for recommendation object distribution strategies that balances immediate benefits and long-term viral potential.

[0138] In one optional embodiment of this specification, the distribution ranking priority of recommended objects is evaluated through a propagation value assessment model, including the following specific steps:

[0139] Obtain the set of candidate recommendation objects and the set of users, where the set of users includes multiple users;

[0140] Select the target recommendation object from the candidate recommendation object set, and select the initiating user from the user set;

[0141] By using a value assessment model, based on the characteristics of the target recommendation object and the user characteristics of the initiating user, we can obtain the probability of the initiating user's posting behavior on the target recommendation object's content, the probability of the reaching user's subsequent behavior on the target recommendation object's content, and the expected value of the subsequent behavior.

[0142] Calculate the propagation value coefficient of the target recommendation object based on the probability of posting behavior, the probability of subsequent behavior, and the expected value of subsequent behavior.

[0143] Return to the step of selecting the initiating user from the user set until the user set is empty, and determine the target propagation value of the target recommendation object based on the propagation value coefficient corresponding to each initiating user;

[0144] Return to the step of selecting target recommendation objects from the candidate recommendation object set until the candidate recommendation object set is empty. Then, determine the distribution ranking priority of each target recommendation object based on the target propagation value of each target recommendation object.

[0145] The candidate recommendation object set is a pool of candidate recommendation objects to be distributed, containing multiple instances of deliverable recommendation objects. Optionally, it can be a dynamic set updated in real time by the advertiser's delivery system, with each recommendation object carrying complete metadata. For example, the candidate recommendation object set may contain 100 ads to be distributed, such as ad IDs A001 (sunscreen) and A002 (sports shoes).

[0146] The user set refers to the currently active and reachable user group. Optionally, it can be a cluster of potentially high-value users selected based on user profiles, grouped by interest tags. For example, the user set could include 5,000 target users such as user K (a beauty enthusiast) and user M (a fitness enthusiast).

[0147] The initiating user is the initial user who is likely to trigger the propagation chain. The initiating user is a seed user with a history of content creation or sharing. For example, user K has published 3 pieces of user-generated content in the past 30 days and is marked as a user with high propagation potential.

[0148] The recommendation features of the target recommendation object are the quantifiable attribute features of the target recommendation object. The recommendation features of the target recommendation object include, but are not limited to: static attributes such as brand identity, content type (text / image / video), SPU / SKU encoding, and historical dissemination effects (such as click-through rate, interaction rate), etc.

[0149] The user characteristics of the initiating user are the quantifiable attributes of the user who triggered the dissemination behavior. The user characteristics of the initiating user include, but are not limited to, user profiles (such as age and gender), behavioral history (such as historical interaction frequency and content preferences), and the strength of social relationships.

[0150] The probability of an initiating user publishing content related to the target recommendation object is determined by the quantifiable attribute features of the content generated by the initiating user and relevant to the recommendation object. The content features of the target recommendation object include content format (text / image / video), sentiment polarity, keyword distribution, and visual features.

[0151] The probability of a user's subsequent behavior after being reached by the content of the target recommendation object is the probability value predicted by the model that the user will interact with the disseminated content. This probability reflects the dynamic potential for the continuation of the dissemination chain, ranging from 0 to 1. For example, the click probability (PCTR) of user U1 on user K's sunscreen note is 0.08.

[0152] The expected value of subsequent behavior is a quantitative score that predicts the value brought by a single subsequent behavior. The expected value of subsequent behavior is a normalized indicator derived from regression of historical behavior data. For example, the expected value of user U2's comment behavior is 1.2 (1.2 times the baseline value of 1.0).

[0153] The propagation value coefficient of the target recommendation object is a dynamic score that quantifies the propagation potential of a single user-recommended object combination. Optionally, it is a quantitative indicator calculated by multiplying PUGC × PCTR × PUGV. For example, the propagation coefficient of user K for advertisement A001 is 0.15 × 0.08 × 1.2 = 0.0144.

[0154] The target dissemination value of a recommended object is an assessment of its overall dissemination potential among the target user group. Note: This is a weighted aggregation of the dissemination coefficients of all seed users. For example, the target dissemination value of advertisement A001 among 5000 users is 72.5.

[0155] Optionally, the target propagation value of the target recommended object is determined based on the propagation value coefficient corresponding to each initiating user. One possible approach is to maximize and sum the propagation value coefficients corresponding to each initiating user to obtain the target propagation value of the target recommended object. The calculation formula is as follows:

[0156]

[0157] For example, for advertisement A001, 5000 seed users are traversed: the propagation coefficient of user K (beauty influencer) = 0.0144; the propagation coefficient of user M (ordinary user) = 0.0021; ... (other user data).

[0158] The dissemination value coefficients corresponding to each initiating user are maximized and aggregated: the dissemination coefficient of a certain user is 0.0125, and the target dissemination value is 72.5 after conversion. Similarly, the target dissemination value of advertisement A002 is calculated to be 89.7.

[0159] Final distribution order: A002(89.7)>A001(72.5).

[0160] In the embodiments described in this specification, the chain propagation potential of user dimensions is dynamically aggregated to achieve precise quantification of the distribution priority of recommended objects, breaking through the limitations of traditional static evaluation. It can intelligently identify content user combinations with high fission potential, and significantly improve the long-term propagation depth while ensuring immediate conversion effect.

[0161] Taking user-generated content as an example, the black-box solution ignores what kind of user-generated content user K will produce, and which users the user-generated content produced by user K will affect, focusing only on two core questions:

[0162] 1. After the target recommendation is distributed to user K, will user K generate user-generated content of the same brand?

[0163] 2. After user K produces user-generated content, how much clicks / interactions will this user-generated content generate to reach users?

[0164] In one optional embodiment of this specification, the propagation value assessment model includes a shared underlying neural network, a first branch network, a second branch network, and a third branch network. Based on the characteristics of the target recommendation object and the user characteristics of the initiating user, the propagation value assessment model obtains the probability of the initiating user's posting behavior towards the target recommendation object's content, the probability of the reaching user's subsequent behavior towards the target recommendation object's content, and the expected value of the subsequent behavior. This includes the following specific steps:

[0165] The recommendation object information of the target recommendation object and the user information of the initiating user are input into the shared underlying neural network to obtain the recommendation object features and user features. The recommendation object features and user features are then fused to obtain the shared features.

[0166] The shared features are input into the first branch network to generate content publication probabilities;

[0167] The shared features are input into the second branch network to generate the probability of subsequent behaviors.

[0168] The shared features are input into the third branch network to generate the expected value of the subsequent behavior.

[0169] A shared underlying neural network serves as the foundational network structure for extracting common features between recommended objects and users. Optionally, the shared underlying network employs a multilayer perceptron to achieve feature cross-compression. For example, the shared underlying neural network may consist of two fully connected layers: one with 256 dimensions and the other with 128 dimensions.

[0170] The first branch network is a dedicated subnetwork for predicting the probability of content publication. Optionally, the first branch network consists of a deep neural network core. For example, after inputting shared features, the output PUGC = 0.15.

[0171] The second branch network is a dedicated subnetwork for predicting the probability of subsequent behaviors. Optionally, the first branch network consists of a deep neural network core. For example, it outputs PCTR = 0.08 after inputting the same shared features.

[0172] The third branch network is a dedicated subnetwork for predicting the expected value of subsequent behavior. Optionally, the first branch network consists of a deep neural network core. For example, after inputting shared features, the output PUGV = 1.2.

[0173] The recommendation information for the target recommendation object is the raw data representation of the recommendation object. This information includes, but is not limited to, structured fields such as brand ID and content type. For example, the raw data for advertisement A001 is {Brand ID: 123, Type: Video}.

[0174] The user information of the initiating user is a raw data representation of the user. This user information includes, but is not limited to, dimensions such as user profile and behavioral history. For example, user K's raw data might be: {Age: 25, Interest Tag: Beauty}.

[0175] The shared features are the coded features resulting from the fusion of the features of the recommendation objects and the user features. These shared features represent the implicit expression of user interaction information with the recommendation objects.

[0176] Figure 6 This specification illustrates a schematic diagram of the communication value assessment model in a method for evaluating the communication of recommended objects, provided in one embodiment:

[0177] The features are embedded and encoded to obtain the embedding codes of multiple target recommendation object groups and the user embedding code of the initiating user. The embedding codes of multiple target recommendation object groups include: the embedding code of target recommendation object group 1, the embedding code of target recommendation object group 2, ..., the embedding code of target recommendation object group n.

[0178] After mean pooling, the embedding codes of multiple target recommendation object groups are fused by Concat and input into the compression activation layer. The Concat fusion features of the embedding codes of multiple target recommendation object groups are further fused and input into feature network layer 1 (256), which is activated by the ReLU activation function.

[0179] The user embedding code of the initiating user is input into neural network layer 1 (256) and activated by the Sigmoid activation function.

[0180] The two are then fused and input into neural network layer 2 (128), and activated by the ReLU activation function.

[0181] The user embedding code of the initiating user is input into neural network layer 2 (128) and activated by the Sigmoid activation function.

[0182] The two are fused to obtain shared features, which are then input into the first branch network (deep neural network core), the second branch network (deep neural network core), and the third branch network (deep neural network core), respectively, and output the content publication probability, the probability of subsequent behavior, and the expected value of subsequent behavior.

[0183] Figure 7 This specification illustrates a schematic diagram of the deep neural network core in a propagation value assessment model provided in one embodiment of the network enhancement method:

[0184] The core structure of any deep neural network is as follows: Input is fed into neural network layer 1 (64) and activated by the ReLU activation function. User cross-embedding features and content cross-embedding features are fused and then fed into neural network layer 3 (1). The two are concatenated and output.

[0185] In the embodiments described in this specification, an architecture design that utilizes a shared underlying network and multi-branch collaboration achieves efficient reuse of user-recommendation object interaction features and task-specific optimization. This structure, while ensuring the independence of each prediction task, fully leverages the common representation of underlying features, significantly improving model training efficiency and prediction accuracy, and enhancing the accuracy of dynamic evaluation of the propagation path.

[0186] In one optional embodiment of this specification, the propagation value coefficient of the target recommendation object is calculated based on the probability of the publishing behavior, the probability of the subsequent behavior, and the expected value of the subsequent behavior, including the following specific steps:

[0187] Based on the probability of publishing behavior, the probability of subsequent behavior, and the expected value of subsequent behavior, a basic quantity of dissemination value is generated.

[0188] Based on the basic quantity of dissemination value and the preset weight coefficient, the dissemination value coefficient of the target recommendation object is calculated.

[0189] The basic value of propagation is the original calculation result reflecting the potential value of a single propagation link. For example, it is obtained by multiplying PUGC, PCTR, and PUGV. When PUGC = 0.15, PCTR = 0.08, and PUGV = 1.2, the basic value is 0.0144.

[0190] The preset weighting coefficient is an adjustable parameter that adjusts the importance of different propagation stages.

[0191] Optionally, based on the probability of the publishing behavior, the probability of the subsequent behavior, and the expected value of the subsequent behavior, a basic quantity of dissemination value is generated, calculated using the following formula:

[0192] PCTR ij ×PUGC ij ×PUGV ij

[0193] Optionally, based on the basic quantity of dissemination value and the preset weight coefficient, the dissemination value coefficient of the target recommendation object is calculated, and the calculation formula is as follows:

[0194] x ij ×PCTR ij ×PUGC ij ×PUGV ij

[0195] In the embodiments described in this specification, the dissemination value coefficient of the target recommended object is calculated based on the basic quantity of dissemination value and the preset weight coefficient, realizing the dynamic weighted evaluation of the value of the dissemination link. This effectively solves the problem that traditional evaluation methods are insufficient in distinguishing the value contribution of different dissemination links, and provides a more granular decision-making basis for precise targeting.

[0196] In one optional embodiment of this specification, the distribution ranking priority of each target recommendation object is determined based on its target propagation value, including the following specific steps:

[0197] Obtain the expected dissemination benefits for each target recommendation audience;

[0198] Based on the expected propagation benefits of each target recommendation object, calculate the total expected propagation benefits of each target recommendation object;

[0199] Calculate the total target communication value of each target recommendation object based on the target communication value of each target recommendation object;

[0200] The ranking weight of each target recommendation object is calculated with the constraint that the total expected dissemination benefit is not less than a preset threshold and the objective of maximizing the total target dissemination value of each target recommendation object.

[0201] The target recommendation objects are sorted in descending order according to their ranking weights to obtain the distribution ranking priority of each target recommendation object.

[0202] The expected revenue from the dissemination of a target audience refers to the direct economic benefit that the target audience can generate during the dissemination process. The expected revenue from the dissemination of a target audience can be calculated based on historical conversion rates and average order value. For example, the expected revenue per thousand impressions for sunscreen advertisement A001 is 500 yuan.

[0203] The total expected revenue from each target audience is the sum of the direct economic benefits that each target audience can generate during the dissemination process. This total expected revenue reflects the baseline for revenue protection of the overall campaign. For example, the total expected revenue per thousand impressions for 100 ads is 50,000 yuan.

[0204] The total target communication value for all recommended targets is an aggregated index of the communication value coefficients of each target recommended target. The total target communication value for all recommended targets comprehensively considers both long-term brand value and short-term conversion returns. For example, the total target communication value of 100 advertisements = 8500.

[0205] Optionally, based on the expected propagation benefits of each target recommendation object, the total expected propagation benefits of each target recommendation object are calculated using the following formula:

[0206]

[0207] Optionally, with the constraint that the total expected propagation revenue is not less than a preset threshold, and with the objective of maximizing the total target propagation value of each target recommendation object, the ranking weight of each target recommendation object is calculated to represent the optimal solution under the i-th propagation. The calculation formula is as follows:

[0208]

[0209] j * =argmax j (PCTR ij ×PUGC ij ×PUGV ij +λECPM ij )

[0210] In this embodiment, a dual-value evaluation system based on a dynamic propagation link is constructed to jointly model the expected propagation benefits and target propagation value of the recommended objects. An optimization ranking mechanism under preset threshold constraints is introduced, and an optimization algorithm with constraints is used to maximize the propagation network effect while ensuring basic benefits. This system can autonomously identify high-quality recommended objects with high conversion rates and high viral propagation synergy. By prioritizing the distribution of recommended objects that can stimulate user-generated content and multi-hop propagation, the distribution efficiency is significantly improved.

[0211] In one optional embodiment of this specification, the preset threshold is determined by calculating a linear function based on the total revenue of the recommendation object delivery system over a historical period.

[0212] A referral targeting system is a technology platform used to manage and execute the distribution of referrals. Optionally, the referral targeting system integrates referral management, user outreach, and performance tracking functions, and supports dynamic adjustments to distribution strategies. For example, an advertising system on a community content platform can manage advertising at the SPU / SKU level.

[0213] A linear function is an nth-order linear relationship modeled based on historical data. It fits the linear trend of the return time series using the least squares method and is used to predict the threshold baseline. For example, a function y = ax + b is fitted using the return data of the past 30 days, and 90% of the y value is taken as the preset threshold.

[0214] Total revenue for a historical period represents the cumulative value generated by the dissemination of recommended content within that historical timeframe. Total revenue for a historical period includes direct conversion revenue and indirect dissemination value, aggregated and calculated by time window. For example, advertiser A's total revenue last week through the system = 5000 yuan in click revenue + 3000 yuan in user-generated content-related orders.

[0215] For example, the total system revenue data for the past 7 days is extracted as [45000, 48000, 51000, 49000, 50000, 52000, 53000] yuan. A baseline is obtained through linear regression: y = 1200x + 44000. 85% of the predicted value of 52400 yuan when x = 7 is taken as the threshold of 44540 yuan. The total expected dissemination revenue of the current 100 advertisements = 50000 yuan > the threshold, triggering optimization to maximize dissemination value.

[0216] In the embodiments described in this specification, a balance between commercial goals and dissemination value is achieved through multi-objective optimization under dynamic revenue threshold constraints. This ensures both the basic revenue requirements of the advertiser and fully leverages the viral spread effect of high-potential recommenders. It can adaptively adjust the evaluation strategy based on market fluctuations, maintaining the feasibility of the distribution strategy in complex dissemination environments.

[0217] See Figure 8 , Figure 8 A flowchart illustrating a method for propagating a recommended object according to an embodiment of this specification is shown, including the following specific steps:

[0218] Step 802: Obtain the distribution ranking priority of each target recommendation object. The distribution ranking priority of each target recommendation object is the priority ranking result determined by the dissemination value evaluation model trained based on the dissemination behavior data. The dissemination behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value.

[0219] For example, in the advertising system of a community content platform, the system acquires 20 user-generated advertising pieces to be distributed in real time. Each advertisement is dynamically scored using a dissemination value assessment model.

[0220] Model input: Characteristics of the propagation chain for each advertisement (current forwarding depth, secondary propagation conversion rate, KOL reach ratio, etc.).

[0221] Dynamic evaluation: A beauty product ad was promoted to the top 3 because it was forwarded by 3 community influencers, forming a "propagation tree"; while the newly released digital ad had a high initial click-through rate, but was temporarily ranked 15th due to the lack of propagation chain data.

[0222] Output: Generate a final priority queue (e.g., [Beauty Ad A, Food Ad B, Travel Ad C...]), where the top 5 items all have at least two layers of propagation links.

[0223] Step 804: Based on the distribution sorting priority of each target recommendation object, perform propagation operations on the content of each target recommendation object to reach users in a specific order.

[0224] For example, the advertising system implements a tiered outreach strategy:

[0225] First wave: Priority 1-3 beauty ads will be launched through the "Influencer Acceleration Channel". At the same time, the ads will be pushed to 500 users with beauty tags and given enhanced exposure to 80 users who forwarded the ads in the original dissemination chain.

[0226] Second wave (30 minutes later): Ads with priority 4-10 will trigger the normal delivery rules.

[0227] Dynamic adjustment: When a digital advertisement is forwarded by a tech influencer, forming a new dissemination chain, its priority immediately jumps to 6th place and enters the next batch of reach queues.

[0228] In the embodiments of this specification, the distribution ranking priority of each target recommendation object is obtained, where the priority is automatically generated by the propagation value assessment model based on the dynamic propagation link characteristics; based on this priority, sequential reach operations are performed on the content of each target recommendation object, realizing the dynamic optimization allocation of propagation resources: the priority ranking result implicitly contains the optimized solution of the propagation chain path, and the end-to-end decision of the black box model directly guides the propagation operation, avoiding the complex process of explicitly modeling user and content characteristics, ensuring that high-value content reaches key propagation nodes first, forming a self-optimizing propagation link structure, which significantly reduces computational complexity while improving propagation efficiency, and achieves a balance between propagation effect and resource consumption.

[0229] The following is in conjunction with the appendix Figure 9 Taking the application of the recommended target communication evaluation method provided in this manual in customer service business as an example, this paper further explains the recommended target communication evaluation method. Figure 9 The flowchart illustrates a method for evaluating the propagation of recommended targets applied to advertising distribution on a content community platform, according to an embodiment of this specification, including the following specific steps:

[0230] Step 902: Obtain the dissemination behavior data of the sample advertisement, wherein the dissemination behavior data includes publishing behavior data, follow-up behavior data and follow-up behavior value. Publishing behavior data represents the note published by the initiating user through distribution behavior after receiving the sample advertisement. Follow-up behavior data represents the influence of the note of the sample advertisement on the reaching user to initiate follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior.

[0231] For example, data on the propagation chain of a sunscreen advertisement was extracted from the behavior logs of a community note-taking platform: User K viewed the advertisement at 10:00 on May 1, 2023, and then posted a text and image note containing the brand's keywords at 14:00 on the same day (posting behavior data); User U1 clicked on the note through the discovery page on May 2, 2023, and stayed for 120 seconds (click behavior in the follow-up behavior data); User U2 commented in the comment section with a positive review of "long-lasting sun protection effect" (interaction behavior in the follow-up behavior data). According to the platform's value conversion table, the click behavior was assigned a value of 0.5 points, and the high-quality comment was assigned a value of 1.2 points, forming a complete two-hop attribution data chain.

[0232] Step 904: Using the dissemination value assessment model, based on the advertising characteristics of the sample advertisement, the user characteristics of the initiating user, the note characteristics of the sample advertisement, the behavioral characteristics of the posting behavior, and the behavioral characteristics of the follow-up behavior, obtain the probability of the initiating user posting a note for the sample advertisement, the probability of the reaching user's follow-up behavior for the sample advertisement note, and the expected value of the follow-up behavior.

[0233] For example, the advertising features of the ad "Brand A Sunscreen" (Brand ID = 123, Note Type = Video), user K's features (historical UGC posting frequency = 5 articles / month, beauty interest tags), note features (keyword "sunscreen" appears 8 times, sentiment score 0.7), and behavioral context features (device type = mobile, exposure time = weekday) are input into the model. A shared underlying neural network extracts feature cross-representations through a 256-dimensional fully connected layer. The first branch outputs PUGC = 0.15, the second branch outputs PCTR = 0.08, and the third branch outputs PUGV = 1.2, thus fully predicting the dynamic value of this propagation link.

[0234] Step 906: Calculate the first loss value of the posting probability and posting behavior data posting tag, calculate the second loss value of the follow-up behavior probability and follow-up behavior data interaction tag, calculate the third loss value of the follow-up behavior expected value and follow-up behavior value tag, and determine the loss value based on the first loss value, second loss value and third loss value.

[0235] For example, the model predicts that user K's PUGC = 0.15 corresponds to the actual posting tag 1 (UGC exists), and the first loss value of 2.12 is calculated using cross-entropy; the predicted user U1's PCTR = 0.08 corresponds to the actual click probability of 0.05, and the second loss value is 0.33; the predicted user PUGV = 1.2 corresponds to the actual value tag of 1.5 (click 0.5 + comment 1.2 - decay coefficient 0.2), and the third loss value is 0.09. The joint loss value is calculated using weighted coefficients, and the model parameters are adjusted and updated.

[0236] Step 908: Adjust the model parameters of the propagation value assessment model based on the loss value.

[0237] For example, an optimizer was used to adjust the weight matrix W1 (256×128) and bias vector b1 of the shared underlying neural network, simultaneously optimizing the parameters of the last layer of the three branch networks. After 500 iterations, the cross-entropy loss predicted by PUGC decreased to 0.89, and the mean squared error of PUGV decreased to 0.05, indicating that the model has learned the dynamic patterns of the propagation path.

[0238] Step 910: Obtain the candidate ad set and the user set, wherein the user set includes multiple users.

[0239] For example, the candidate ad pool is obtained in real time from the ad delivery system, containing 100 SPU-level ads to be distributed (such as sunscreen B, sports shoes C, etc.). At the same time, based on user profiles, 5,000 potential high-value users are selected to form a user set, including 1,200 beauty users, 800 sports users, and the rest are users with general interests. All users meet the filtering condition of having been active ≥ 3 times in the past 7 days.

[0240] Step 912: Select the target ad from the candidate ad set and the initiating user from the user set.

[0241] For example, a polling strategy is used to select ad A001 (sunscreen SPU) from the candidate pool, and user K (user ID = 9527) is randomly selected from the beauty user group as the initiating user. User K's characteristics include: age 25, number of followers 1500, historical UGC interaction rate 12.3%, and ad A001's characteristics include: brand awareness index 85, historical average CTR 3.2%, and video completion rate 65%.

[0242] Step 914: Input the advertising information of the target advertisement and the user information of the initiating user into the shared underlying neural network to obtain the corresponding advertising features and user features. Perform feature fusion on the advertising features and user features to obtain shared features. Input the shared features into the first branch network to generate the note publishing probability. Input the shared features into the second branch network to generate the subsequent behavior probability. Input the shared features into the third branch network to generate the expected value of the subsequent behavior.

[0243] For example, the original data of advertisement A001 {SPU_ID:789, video duration:30s} and the original data of user K {user ID:9527, device model:iPhone13} are input into the embedding layer to generate a 256-dimensional feature vector. After weighting by the SENET module, the shared features fuse a 128-dimensional implicit representation of user preferences and advertisement attributes. The first branch network outputs PUGC=0.18 through a 64-dimensional fully connected layer, the second branch outputs PCTR=0.07, and the third branch regresses to obtain PUGV=1.35.

[0244] Step 916: Based on the probability of posting behavior, the probability of subsequent behavior, and the expected value of subsequent behavior, generate a basic quantity of communication value. Based on the basic quantity of communication value and the preset weight coefficient, calculate the communication value coefficient of the target advertisement.

[0245] For example, the base value of dissemination is calculated as 0.18 × 0.07 × 1.35 = 0.01701. Combining this with the current time period weighting coefficient λ = 0.85 (added for evening active periods), the final dissemination value coefficient is 0.01701 × 0.85 = 0.0145. This value represents the expected viral value of user K disseminating ad A001, a 45% increase compared to the baseline level (0.01).

[0246] Step 918: Return to the step of selecting the initiating user from the user set until the user set is empty. Based on the propagation value coefficient corresponding to each initiating user, determine the target propagation value of the target advertisement.

[0247] For example, by iterating through 5000 users, the communication value coefficient of advertisement A001 is calculated: User K (beauty influencer) = 0.0145, User M (ordinary user) = 0.0021, ..., User N (beauty KOC) = 0.0213. The aggregated target communication value is 72.5 points, reflecting the overall communication potential of the advertisement among the target group.

[0248] Step 920: Return to the step of selecting target ads from the candidate ad set until the candidate ad set is empty. Based on the expected dissemination revenue of each target ad, calculate the total expected dissemination revenue of each target ad. Based on the target dissemination value of each target ad, calculate the total target dissemination value of each target ad. With the constraint that the total expected dissemination revenue is not lower than a preset threshold, and with the goal of maximizing the total target dissemination value of each target ad, calculate the ranking weight of each target ad. Sort each target ad in descending order according to the ranking weight to obtain the distribution ranking priority of each target ad.

[0249] For example, the total expected revenue from 100 advertisements is calculated as Σ(ECPM_i × Exposure_i) = 50,000 yuan, and the total target communication value is 500 points. With a revenue threshold of 45,000 yuan as a constraint, a linear programming problem is solved to obtain the optimal ranking weights: sunscreen advertisement A001 weight = 0.85 > sneaker advertisement A002 weight = 0.79... Finally, a priority queue [A001, A003, A005, ...] is generated to ensure that the communication depth is maximized while meeting the revenue target.

[0250] In the embodiments described in this specification, a quantitative evaluation system for word-of-mouth marketing is constructed. A pioneering two-hop attribution scheme is adopted to transform the ambiguous word-of-mouth dissemination process into a clear data link of "advertising touchpoint → user creation → secondary dissemination," providing a reliable data foundation for algorithm applications. Furthermore, a black-box modeling method is proposed, using neural networks to automatically learn the patterns of user creation and dissemination. This achieves intelligent evaluation and dynamic optimization of dissemination value without altering the original advertising system architecture. This technical solution enables the advertising system to balance immediate conversion effects and long-term dissemination value, significantly improving the natural spread of high-quality content and user engagement.

[0251] Corresponding to the above method embodiments, this specification also provides recommended object platform embodiments. Figure 10 A schematic diagram of the structure of a recommended object platform provided in one embodiment of this specification is shown. For example... Figure 10 As shown, the recommendation platform 1000 includes an offline module 1010 and an online module 1020;

[0252] Offline module 1010, used for:

[0253] Acquire the propagation behavior data of the sample recommended object. The propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value. Publishing behavior data represents the content of the sample recommended object published by the initiating user through distribution behavior after receiving the sample recommended object. Follow-up behavior data represents the influence of the content of the sample recommended object on the reaching user to initiate follow-up behavior. Follow-up behavior value represents the value corresponding to the follow-up behavior.

[0254] Train a communication value assessment model based on communication behavior data;

[0255] By using a propagation value assessment model, the distribution ranking priority of each target recommendation object is evaluated and obtained.

[0256] Online module 1020 is used for:

[0257] Based on the distribution ranking priority of each target recommendation object, the content of each target recommendation object is disseminated to users in a specific order.

[0258] In the embodiments of this specification, the collaborative architecture design of offline and online modules achieves full-process optimization of dissemination value assessment and content distribution: the offline module collects complete dissemination behavior data, including publishing behavior, follow-up behavior and their value, to construct a training sample set covering the entire dissemination chain cycle, enabling the dissemination value assessment model to learn the deep features of the dynamic dissemination chain; the online module implements real-time content distribution based on the dynamic priority output by the model, forming a closed-loop system combining offline training and online prediction.

[0259] Batch training of the offline module ensures the model accurately captures the dynamic characteristics of the propagation chain, solving the computational pressure problem of real-time modeling; the lightweight deployment of the online module enables efficient distribution decision response, meeting the timeliness requirements of propagation; the asynchronous processing mechanism of the two modules significantly reduces resource consumption while ensuring propagation effect, realizing efficient value assessment and accurate content delivery for large-scale propagation chains.

[0260] The above is an illustrative scheme of a recommendation platform according to this embodiment. It should be noted that the technical solution of this recommendation platform belongs to the same concept as the aforementioned recommendation object propagation evaluation method and recommendation object propagation method. Details not described in detail in the technical solution of the recommendation platform can be found in the descriptions of the aforementioned recommendation object propagation evaluation method or recommendation object propagation method.

[0261] Figure 11 A structural block diagram of a computing device according to one embodiment of this specification is shown. The components of the computing device 1100 include, but are not limited to, a memory 1110 and a processor 1120. The processor 1120 is connected to the memory 1110 via a bus 1130, and a database 1150 is used to store data.

[0262] The computing device 1100 also includes an access device 1140, which enables the computing device 1100 to communicate via one or more networks 1160. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 1140 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0263] In one embodiment of this specification, the aforementioned components of the computing device 1100 and Figure 11 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 11 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0264] The computing device 1100 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 1100 can also be a mobile or stationary server.

[0265] The processor 1120 is configured to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-mentioned propagation evaluation method for the recommended object or the propagation method for the recommended object.

[0266] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the aforementioned method for evaluating the propagation of recommended objects and the method for propagating recommended objects. Details not described in detail in the technical solution of the computing device can be found in the description of the aforementioned method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0267] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0268] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the aforementioned method for evaluating the propagation of recommended objects and the method for propagating recommended objects. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the aforementioned method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0269] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method for evaluating the propagation of recommended objects or the method for propagating recommended objects.

[0270] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the aforementioned method for evaluating the propagation of recommended objects and the technical solution of the method for propagating recommended objects. Details not described in detail in the technical solution of the computer program product can be found in the description of the aforementioned method for evaluating the propagation of recommended objects or the technical solution of the method for propagating recommended objects.

[0271] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0272] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0273] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0274] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0275] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for evaluating the propagation of recommendations, characterized in that, include: Acquire the propagation behavior data of the sample recommendation object, wherein the propagation behavior data includes publishing behavior data, follow-up behavior data and follow-up behavior value. The publishing behavior data represents the content of the sample recommendation object published by the initiating user through distribution behavior after receiving the sample recommendation object. The follow-up behavior data represents the influence of the content of the sample recommendation object on the reaching user to initiate follow-up behavior. The follow-up behavior value represents the value corresponding to the follow-up behavior. Based on the aforementioned dissemination behavior data, a dissemination value assessment model is trained; The distribution ranking priority of each target recommendation object is evaluated using the aforementioned dissemination value assessment model.

2. The method according to claim 1, characterized in that, The step of training a communication value assessment model based on the communication behavior data includes: By using a propagation value assessment model, based on the characteristics of the recommended object of the sample recommended object, the user characteristics of the initiating user, the content characteristics of the content of the sample recommended object, the behavioral characteristics of the publishing behavior, and the behavioral characteristics of the subsequent behavior, the probability of the initiating user publishing content of the sample recommended object, the probability of the reaching user's subsequent behavior of the content of the sample recommended object, and the expected value of the subsequent behavior are obtained. The loss value is determined based on the difference between the content publication probability, the subsequent behavior probability, the expected value of the subsequent behavior, and the dissemination behavior data. Based on the loss value, adjust the model parameters of the propagation value assessment model.

3. The method according to claim 2, characterized in that, The determination of the loss value based on the difference between the content publication probability, the probability of subsequent behavior, the expected value of subsequent behavior, and the dissemination behavior data includes: Calculate the first loss value between the content publication probability and the publication tag of the publication behavior data; Calculate a second loss value between the probability of the subsequent behavior and the interaction label of the subsequent behavior data; Calculate the third loss value between the expected value of the subsequent behavior and the value label of the subsequent behavior; The loss value is determined based on the first loss value, the second loss value, and the third loss value.

4. The method according to any one of claims 1-3, characterized in that, The step of evaluating the distribution ranking priority of recommended objects through the propagation value assessment model includes: Obtain a set of candidate recommendation objects and a set of users, wherein the set of users includes multiple users; Select the target recommendation object from the set of candidate recommendation objects, and select the initiating user from the set of users; By using a propagation value assessment model, based on the characteristics of the target recommendation object and the user characteristics of the initiating user, the probability of the initiating user's posting behavior on the content of the target recommendation object, the probability of the reaching user's subsequent behavior on the content of the target recommendation object, and the expected value of the subsequent behavior are obtained. Based on the publication behavior probability, the subsequent behavior probability, and the expected value of the subsequent behavior, the propagation value coefficient of the target recommendation object is calculated; Return to the step of selecting the initiating user from the user set, until the user set is empty, and determine the target propagation value of the target recommendation object based on the propagation value coefficient corresponding to each initiating user; Return to the step of selecting a target recommendation object from the candidate recommendation object set until the candidate recommendation object set is empty, and determine the distribution ranking priority of each target recommendation object based on the target propagation value of each target recommendation object.

5. The method according to claim 4, characterized in that, The propagation value assessment model includes a shared underlying neural network, a first branch network, a second branch network, and a third branch network. The propagation value assessment model, based on the recommendation object characteristics of the target recommendation object and the user characteristics of the initiating user, obtains the probability of the initiating user's posting behavior towards the content of the target recommendation object, the probability of the reaching user's subsequent behavior towards the content of the target recommendation object, and the expected value of the subsequent behavior, including: The recommendation object information of the target recommendation object and the user information of the initiating user are input into a shared underlying neural network to obtain the recommendation object features and user features. The recommendation object features and user features are then fused to obtain shared features. The shared features are input into the first branch network to generate content publishing probabilities; The shared features are input into the second branch network to generate the probability of subsequent behaviors. The shared features are input into the third branch network to generate the expected value of the subsequent behavior.

6. The method according to claim 4, characterized in that, The calculation of the propagation value coefficient of the target recommendation object based on the publication behavior probability, the subsequent behavior probability, and the expected value of the subsequent behavior includes: Based on the probability of the publishing behavior, the probability of the subsequent behavior, and the expected value of the subsequent behavior, a basic quantity of dissemination value is generated. Based on the aforementioned basic value of dissemination and the preset weight coefficient, the dissemination value coefficient of the target recommendation object is calculated.

7. The method according to claim 4, characterized in that, The step of determining the distribution ranking priority of each target recommendation object based on its target propagation value includes: Obtain the expected dissemination benefits for each target recommendation audience; Based on the expected propagation benefits of each target recommendation object, calculate the total expected propagation benefits of each target recommendation object; Based on the target propagation value of each target recommendation object, calculate the total target propagation value of each target recommendation object; The ranking weight of each target recommendation object is calculated with the constraint that the total expected dissemination revenue is not less than a preset threshold, and with the objective of maximizing the total target dissemination value of each target recommendation object. The target recommendation objects are arranged in descending order according to the ranking weight to obtain the distribution ranking priority of each target recommendation object.

8. The method according to claim 7, characterized in that, The preset threshold is the total revenue of the recommendation object delivery system over a historical period, and is determined by calculating a linear function.

9. A method for propagating recommendation objects, characterized in that, include: Obtain the distribution ranking priority of each target recommendation object, wherein the distribution ranking priority of each target recommendation object is determined by the priority ranking result based on the propagation value evaluation model trained based on propagation behavior data, and the propagation behavior data includes publishing behavior data, follow-up behavior data, and follow-up behavior value; Based on the distribution ranking priority of each target recommendation object, the content of each target recommendation object is disseminated to users in a specific order.

10. A recommendation platform, characterized in that, Includes offline modules and online modules; The offline module is used for: Acquire the propagation behavior data of the sample recommendation object, wherein the propagation behavior data includes publishing behavior data, follow-up behavior data and follow-up behavior value. The publishing behavior data represents the content of the sample recommendation object published by the initiating user through distribution behavior after receiving the sample recommendation object. The follow-up behavior data represents the influence of the content of the sample recommendation object on the reaching user to initiate follow-up behavior. The follow-up behavior value represents the value corresponding to the follow-up behavior. Based on the aforementioned dissemination behavior data, a dissemination value assessment model is trained; The distribution ranking priority of each target recommendation object is evaluated and obtained through the aforementioned dissemination value assessment model. The online module is used for: Based on the distribution ranking priority of each target recommendation object, the content of each target recommendation object is disseminated to users in a specific order.

11. A computing device, characterized in that, include: memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, It stores a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, Includes a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 9.

Images