Recommended object optimization method and device, electronic equipment and program product

Abstract

The invention discloses a recommendation object optimization method and device, electronic equipment and a program product, and relates to the technical field of content recommendation, the method comprises the following steps: obtaining theme features of a recommendation set, the recommendation set being used for representing multi-modal recommendation information; obtaining a matching relationship between the theme feature and an interest portrait of the first recommendation object, wherein the interest portrait is determined based on first interaction data of the first recommendation object; screening the first recommendation object based on the matching relationship to obtain a second recommendation object; recommending the recommendation set to a second recommendation object, and obtaining second interaction data of the second recommendation object for the recommendation set; and optimizing the second recommendation object by using the second interaction data. By implementing the technical scheme of the application, the target users can be screened and recommended based on accurate matching of the recommended collection theme and the user interests, and then the target user group is continuously and dynamically optimized by using the feedback data, so that the recommendation accuracy and the interaction conversion rate are effectively improved.

Description

Technical Field

[0001] This application relates to the field of content recommendation technology, specifically to methods, apparatus, electronic devices, and program products for optimizing recommended objects. Background Technology

[0002] With the rapid development of digital content consumption, various platforms are pushing integrated content to users through recommendation collections and other means, which has become an important way to improve user stickiness and activity. However, when pushing recommendation collections to a wide range of users, platforms usually recommend the same collection to all users without any personalized differentiation, resulting in poor recommendation effectiveness. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, electronic device, and program product for optimizing recommended objects, in order to solve the problem of poor recommendation performance of recommendation collections.

[0004] In a first aspect, this application provides a method for optimizing recommended objects, comprising: obtaining topic features of a recommendation set, the recommendation set being used to represent multimodal recommendation information; obtaining a matching relationship between the topic features and the interest profile of a first recommended object, the interest profile being determined based on first interaction data of the first recommended object; filtering the first recommended object based on the matching relationship to obtain a second recommended object; recommending the recommendation set to the second recommended object, and obtaining second interaction data of the second recommended object on the recommendation set; and optimizing the second recommended object using the second interaction data.

[0005] In one optional implementation, optimizing the second recommended object using the second interaction data includes: processing the second interaction data to determine the evaluation index of the recommended collection after it is recommended; and optimizing the second recommended object based on the evaluation index.

[0006] In one optional implementation, the recommendation set includes at least one recommendation item; the second interaction data is processed to determine the evaluation metrics after the recommendation set is recommended, including: determining the target consumption completion rate of each recommendation item based on the content consumption records in the second interaction data; determining the target weighting coefficient corresponding to each target consumption completion rate based on the mapping relationship between consumption completion rate and weighting coefficient; determining the consumption duration weighted score corresponding to the recommendation set based on the superposition result of each target weighting coefficient; determining the preference expression rate corresponding to the recommendation set based on the number of preference expression operations in the second interaction data, and determining the attention conversion rate corresponding to the recommendation set based on the number of attention operations in the second interaction data; and determining the evaluation metrics based on the consumption duration weighted score, preference expression rate, and attention conversion rate.

[0007] In one optional implementation, the second interaction data is parsed to determine the operation behavior sequence corresponding to each second recommendation object; the frequency of occurrence of each operation behavior sequence is counted; and the operation behavior sequence whose occurrence frequency exceeds a preset frequency threshold and whose evaluation index exceeds a preset index threshold is determined as the target behavior sequence.

[0008] In one alternative implementation, the evaluation metrics and target behavior sequence are analyzed to generate optimization suggestions for the recommendation set; in response to the adjustment operation performed on the recommendation set based on the optimization suggestions, a modified recommendation set is obtained.

[0009] In one optional implementation, obtaining the matching relationship between topic features and the interest profile of the first recommended object includes: extracting behavioral feature data under the content consumption dimension, interactive operation dimension, and content topic preference dimension from the first interaction data; processing the behavioral feature data of the content consumption dimension to obtain the content preference value of the first recommended object for each content type; processing the behavioral feature data of the interactive operation dimension to obtain the interactive behavior intensity value of the first recommended object for multiple interactive types; processing the behavioral feature data of the content topic preference dimension to obtain the topic preference degree of the first recommended object for each content topic; constructing the interest profile of the first recommended object based on the content preference value, interactive behavior intensity value, and topic preference degree, and establishing the matching relationship between topic features and interest profile.

[0010] In one optional implementation, filtering the first recommended object based on the matching relationship to obtain the second recommended object includes: matching the target content topic represented by the topic feature with the topic preference degree in the interest profile to obtain a matching result; if the matching result indicates that the first recommended object is successfully matched, then determining whether the interaction behavior intensity value corresponding to the first recommended object exceeds a preset intensity threshold, and whether the content preference value of any content type exceeds a preset preference threshold, to obtain a judgment result; if the judgment result indicates that the interaction behavior intensity value exceeds the preset intensity threshold, and the content preference value of any content type exceeds the preset preference threshold, then the first recommended object is determined as the second recommended object.

[0011] Secondly, this application provides an optimization apparatus for recommended objects, comprising: a first acquisition module for acquiring topic features of a recommendation set, the recommendation set being used to represent multimodal recommendation information; a second acquisition module for acquiring the matching relationship between the topic features and the interest profile of a first recommended object, the interest profile being determined based on the first interaction data of the first recommended object; a filtering module for filtering the first recommended object based on the matching relationship to obtain a second recommended object; a recommendation module for recommending the recommendation set to the second recommended object and acquiring the second interaction data of the second recommended object regarding the recommendation set; and an optimization module for optimizing the second recommended object using the second interaction data.

[0012] Thirdly, this application provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the optimization method of the first aspect or any of its corresponding embodiments described above.

[0013] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to perform the optimized method of the first aspect or any of its corresponding embodiments.

[0014] Fifthly, this application provides a computer program product, including computer instructions for causing a computer to execute the optimized method of the recommended object of the first aspect or any corresponding embodiment described above.

[0015] The method for optimizing recommended objects provided in this application overcomes the limitations of traditional single-content recommendations by acquiring thematic features of a multimodal recommendation set, enabling a more comprehensive and accurate characterization of content semantics and sentiment. Constructing interest profiles based on actual user interaction data ensures the real-time nature and behavioral authenticity of the profiles, avoiding biases caused by reliance on static tags. Furthermore, by matching thematic features with interest profiles and filtering recommended objects based on this matching relationship, the recommendation process shifts from broad exposure to targeted selection based on feature matching, significantly improving the relevance between content and users. By collecting user interaction data after recommendations and using it for continuous optimization, a closed-loop learning mechanism of "deployment-feedback-iteration" is formed. This not only dynamically adapts to changes in user interests but also continuously improves deployment effectiveness and resource utilization efficiency in actual operation, thereby effectively improving recommendation accuracy and enhancing user experience. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this application, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of the first method for optimizing the recommended object according to an embodiment of this application; Figure 2 This is a schematic diagram of a homepage banner promotion according to an embodiment of this application; Figure 3This is a schematic diagram of a second process for optimizing the recommended object according to an embodiment of this application; Figure 4 This is a structural block diagram of the optimization apparatus for the recommended object according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that the information (including but not limited to user input information, such as information entered by the user into input boxes), data (including but not limited to data used for analysis, stored data, and displayed data, such as context code, all code of the current project, the service pressure corresponding to operations performed on all code of the current project, and the code development status of the current project), and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. For example, the context code, operations performed on all code of the current project, the corresponding service pressure, and the code development status involved in this application were all obtained with full authorization.

[0020] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0021] Currently, recommendation systems are widely used on various information platforms, aiming to achieve personalized content distribution by analyzing users' historical behavior. Common recommendation strategies typically rely on users' past interaction data (such as clicks, plays, and purchase records) to build interest models and push potentially interesting content to users based on these models. However, such methods often focus on responding to single types of content or single user behaviors, failing to fully integrate the characteristics of multimodal and thematic recommendation collections, and lacking a closed-loop optimization mechanism from "content topic—user interest—real-time feedback".

[0022] In related technologies, a typical approach is to employ a broad-area exposure strategy, pushing a unified set of recommendations (such as special topics, events, and multimodal information feeds) to a large number of users. While this method has broad coverage, it neglects the deep matching between individual interests and the theme of the set, resulting in limited reach accuracy, low user interaction rates, and difficulty in improving overall promotional efficiency (such as eCPM). Another improvement attempt is to introduce personalized recommendation models, targeting based on predicted click-through rates. However, its modeling process is often limited to single content types or simple user tags, failing to adequately identify composite recommendation sets containing multiple formats (such as music, text / images, and videos), failing to incorporate user consumption habits for content combinations into tiered filtering, and unable to continuously optimize based on multidimensional interaction metrics after the push notification.

[0023] The method for optimizing recommendation objects provided in this application overcomes the shortcomings of traditional wide-area exposure strategies that neglect the matching of content themes and user interests by acquiring the thematic features of recommendation collections and using multimodal, thematic recommendation information as the basis for recommendations. It constructs interest profiles based on users' historical interaction data and matches them with thematic features to achieve precise selection of target groups and improve reach accuracy. By pushing recommendation collections to selected users and collecting their subsequent interaction data, a closed-loop optimization mechanism of "content theme—user interest—real-time feedback" is formed. This allows for continuous optimization of recommendation objects based on interaction data after push notifications, achieving dynamic improvement of recommendation performance and iterative refinement of the user model.

[0024] According to an embodiment of this application, an optimization method for recommending objects is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] This embodiment provides an optimization method for recommending objects, which can be used in electronic devices, such as servers. Figure 1 This is a flowchart of an optimization method for a recommended object according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain the topic features of the recommendation set, which is used to represent multimodal recommendation information.

[0026] A recommendation collection refers to a multimodal content set planned and integrated by a platform (such as a music platform or a shopping platform), containing at least two types of content elements (such as text, audio, video, and images), used for joint recommendations to users. For example, it could be a music weekly planned by a music platform, which includes songs and selected text content (such as editor's notes, excerpts from music reviews, and themed stories). Thematic features refer to the classification or tag attributes of a recommendation collection, used to characterize its content attributes and audience targeting basis, such as content theme, style type, language category, applicable scenarios, and emotional tags. Recommendation information refers to the specific content elements contained in the recommendation collection, which can be one or more forms such as text, audio, video, and images. Specifically, thematic features of a recommendation collection refer to abstract tags or attributes extracted from the multimodal content contained in the collection that can characterize its core content and audience direction. These features can be acquired through content-based analysis, such as extracting text keywords using natural language processing, identifying styles or categories through audio / video analysis, and understanding visual themes through image recognition; or through manual or semi-manual annotation combined with prior knowledge or operational strategies, such as defining the content's domain, emotional tone, applicable scenarios, or target audience. Ultimately, these features are structured into a tagging system with one or more dimensions (such as theme, style, language, difficulty level, etc.) as a benchmark for subsequent matching with user interests.

[0027] Step S102: Obtain the matching relationship between topic features and the interest profile of the first recommended object. The interest profile is determined based on the first interaction data of the first recommended object.

[0028] The primary target audience refers to all potential users on the platform before optimization and filtering, i.e., the initially defined audience set. Interest profiles refer to multi-dimensional preference models built based on users' historical behavioral data, which may include user preferences for content types, interaction intensity, content theme preferences, and behavioral path patterns. Primary interaction data refers to users' past behavioral data on the platform, including content consumption records, interaction behaviors (such as clicks, plays, comments, likes, favorites, and shares), and dwell time. Specifically, obtaining matching relationships essentially involves calculating the correlation or fit between the theme features of the recommendation set and the interest profiles of the primary target audience (i.e., the potential user group). Interest profiles are constructed by analyzing users' historical primary interaction data (such as browsing, plays, likes, comments, and favorites), which includes the intensity and patterns of user preferences across multiple dimensions. Obtaining matching relationships can be done by comparing the user's interest tag vector with the theme feature vector of the recommendation set and calculating similarity, or by using more complex machine learning models to assess the probability of a specific theme's attractiveness to the user group. This process outputs a quantitative matching score or a classification matching result.

[0029] Step S103: Filter the first recommended object based on the matching relationship to obtain the second recommended object.

[0030] The second recommendation group refers to the highly matched user group selected from the first recommendation group based on the matching degree between the theme characteristics of the recommendation collection and the user's interest profile. Specifically, based on the matching relationship obtained above, a filtering strategy is executed to accurately locate the second recommendation group from the broad pool of first recommendation groups. The core logic of the filtering can be to set a threshold or sorting rules: for example, only users whose interest profile matches the theme characteristics of the current recommendation collection with a degree exceeding a preset threshold are selected; or, users are sorted in descending order according to their matching degree scores, and the top N% of users are selected. In addition, the filtering strategy may not be a single one, but a combination of multiple conditions, such as requiring both high matching degree and recent user activity, or preference for a certain specific interaction behavior pattern. Through this step, a transition from casting a wide net to precise targeting is achieved, ensuring that recommendation resources are concentrated on the user group most likely to generate positive feedback.

[0031] Step S104: Recommend the recommendation collection to the second recommendation object, and obtain the second interaction data of the second recommendation object for the recommendation collection.

[0032] Secondary interaction data refers to the real-time behavioral data generated by the second recommended user after receiving the recommendation collection. This data is used to evaluate the recommendation effect and subsequent optimization, including behaviors such as clicks, plays, interactions, and conversions. Specifically, after selecting the secondary recommended users, electronic devices will display the recommendation collection to these target users through platform-specified exposure channels (such as homepage banners, news feed recommendations, push notifications, etc.). For example, ... Figure 2As shown, when the recommended collection is a music weekly, the placement can be the main exposure position on the homepage banner (the dotted box in the figure), ensuring efficient reach of the content to core traffic entry points. The content format is a composite format, including a combination of song audio content and selected text content (such as editor's introduction, music review excerpts, and theme stories), achieving an integrated presentation of multimodal content. The specific placement logic is as follows: from the pre-selected second recommendation target (i.e., the high-matching user group), the target exposure sample for this placement is extracted according to a preset ratio or rule. Through the placement scheduling algorithm, it is ensured that each user identifier (UID) is exposed only once within the placement period of a single music weekly, to avoid excessive interference, and the system load and user experience are balanced through time window scheduling (such as uniform distribution). At the same time, a combination of exposure tracking and behavior tracking is adopted to automatically record logs at key nodes in the entire link of "seeing the recommendation (exposure) → clicking → playing → interacting (likes, comments, collections, etc.)", and these logs are aggregated through a unified log collection module to ensure that the user behavior response corresponding to each exposure can be completely and accurately collected, forming second interaction data for effect evaluation.

[0033] To achieve this process, electronic devices rely on a delivery scheduling system that is responsible for delivering content to a specified list of users at the right time and in the right place, and controlling the exposure strategy based on the logic described above.

[0034] Step S105: Optimize the second recommended object using the second interaction data.

[0035] Electronic devices define a set of core performance metrics (e.g., conversion rate to measure deep engagement, interaction duration to measure content appeal, and equivalent cost-per-million (eCPM) to measure commercial value) and use collected secondary interaction data to calculate these metrics. Secondly, based on the evaluation results, the electronic devices initiate optimization mechanisms. This can be model optimization: for example, using positive or negative feedback data from the current campaign as new training samples to feed back into the user interest profile model or matching algorithm used to calculate matching relationships, making future predictions more accurate; or strategy optimization: for example, automatically adjusting the matching threshold used when selecting secondary recommendation targets, updating the behavioral path definition of high-value users, or dynamically assigning different audience strategies to different content themes. Through this continuous cycle of "campaign-collection-evaluation-adjustment," it can continuously approach better recommendation results and achieve adaptive audience optimization.

[0036] The method for optimizing recommendation objects provided in this application overcomes the shortcomings of traditional wide-area exposure strategies that neglect the matching of content themes and user interests by acquiring the thematic features of recommendation collections and using multimodal, thematic recommendation information as the basis for recommendations. It constructs interest profiles based on users' historical interaction data and matches them with thematic features to achieve precise selection of target groups and improve reach accuracy. By pushing recommendation collections to selected users and collecting their subsequent interaction data, a closed-loop optimization mechanism of "content theme—user interest—real-time feedback" is formed. This allows for continuous optimization of recommendation objects based on interaction data after push notifications, achieving dynamic improvement of recommendation performance and iterative refinement of the user model.

[0037] This embodiment provides an optimization method for recommending objects, which can be used in electronic devices, such as servers. Figure 3 This is a flowchart of an optimization method for a recommended object according to an embodiment of this application, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain the topic features of the recommendation set. The recommendation set is used to represent multimodal recommendation information. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0038] Step S302: Obtain the matching relationship between topic features and the interest profile of the first recommended object. The interest profile is determined based on the first interaction data of the first recommended object.

[0039] Specifically, step S302 includes: Step a1: Extract behavioral feature data from the first interaction data under the dimensions of content consumption, interactive operation, and content theme preference.

[0040] The content consumption dimension refers to the data category corresponding to the broad category of user behavior of passively receiving or watching content on the platform. It focuses on what users "watched," "listened to," "read," and "for how long." For example, records of browsing articles, playing videos, listening to audio, and reading comments all fall under this dimension. The interaction operation dimension refers to the data category corresponding to user behavior of actively expressing preferences, attitudes, or engaging in social communication regarding platform content or others. It focuses on what interactions users performed. For example, records of actions such as liking, disliking, saving, forwarding, commenting, replying, sharing, and following all fall under this dimension. The content theme preference dimension refers to the data category used to characterize users' preferences for the inherent attributes or categories of content. It focuses on what categories / styles of content users like. These themes are abstract labels for content, such as technology, entertainment, sports, classical music, rock music, and Chinese content. Specifically, extracting behavioral feature data from the first interaction data first requires defining and identifying behavioral categories under different dimensions. Electronic devices will preprocess and classify the raw user behavior logs. For content consumption, the system filters out user behavior records of passively receiving content, such as the start and end times, content identifiers, and consumption progress of events like content playback, browsing, and reading. For interaction, it extracts records of users actively expressing attitudes or engaging in social activities, such as the type, timestamp, and target of actions like liking, commenting, saving, sharing, and following. For content topic preference, the system associates these behaviors with the metadata (such as category tags and topic tags) carried by the content to extract the content topic tags corresponding to each user consumption or interaction. This process typically relies on data cleaning, event type filtering, and metadata association techniques to ultimately form a structured, dimension-based behavioral feature dataset.

[0041] Step a2: Process the behavioral feature data of content consumption dimension to obtain the content preference values ​​of the first recommendation object for each content type.

[0042] Content type refers to the classification of different forms or carriers of content within a platform; it answers the question "What form is the content?". For example, when a recommended collection is a weekly music review, the corresponding content types could include listening to songs, music reviews, notes, banner consumption, etc. Content preference value is used to characterize a user's relative liking or consumption habit intensity for a specific content type. It is calculated by analyzing users' historical behavioral data on different content types across the content consumption dimension; a higher value indicates a stronger preference for that type of content. Specifically, the core of processing behavioral characteristic data on the content consumption dimension to obtain the content preference value is to quantify the relative investment of users in various types of content. First, the number of historical consumption behaviors or the total duration for each user under different content types are counted. Then, normalization is usually used, such as calculating the proportion of a user's consumption on a certain content type to their total consumption across all content types. More refined models can assign different weights to different types of consumption behavior (e.g., full playback has a higher weight than short stays), and calculate the preference value through weighted summation. Finally, each user will receive a vector, where each component represents their preference for a particular content type; a higher value indicates a stronger preference.

[0043] In some alternative implementations, the content preference value can be calculated using the following formula:

[0044] in, Indicates user In terms of content type The next Number of times a class of behaviors. The weight of the corresponding behavior can be initially set to 1, and can be dynamically adjusted according to the business results.

[0045] Step a3: Process the behavioral feature data of the interactive operation dimension to obtain the interaction intensity values ​​of the first recommended object for multiple interaction types.

[0046] Interaction type refers to the specific types of interactive actions within the dimension of interactive operations. For example, liking, commenting, saving, and following are each independent interaction types. The interaction behavior intensity value is used to characterize a user's overall interaction activity or depth. It is calculated by weighted summation of the user's historical behavior frequency or quantity across different interaction types; a higher value indicates a greater inclination to participate in the interaction. Specifically, behavioral characteristic data from the interactive operation dimension is processed to obtain the interaction behavior intensity value, aiming to comprehensively evaluate a user's interaction activity and depth. Electronic devices statistically analyze the frequency or number of times each user performs various interactive operations (such as liking, commenting, saving, and forwarding) over a period of time. Then, a weighted summation model is used to integrate the statistical values ​​of these different interaction types into a single intensity value. Each interaction type in the model is assigned a weight coefficient, which reflects the user's level of engagement represented by that interactive behavior (e.g., commenting has a higher weight than liking). The weighted values ​​of all interaction types are summed to obtain the user's interaction behavior intensity value; a higher value indicates a greater inclination to engage and interact deeply.

[0047] In some alternative implementations, the intensity value of the interactive behavior can be calculated using the following formula:

[0048] in, For the number of comments, For the number of likes, To the number of times it is collected, This represents the number of replies. , , , The corresponding weighting coefficient can be initially set to 1.

[0049] Step a4: Process the behavioral feature data of the content topic preference dimension to obtain the topic preference degree of the first recommendation object for each content topic.

[0050] Content themes refer to the specific categories or tags used when categorizing or labeling content. For example, artificial intelligence, basketball games, and folk music can all be content themes. Theme preference is used to characterize a user's degree of liking for a particular content theme. It is calculated by analyzing the user's historical behavioral data (such as number of purchases, completion rate, etc.) of content tagged with that theme that they have consumed. Specifically, the theme preference dimension of behavioral feature data is processed to obtain the theme preference score, the key being to quantify the intensity of a user's liking for a specific content theme. Electronic devices analyze all content items that users have historically consumed and summarize the theme tags associated with these content items. For each theme, user behavior indicators related to that theme are calculated, such as: the total number of times content on that theme has been consumed, the total duration, the average completion rate, and whether there is deep interaction such as saving or adding to favorites. These raw indicators are fused through a calculation model (which can be a simple weighted average or a more complex machine learning model), and a time decay factor can be introduced to make recent behavior have a greater impact on preference. Finally, each user's preference score for each content theme is output, with a higher score indicating that the user is more interested in that theme.

[0051] In some alternative implementations, theme preference (such as genre preference) can be calculated using the following formula:

[0052] in, Indicates user preference for music style The The behavior of playing the first song As a weight for playback completion rate, This is the time decay coefficient.

[0053] Step a5: Based on content preference value, interaction behavior intensity value and topic preference degree, construct the interest profile of the first recommended object, and establish the matching relationship between topic features and interest profile.

[0054] Building an interest profile involves integrating calculated quantitative indicators from multiple dimensions, such as content preference values, interaction intensity values, and topic preference degrees, into a structured user model. Specifically, this model can be represented as a multi-dimensional vector or a feature set containing multiple fields, comprehensively depicting the user's characteristics in content consumption habits, interaction tendencies, and topic interests. Establishing a matching relationship involves calculating the similarity between the topic features (also one or more topic tags) of the recommended collection and the topic preference vector in the user's interest profile. This is typically achieved by calculating cosine similarity, Euclidean distance, or directly finding the user's preference score for that specific topic. The result of the matching relationship is a quantitative matching score, or a match / mismatch status based on a threshold, used to assess the degree of fit between the collection and the user's interests.

[0055] The method for optimizing recommended users provided in this application first extracts three key behavioral features from the first interaction data: content consumption dimension, interaction operation dimension, and content theme preference dimension. This division comprehensively covers the main aspects of user-content interaction from three complementary perspectives: what is consumed, how is it interacted, and what themes are preferred, laying a comprehensive data foundation for constructing a three-dimensional interest profile. Secondly, the data for each dimension is specifically processed to generate corresponding quantitative indicators—content preference value, interaction behavior intensity value, and theme preference degree. This process transforms the raw behavioral data into structured, comparable numerical features, enabling the previously vague user interests to be accurately characterized and measured. Furthermore, based on the quantitative results of these three dimensions, an interest profile is comprehensively constructed and its matching relationship with theme features is established. This profile integrates user type preferences, interaction tendencies, and theme interests, forming a multi-factor weighted composite model that more profoundly reveals the user's intrinsic interest-driving factors. This achieves efficient and refined transformation of user interests from data to model, providing a technical foundation for accurate matching.

[0056] Step S303: Filter the first recommended object based on the matching relationship to obtain the second recommended object.

[0057] Specifically, step S303 includes: Step b1: Match the target content theme represented by the theme features with the theme preference in the interest profile to obtain the matching result.

[0058] The matching result is the output generated by comparing the topic features (i.e., the content topics to which the recommended collection belongs) with the user's topic preferences. Specifically, matching topic features with interest profiles is a direct query and comparison process. First, the target content topics (one or more tags) to which the recommended collection belongs are obtained from its metadata. Then, the topic preference scores corresponding to these topics are found in the target user's interest profile. The matching logic can be exact matching: if the user's interest profile contains the topic and the preference score exceeds a preset matching threshold, it is considered a successful match; or it can be fuzzy matching or Top-N matching: the similarity between the collection's topic and all of the user's interest topics is calculated, and if the highest similarity exceeds the threshold, it is considered a match. The matching result is usually a Boolean value (true / false) or a discrete matching level (such as high, medium, low).

[0059] Step b2: If the matching result indicates that the first recommended object is successfully matched, then determine whether the intensity value of the interaction behavior corresponding to the first recommended object exceeds the preset intensity threshold, and whether the content preference value of any content type exceeds the preset preference threshold, and obtain the judgment result.

[0060] The judgment result is a logical conclusion reached after a successful match, further checking whether the user meets other filtering conditions (such as whether the interaction intensity value and content preference value exceed the threshold). Specifically, based on a successful topic match, additional condition judgments are performed to further filter out high-value, high-potential users. This process is a sequential logical AND operation. The user's interaction intensity value is read and compared with a preset intensity threshold to determine if it meets the high-interaction standard. The user's content preference value vector is traversed to check if there is a preference value for at least one content type exceeding the corresponding preset preference threshold, i.e., to determine if the user has a strong consumption habit for a certain type of content. The results of these two condition checks (yes / no) are then logically ANDed. The final judgment result is a Boolean value: the judgment is passed if and only if both the interaction intensity value and the preference value for at least one content type exceed the threshold; otherwise, it is failed.

[0061] Step b3: If the judgment result indicates that the intensity value of the interactive behavior exceeds the preset intensity threshold, and the content preference value of any content type exceeds the preset preference threshold, then the first recommended object is determined as the second recommended object.

[0062] When the judgment result is "pass," it means that the first recommended user simultaneously meets the three core screening criteria: interest in the recommended topic, high interaction activity, and clear content consumption preferences. The electronic device will then mark this user from a potential, broad audience pool (the first recommended user) and add them to the precise target audience set that this recommendation campaign ultimately aims to reach. This set is the second recommended user. This process essentially applies multi-dimensional, multi-threshold logical rules to user characteristic data, enabling the automatic and precise identification of the high-quality target audience most likely to engage deeply with complex content from a massive user base.

[0063] The optimization method for recommending users provided in this application first performs topic matching, comparing the topic features of the recommended content with the topic preferences in the user profile to ensure that candidate users have a basic interest in the content topic, completing the first layer of coarse screening. Based on this, instead of directly accepting all users with topic matching, two parallel refined judgments are further performed: first, checking whether their interaction intensity value exceeds a preset threshold to filter out users with high activity and strong interaction willingness, ensuring the interaction potential of the target group; second, checking whether they have a high preference for any content type, ensuring that users have significant content consumption habits. Only when a user simultaneously meets the three conditions of successful topic matching, high interaction intensity, and strong content preference will they be finally determined as a second recommended user. This screening logic, through multiple constraints, aggregates three key characteristics—interest in the topic, habitual active interaction, and high content consumption tendency—into the same target group, making the selected second recommended users not only likely to click on content but also likely to complete in-depth playback, interaction, and even follow behavior, greatly improving the effectiveness of subsequent recommendations.

[0064] In some optional implementations, the selection logic for the first recommended users is a multi-condition combination filter, specifically including topic matching (the user's preference for the topic of the recommended collection exceeds a preset threshold), interaction intensity (the user's interaction intensity exceeds a preset intensity threshold), and content preference (the user's preference for at least one content type (such as music reviews, notes, or banner consumption) exceeds a preset preference threshold). Users who meet all the conditions will be selected as the second recommended users. The user selection results are updated daily to reflect the latest changes in user behavior.

[0065] Step S304: Recommend the recommendation collection to the second recommendation object, and obtain the second interaction data of the second recommendation object for the recommendation collection. For details, please refer to [link to details]. Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0066] Step S305: Optimize the second recommended object using the second interaction data.

[0067] Specifically, step S305 includes: Step S3051: Process the second interactive data to determine the evaluation metrics after the recommended collection is recommended.

[0068] Evaluation metrics are core quantitative standards used to quantify the overall commercial effectiveness and user engagement of a recommendation campaign within a target audience. For example, eCPM (Effective Cost Per Mille) is a composite metric. Specifically, user behavior records directly related to the recommendation campaign are extracted from the second-level interaction data. These records, collected through a tracking system, encompass the complete chain from user exposure to subsequent deep interaction. The processing includes data cleaning, deduplication, correlation, and aggregation. Then, based on predefined metric calculation logic, core elements are extracted from the cleaned data: for example, summarizing all exposure events to obtain total exposure; analyzing playback or reading behavior to calculate consumption depth; statistically analyzing likes, favorites, etc., to obtain preference expression data; and statistically analyzing attention behavior to obtain relationship establishment data. Finally, these elements are calculated using a comprehensive formula (such as a weighted product or composite function) to output a quantitative evaluation metric value. This value aims to comprehensively reflect the overall effect of the recommendation campaign in terms of user engagement depth, emotional identification, and long-term value conversion.

[0069] In some alternative implementations, the recommendation set includes at least one recommendation item, and step S3051 above includes: Step c1: Based on the content consumption records in the second interaction data, determine the target consumption completion rate for each recommendation item.

[0070] Content consumption records refer to detailed log data generated by users during the consumption of recommended content collections, including records of consumption progress or duration for each independent content unit within the collection. A recommended item is the smallest independent content unit constituting a recommended collection. A recommended collection consists of at least one recommended item. For example, a single article in a collection of themed articles, or a single song in a music weekly. Target consumption completion is the percentage of the actual length consumed by a user (e.g., playback time, reading progress) to the total length of that recommended item, used to quantify the depth of user consumption of that single content item. Specifically, determining the target consumption completion for each recommended item requires fine-grained quantification of user consumption behavior on individual content units. Electronic devices analyze consumption records for each recommended item (e.g., songs, videos, articles, etc.) in the second interaction data. These records typically include user identifier, recommended item identifier, consumption start time, end time, or consumption progress percentage. For each recommended item, the progress of each user's current consumption is calculated (e.g., playback time / total duration, number of pages read / total number of pages, etc.). Then, an aggregation strategy can be used to determine the overall target consumption completion rate of the recommendation: for example, calculating the average consumption progress of all users who have consumed the recommendation; or considering deduplication, calculating the proportion of users whose consumption progress exceeds a certain threshold (such as 50%). This completion rate represents the average ability of the recommendation to attract users to complete their consumption.

[0071] Step c2: Based on the mapping relationship between consumption completion rate and weighting coefficient, determine the target weighting coefficient corresponding to each target consumption completion rate.

[0072] The target weighting coefficient is a weight value assigned to different target consumption completion ranges according to preset rules. Higher completion rates typically receive a higher weighting coefficient to highlight the value of in-depth consumption. Specifically, electronic devices internally use a preset rule table or function to map consumption completion ranges to different weighting coefficients. This mapping aims to distinguish the value difference between superficial and in-depth consumption. After calculating the target consumption completion rate for a recommended item, this value is compared with the preset range. For example, as shown in Table 1, the rules stipulate: 0-10% completion corresponds to a weight of 0, 10%-30% to a weight of 0.3, 30%-60% to a weight of 0.6, and above 60% to a weight of 1.0. Based on the range the target consumption completion rate falls into, the corresponding target weighting coefficient is automatically retrieved or calculated. This coefficient is a scalar used to assign different levels of importance to different levels of consumption behavior in subsequent comprehensive score calculations.

[0073] Table 1

[0074] Step c3: Based on the summation of the weighted coefficients of each objective, determine the consumption duration weighted score corresponding to the recommended set.

[0075] The Consumption Duration Weighted Score is a comprehensive metric that quantifies a user's consumption depth across the entire recommendation collection. It is calculated by mapping the target consumption completion rate of each recommendation item in the collection to a target weighted coefficient, and then summing the weighted coefficients of all recommendations. Specifically, the target weighted coefficients calculated for each recommendation item (or for each user's consumption behavior for each recommendation item) are summed. If the calculation is performed at the user-recommendation item granularity, the sum of the weighted coefficients for each user for the collection can be calculated first, and then averaged over all users; or all weighted coefficients for all users can be directly summed. This summation (or averaging) result is the Consumption Duration Weighted Score. It avoids the bias that may arise from simply summing the original duration, emphasizing the value of content items that trigger deep user consumption, thus more accurately measuring the overall content attractiveness of the collection.

[0076] For example, when the recommended collection is a weekly music magazine, its consumption time-weighted score is specifically reflected as a playback time-weighted score. This score is calculated by assigning a corresponding weight coefficient to each user's playback of songs within the magazine based on their playback completion rate, and then summing the weights of all playback actions. The specific formula is as follows:

[0077] in, For the first The completion weight corresponding to each playback action is determined based on the completion percentage of this playback and with reference to a preset mapping table (e.g., completion < 10% corresponds to weight 0, completion 10%-30% corresponds to weight 0.3, etc.). This represents the total number of plays generated by all target users for this music weekly.

[0078] Step c4: Determine the preference expression rate corresponding to the recommendation set based on the number of preference expression operations in the second interaction data, and determine the attention conversion rate corresponding to the recommendation set based on the number of attention operations in the second interaction data.

[0079] The number of preference expression actions refers to the total number of interactive actions performed by users after they have shown interest in a recommended collection, expressing positive emotions or saving their interest, such as the total number of "likes". The preference expression rate is used to measure the ratio of positive emotional feedback elicited by recommended content from users. Specifically, the preference expression rate is a metric that measures the ratio of positive emotional feedback elicited by recommended content from users. First, all preference expression action records, such as the logs of "like" events, are filtered from the secondary interaction data. Then, the total number of these actions is counted to obtain the number of preference expression actions. When calculating the preference expression rate, this number is usually correlated with the total exposure or total clicks of the recommendation campaign. For example, the calculation method could be: Preference Expression Rate = (Total Number of Preference Expression Actions / Total Exposures) × 1000 (for per mille), or Preference Expression Rate = (Number of Users Who Performed at Least One Preference Expression Action / Total Number of Exposures) × 100% (for percentage). The specific formula depends on the business definition, but its core is to establish a conversion relationship between exposure and positive emotional feedback.

[0080] The number of follow actions represents the total number of interactive actions a user performs to establish a long-term subscription relationship after showing interest in a recommended collection. Examples include following an author, following a topic, and subscribing to a column. The follow conversion rate measures the percentage of recommended content that successfully encourages users to establish a long-term connection. Specifically, the follow conversion rate quantifies the effectiveness of recommended content in encouraging users to establish long-term subscription relationships. First, all follow action records are extracted from the second interaction data, such as following creators, subscribing to columns, and following topics. Then, the total number of these actions is counted to obtain the number of follow actions. The calculation logic for the follow conversion rate is similar to that of the preference expression rate, typically comparing the number of follow actions to a base number. For example, the calculation method could be: Follow Conversion Rate = (Total Number of Follow Actions / Total Exposures of the Recommended Collection) × 1000 (commonly used in eCPM calculations), or Follow Conversion Rate = (Number of Users Who Performed at Least One Follow Action / Total Number of Clicks) × 100%. This metric reflects the potential of recommended content to promote long-term user retention and continuous interaction.

[0081] Step c5: Determine the evaluation indicators based on the consumption duration weighted score, preference expression rate, and attention conversion rate.

[0082] The process of determining the evaluation metric involves fusing the three metrics—representing different dimensions of consumption depth, emotional feedback, and long-term value—into a single, comparable comprehensive value using a predefined mathematical model. For example, when the evaluation metric is eCPM, the mathematical model used is: Evaluation Metric = (Consumption Duration Weighted Score × Preference Expression Rate × Attention Conversion Rate) / Exposure Volume × 1000 Electronic devices input the calculated consumption time-weighted score, preference expression rate, and attention conversion rate into the formula, perform multiplication, divide by the total exposure, and amplify by a thousand times to finally obtain the evaluation index value in units of equivalent revenue per thousand exposures.

[0083] In the above implementation, the recommended collection is broken down into its constituent recommendation items, and the target consumption completion rate of each recommendation item is calculated by analyzing content consumption records. This design allows the evaluation to penetrate the entire collection, revealing the user's true consumption depth of specific content and avoiding the obscuring of high-value items by the overall data. Secondly, a mapping relationship between consumption completion rate and weighting coefficients is introduced, and different weights are assigned to different completion rate intervals. This mechanism scientifically acknowledges that deep consumption (such as high-completion playback) has higher user interest value than shallow reach, so that the final calculated consumption duration weighted score is no longer a simple accumulation of time, but a weighted measure of the user's effective stay and true interest. Furthermore, two key behavioral dimensions, preference expression rate and attention conversion rate, are also incorporated in parallel. The former quantifies the user's immediate emotional feedback (such as likes and hearts), while the latter measures the user's ability to convert into long-term interest bonds (such as following), thus jointly covering the entire user psychological conversion stage from shallow interaction to deep recognition. Ultimately, the evaluation metric is determined by combining the weighted score representing consumption depth, the preference expression rate representing immediate emotional feedback, and the attention conversion rate representing long-term stickiness. This application utilizes a multi-dimensional, weighted, and sophisticated computational model to transform complex user post-hoc behaviors into a composite evaluation metric that simultaneously measures consumption depth, emotional resonance, and long-term value. This allows subsequent optimization decisions to be based on a rich and accurate performance profile, significantly improving the scientific rigor and effectiveness of the optimization direction.

[0084] Step S3052: Optimize the second recommended object based on the evaluation indicators.

[0085] Optimizing the second-tier recommended users based on evaluation metrics is a process of iteratively refining the target audience selection strategy using performance feedback data. Specifically, the electronic device compares and analyzes the evaluation metrics calculated for this recommendation campaign with historical baselines or metrics from different audience segments. Optimization actions mainly include: 1) Audience Package Adjustment: For a certain type of second-tier recommended users (those with specific interest profile characteristics) whose evaluation metrics are significantly higher than the average, the electronic device marks them as high-value users, allowing for priority use or expanded targeting in subsequent recommendations on similar topics. 2) Filtering Rule Optimization: For underperforming users, the electronic device back-analyzes their interest profile characteristics (such as topic preference, interaction intensity, and content preference) and attempts to adjust the preset intensity or preference thresholds used in step S303, making future filtering conditions more stringent or targeted, thereby eliminating low-potential users and improving the overall quality of the audience package. This forms a closed-loop optimization mechanism of "targeting-evaluation-adjustment".

[0086] In some optional implementations, performance analysis reports are generated periodically based on evaluation metrics. The report cycle can be bi-weekly and includes performance comparison data for different themes and different demographic groups.

[0087] The method for optimizing recommended objects provided in this application, by processing the second interaction data and determining evaluation indicators, extracts the complex user behavior data into core metrics that objectively and comprehensively reflect the recommendation effect, thus establishing optimization decisions on a solid foundation of quantitative analysis rather than subjective experience. Secondly, optimizing the second recommended objects based on these specific evaluation indicators makes the optimization operation clearly targeted and operable. The accuracy of the currently selected second recommended object group can be determined based on the performance of the indicators, and adjustments can be made accordingly. This not only achieves standardization and automation of the optimization process but also, by introducing the evaluation indicators as a key intermediate layer, makes the iterative optimization of the entire recommendation system observable and controllable, significantly improving the scientific nature of the optimization strategy and the effectiveness of the system's self-evolution.

[0088] In some optional implementations, the optimization method for the above-mentioned recommended objects further includes: Step d1: parse the second interaction data and determine the sequence of operation behaviors corresponding to each second recommendation object.

[0089] An action sequence is a chain of discrete actions performed by a user in chronological order during interaction with a recommended collection. It describes the specific path from user exposure to subsequent behavior, such as "click → play → like → follow". One or more high-value action path templates are pre-defined or configured using rules. These templates describe ideal action sequences from exposure to deep interaction, such as "play → interact → like", "play → like → follow", "play → interact → follow", etc. Electronic devices refer to these templates for matching or similarity calculations when counting the frequency of action sequences.

[0090] Specifically, determining the sequence of operational behaviors requires sorting and pattern recognition of a single user's discrete behaviors over a period of time. The electronic device extracts logs of all behavioral events generated by that user from the second interaction data, starting from the moment they receive the current recommendation exposure, on a per-secondary recommendation object (user ID) basis. These logs include behavior types (such as exposure, click, start playback, pause, resume playback, like, favorite, comment, follow) and timestamps. The electronic device first sorts these events according to the order of their timestamps. Then, irrelevant behaviors are filtered out, and the behavior types are abstracted into codes, forming a list of behavior codes arranged by time. This is the sequence of operational behaviors for that user in this instance.

[0091] Step d2: Count the frequency of occurrence of each operation sequence.

[0092] Frequency of occurrence refers to the number of times a specific sequence of user actions is observed within a statistical period when analyzing large amounts of user data. After obtaining the action sequences of all second-recommended users, the electronic device needs to identify common patterns, i.e., statistical frequency. Specifically, it iterates through all user sequences, grouping identical sequences (i.e., every item in the action code list and its order are the same) into the same category. Then, it calculates the number of times each category of sequences appears; this number is the frequency of occurrence for that category. For example, if there are 1000 users, and 50 users have the sequence "click → play → like → follow," then the frequency of occurrence for that sequence is 50. By analyzing frequency, the most typical behavioral paths within a user group can be identified.

[0093] Step d3 identifies the sequence of operational behaviors that occurs more frequently than a preset frequency threshold and whose evaluation metrics exceed a preset metric threshold as the target behavior sequence.

[0094] Target behavior sequences refer to sequences of operational behaviors identified as having high value. Specifically, electronic devices filter out sequences from all statistically analyzed operational behavior sequences that occur more frequently than a preset threshold; these are common patterns. Next, for these common sequences, the value generated by the user groups following them needs further examination. The electronic device calculates the average evaluation metric (such as average eCPM) for the user subset corresponding to each sequence. Then, only sequences whose average evaluation metric also exceeds another preset threshold are retained. Sequences that simultaneously meet the conditions of high frequency and high value are formally identified as target behavior sequences. They identify ideal user interaction paths that are both widely occurring and generate high returns.

[0095] In the above implementation, by parsing the second interaction data, the ordered sequence of actions performed by each second recommended object after contacting the recommended content is reconstructed. This achieves the transformation from discrete action points to continuous action paths, and can completely depict the typical user interaction process. Secondly, by statistically analyzing the frequency of occurrence of each action sequence, accidental and individual action combinations can be filtered out, focusing on those high-frequency common paths commonly used by a large number of users, ensuring the universality and representativeness of the discovered patterns. Setting dual screening conditions—an occurrence frequency exceeding a preset frequency threshold and an evaluation index exceeding a preset index threshold—not only requires the target action sequence to be common (high frequency) but also requires it to be effective (high evaluation index), thereby accurately identifying target action sequences that are not only frequently executed but also truly bring high-quality interaction results (such as high consumption and high conversion). This application can intelligently discover and lock onto the key user behavior paradigms that drive successful recommendations, providing a valuable, data-driven decision-making basis for a deeper understanding of user preferences, optimizing content design, and adjusting recommendation strategies.

[0096] In some optional implementations, the optimization method for the above-mentioned recommended objects further includes: Step e1 involves analyzing the evaluation metrics and target behavior sequences to generate optimization suggestions for the recommended set.

[0097] Optimization suggestions are automatically generated strategic hints on how to improve future recommendation performance, based on the analysis of evaluation metrics and target behavior sequences. Specifically, electronic devices comprehensively analyze the composition and distribution of overall evaluation metrics, combined with identified characteristics of target behavior sequences. For example, if a low evaluation metric is due to a low conversion rate, and the target behavior sequence shows that "play → like → follow" is a high-conversion path, but most users currently stop at playing, the analysis might conclude that content needs to be strengthened to encourage users to transition from playing to liking. Based on such insights, electronic devices can automatically generate structured optimization suggestions, such as adding interactive guidance elements (e.g., copy encouraging liking) within collections, optimizing the content ending to strengthen the motivation to follow, adjusting the order of recommended items, and placing the most relatable content first. These suggestions will target specific and actionable aspects such as content planning, presentation format, or delivery strategy.

[0098] Step e2, in response to the adjustment operation performed on the recommended set based on the optimization suggestions, yields the modified recommended set.

[0099] Once optimization suggestions are generated (whether provided to human operations staff or an automated system), the entity responsible for content management will respond to these suggestions and modify the original recommendation collection. These adjustments include replacing some recommendations, modifying titles or cover text, adjusting content layout or order, and adding guiding text or buttons. These operations are performed through a content management system (CMS) or similar configuration tools. Once the adjustments are saved and published, a new version of the original recommendation collection is created—the modified collection. This new collection can be used for a new round of A / B testing to compare its performance with the original version, or it can be directly deployed as the optimized product, thus completing a full optimization loop from analysis to action to validation.

[0100] In the above implementation, by performing correlation analysis on evaluation indicators and target behavior sequences, specific optimization suggestions for the recommendation set can be automatically generated. This process combines quantitative evaluation results with high-value user behavior patterns, ensuring that the generation of optimization suggestions is no longer based on fuzzy experience or a single indicator, but rather stems from deep data mining of which content features or structures are associated with better user effects, thus ensuring the scientific nature and relevance of the suggestions. Secondly, in response to these data-driven optimization suggestions, adjustments are directly performed on the recommendation set, generating a modified recommendation set. This mechanism ensures that the optimization logic no longer remains at the level of analysis reports, but can automatically and promptly transform into iteration and improvement of the recommended content itself, forming a rapid response closed loop of "analysis-suggestion-adjustment". This application not only greatly improves the efficiency and agility of content operation and reduces the delay and subjective bias of manual decision-making, but also enables the recommendation set to dynamically evolve and continuously optimize based on real user feedback and behavior patterns, thereby achieving self-improvement and a virtuous cycle of recommended content quality at the system level.

[0101] In some optional implementations, the electronic device supports simultaneous audience targeting and delivery for multiple recommendation collections with different themes (e.g., parallel language-based weekly magazines, scene-based weekly magazines, mood-based weekly magazines, etc. on a music streaming platform). Specifically, the electronic device can simultaneously acquire the theme characteristics of each of the multiple recommendation collections, and based on the method of the embodiments of this application, independently select its corresponding second recommendation object for each recommendation collection, and then deliver each recommendation collection to its corresponding second recommendation object. Through multi-theme parallel targeting, system resources can be efficiently utilized to achieve synchronous and accurate promotion of diverse thematic content within the platform.

[0102] In some optional implementations, to further improve the matching degree between the recommended collection and the target users, artificial intelligence technology can be introduced when generating the content of the recommended collection. Specifically, based on the group interest profile characteristics of the target second recommended audience (e.g., shared text preferences, visual style preferences), AI models can be used to automatically generate or optimize elements such as the introductory text, cover copy, and accompanying images of the recommended collection, making them more aligned with the aesthetics and interests of the target audience in terms of style and content. By combining precise audience selection with personalized content generation, dual optimization from people to content can be achieved, thereby more effectively improving the depth of user interaction and satisfaction.

[0103] In the following embodiments, the optimization method for the above-mentioned recommended objects will be illustrated by taking the application scenario of large-scale online music streaming platforms delivering music weekly as an example.

[0104] Each week, the platform's editorial team plans and produces several themed music weekly publications. Each publication revolves around a core theme (such as emerging Chinese folk music, bedtime instrumental music, or classic rock retrospectives), selecting 3-5 songs that fit the theme, and accompanying them with in-depth text introductions, music reviews, or musician stories written by the editors, forming a multimodal collection of recommendations. These weekly publications aim to deeply convey the value of music culture, enhance user engagement and interaction, and their promotional effectiveness directly impacts the platform's content ecosystem activity and commercial efficiency.

[0105] When the platform was preparing to launch a new weekly magazine called "Urban Night Owls: Ambient Electronic Music," it first obtained the magazine's thematic characteristics through its Content Management System (CMS), including preset tags: electronic music, ambient sound, nighttime setting, and immersive experience. Next, it initiated a user segmentation process: based on the platform's users' first interaction data over the past 30 days (listening to music, reading music reviews, viewing notes, consuming banners, etc.), it calculated each user's thematic preference for electronic and ambient music, their content preference values ​​for content types such as music reviews and notes, and their overall interaction intensity.

[0106] Based on a pre-defined selection logic, users who simultaneously meet all three of the following conditions are automatically selected from hundreds of millions of primary recommendation targets (all active users) to form secondary recommendation targets: 1) Their preference for electronic or ambient music exceeds threshold A; 2) Their interaction intensity exceeds threshold B, indicating they are deeply interactive users who enjoy commenting and liking; 3) Their preference for music reviews or note content types exceeds threshold C, indicating they have a habit of reading text content. This combination of conditions ensures that the selected users are not only interested in music themes but are also highly likely to consume and interact with the weekly composite content format. This selection result is updated daily to reflect the latest changes in user interests.

[0107] After the selection process is complete, the "Urban Night Owls: Ambient Electronic Music" weekly magazine is precisely placed on the main banner of the homepage during peak hours the following day, using the campaign scheduling module. It is only displayed to the selected secondary target audience, ensuring that each user's UID is only exposed once during this promotional period. Simultaneously, the end-to-end tracking system begins working, comprehensively recording a series of secondary interaction data from the target user's initial view of the banner to clicking into the magazine's details page, playing songs, reading text, clicking the heart icon, and following the musician.

[0108] After the one-week campaign period ends, the eCPM evaluation and feedback module is activated. It first calculates the overall performance metrics for this campaign: based on the playback records of all exposed users, a playback duration-weighted score is calculated according to playback completion rate (e.g., a full playback has a weight of 1.0); the "heart rate" is calculated as the ratio of the total number of "likes" to the number of impressions; and the "follower conversion rate" is calculated as the ratio of the total number of "followers" to the number of impressions. Finally, the eCPM for this weekly campaign is calculated using the formula (playback duration-weighted score × heart rate × follower conversion rate / impressions × 1000).

[0109] Based on the evaluation results, the system automatically enters the optimization and retraining phase. If the eCPM of this campaign is higher than the historical baseline for similar themes, the user demographics selected this time (such as specific theme preferences and interaction intensity combinations) will be solidified into a high-value ambient electronic music audience tag package for direct reuse in future similar themes. Simultaneously, analysis of user behavior logs reveals that the frequency and corresponding average eCPM of the behavior path "play the entire song → read the review → click the heart" are significantly higher than other paths, thus marking it as a high-value behavior pattern.

[0110] Finally, based on comprehensive evaluation metrics and high-value behavioral patterns, an optimization suggestion report was generated, indicating to the operations team: "For the 'ambient electronic music' theme, users have a high acceptance of in-depth text content (music reviews), and the 'play-read review-like' path is an efficient conversion path. It is recommended that future weekly issues increase the proportion and placement of high-quality music reviews, and add prompts guiding users to read music reviews on the song playback page." The operations team adopted the suggestions and adjusted the content structure of the next issue of the electronic music weekly. The revised version of the weekly will undergo A / B testing with the original version, thus completing a full data-driven closed loop from "precise selection → targeted delivery → quantitative evaluation → strategy optimization."

[0111] The optimization method for recommending users provided in this application not only greatly improves the efficiency and revenue (eCPM) of weekly promotions, but also accumulates reusable, high-value user group insights and content planning experience for the platform through continuous learning.

[0112] This embodiment also provides an optimization device for recommending objects, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0113] This embodiment provides an optimization device for recommending objects, such as... Figure 4 As shown, it includes: The first acquisition module 401 is used to acquire the topic features of the recommendation set, which is used to represent multimodal recommendation information; The second acquisition module 402 is used to acquire the matching relationship between topic features and the interest profile of the first recommended object, wherein the interest profile is determined based on the first interaction data of the first recommended object; The filtering module 403 is used to filter the first recommended object based on the matching relationship to obtain the second recommended object; The recommendation module 404 is used to recommend the recommendation set to the second recommendation object and obtain the second interaction data of the second recommendation object for the recommendation set; The optimization module 405 is used to optimize the second recommended object using the second interaction data.

[0114] In some alternative implementations, the optimization module 405 includes: The first processing submodule is used to process the second interactive data and determine the evaluation metrics after the recommended collection is recommended. The optimization submodule is used to optimize the second recommended object based on the evaluation metrics.

[0115] In some optional implementations, the recommendation set includes at least one recommendation; the processing submodule includes: The first determining unit is used to determine the target consumption completion rate of each recommendation item based on the content consumption records in the second interaction data. The second determining unit is used to determine the target weighting coefficient corresponding to each target consumption completion degree based on the mapping relationship between consumption completion degree and weighting coefficient; The third determining unit is used to determine the consumption duration weighted score corresponding to the recommendation set based on the superposition result of the weighted coefficients of each target. The fourth determining unit is used to determine the preference expression rate corresponding to the recommendation set based on the number of preference expression operations in the second interaction data, and to determine the attention conversion rate corresponding to the recommendation set based on the number of attention operations in the second interaction data. The fifth determination unit is used to determine the evaluation indicators based on the consumption duration weighted score, preference expression rate, and attention conversion rate.

[0116] In some alternative implementations, the optimization apparatus for the recommended objects further includes: The parsing module is used to parse the second interaction data and determine the sequence of operation behaviors corresponding to each second recommendation object. The statistics module is used to count the frequency of occurrence of each operation sequence; The determination module is used to identify the sequence of operational behaviors that occurs more frequently than a preset frequency threshold and whose evaluation indicators exceed a preset indicator threshold as the target behavior sequence.

[0117] In some alternative implementations, the optimization apparatus for the recommended objects further includes: The generation module is used to analyze the evaluation metrics and target behavior sequences to generate optimization suggestions for the recommendation set; The modification module is used to respond to the adjustment operations performed on the recommendation set based on optimization suggestions, and to obtain the modified recommendation set.

[0118] In some optional implementations, the second acquisition module 402 includes: The extraction submodule is used to extract behavioral feature data from the first interaction data under the dimensions of content consumption, interactive operation, and content theme preference. The second processing submodule is used to process the behavioral feature data of content consumption dimension to obtain the content preference values ​​of the first recommendation object for each content type; The third processing submodule is used to process the behavioral feature data of the interactive operation dimension to obtain the interaction intensity value of the first recommended object for multiple interaction types. The fourth processing submodule is used to process the behavioral feature data of the content topic preference dimension to obtain the topic preference degree of the first recommendation object for each content topic; The construction submodule is used to build an interest profile of the first recommended object based on content preference value, interaction behavior intensity value and topic preference degree, and to establish a matching relationship between topic features and interest profile.

[0119] In some alternative implementations, the filtering module 403 includes: The matching submodule is used to match the target content topics represented by the topic features with the topic preference in the interest profile to obtain the matching results; The judgment submodule is used to determine whether the intensity value of the interaction behavior corresponding to the first recommended object exceeds the preset intensity threshold and whether the content preference value of any content type exceeds the preset preference threshold if the matching result indicates that the first recommended object is successfully matched, and to obtain the judgment result. The determination submodule is used to determine the first recommended object as the second recommended object if the judgment result indicates that the intensity value of the interactive behavior exceeds the preset intensity threshold, and the content preference value of any content type exceeds the preset preference threshold.

[0120] The optimization apparatus for the recommended objects provided in this application can execute the optimization method for the recommended objects provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0121] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0122] The following is a detailed reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 501, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 502 or a program loaded from memory 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device. The processor 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0123] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 508 including, for example, magnetic tapes, hard disks, etc.; and communication devices 509. Communication device 509 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0124] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 509, or installed from memory 508, or installed from ROM 502. When the computer program is executed by processor 501, it performs the functions defined in the optimized method of the recommended object of this application.

[0125] Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0126] This application also provides a computer-readable storage medium. The methods described above according to this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the optimized methods of the recommended objects shown in the above embodiments are implemented.

[0127] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0128] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for optimizing recommendation objects, characterized in that, The method includes: Obtain the topic features of the recommendation set, which is used to represent multimodal recommendation information; Obtain the matching relationship between the topic features and the interest profile of the first recommended object, wherein the interest profile is determined based on the first interaction data of the first recommended object; Based on the matching relationship, the first recommended object is filtered to obtain the second recommended object; The recommended collection is recommended to the second recommended object, and the second interaction data of the second recommended object for the recommended collection is obtained; The second recommended object is optimized using the second interaction data.

2. The method according to claim 1, characterized in that, The optimization of the second recommended object using the second interaction data includes: The second interactive data is processed to determine the evaluation metrics after the recommended collection is recommended; The second recommended object is optimized based on the evaluation indicators.

3. The method according to claim 2, characterized in that, The recommendation set includes at least one recommendation item; The step of processing the second interaction data to determine the evaluation metrics for the recommended collection after it has been recommended includes: Based on the content consumption records in the second interaction data, the target consumption completion rate of each of the recommended items is determined; Based on the mapping relationship between consumption completion rate and weighting coefficient, the target weighting coefficient corresponding to each target consumption completion rate is determined; Based on the summation of the weighted coefficients of each target, the consumption duration weighted score corresponding to the recommendation set is determined; The preference expression rate corresponding to the recommendation set is determined based on the number of preference expression operations in the second interaction data, and the attention conversion rate corresponding to the recommendation set is determined based on the number of attention operations in the second interaction data. The evaluation metrics are determined based on the consumption duration weighted score, the preference expression rate, and the attention conversion rate.

4. The method according to claim 1, characterized in that, The method further includes: Analyze the second interaction data to determine the sequence of operation behaviors corresponding to each of the second recommended objects; Count the frequency of occurrence of each of the aforementioned operation sequences; The sequence of operational behaviors that occurs more frequently than a preset frequency threshold and whose evaluation indicators exceed a preset indicator threshold is identified as the target behavior sequence.

5. The method according to claim 4, characterized in that, The method further includes: The evaluation metrics and the target behavior sequence are analyzed to generate optimization suggestions for the recommended set. In response to the adjustment operation performed on the recommended set based on the optimization suggestions, a modified recommended set is obtained.

6. The method according to claim 1, characterized in that, The step of obtaining the matching relationship between the topic features and the interest profile of the first recommended object includes: From the first interaction data, extract behavioral feature data under the dimensions of content consumption, interactive operation, and content theme preference; The behavioral feature data of the content consumption dimension are processed to obtain the content preference values ​​of the first recommendation object for each content type; The behavioral feature data of the interactive operation dimension are processed to obtain the interaction intensity values ​​of the first recommended object for multiple interaction types; The behavioral feature data of the content topic preference dimension are processed to obtain the topic preference degree of the first recommendation object for each content topic. Based on the content preference value, the interaction behavior intensity value, and the topic preference degree, an interest profile of the first recommended object is constructed, and a matching relationship between the topic features and the interest profile is established.

7. The method according to claim 6, characterized in that, The step of filtering the first recommended object based on the matching relationship to obtain the second recommended object includes: The target content theme represented by the theme feature is matched with the theme preference degree in the interest profile to obtain the matching result; If the matching result indicates that the first recommended object is successfully matched, then it is determined whether the intensity value of the interaction behavior corresponding to the first recommended object exceeds a preset intensity threshold, and whether the content preference value of any content type exceeds a preset preference threshold, and a judgment result is obtained. If the judgment result indicates that the intensity value of the interactive behavior exceeds the preset intensity threshold, and the content preference value of any of the content types exceeds the preset preference threshold, then the first recommended object is determined as the second recommended object.

8. An optimization device for recommending objects, characterized in that, The device includes: The first acquisition module is used to acquire the topic features of the recommendation set, which is used to represent multimodal recommendation information; The second acquisition module is used to acquire the matching relationship between the topic features and the interest profile of the first recommended object, wherein the interest profile is determined based on the first interaction data of the first recommended object; A filtering module is used to filter the first recommended object based on the matching relationship to obtain a second recommended object; The recommendation module is used to recommend the recommendation set to the second recommendation object and obtain the second interaction data of the second recommendation object for the recommendation set; An optimization module is used to optimize the second recommended object using the second interaction data.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the optimization method for the recommended object according to any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes computer instructions for causing a computer to perform the optimization method for the recommended object according to any one of claims 1 to 7.

Images