Multi-target content recommendation method and device, computer equipment and storage medium

Abstract

The embodiment of the invention discloses a multi-target content recommendation method and device, computer equipment and a storage medium, and relates to the field of artificial intelligence. The method comprises the following steps: collecting and preprocessing user, content multi-mode and scene data; extracting and utilizing a cross-modal attention mechanism to fuse the preprocessed user, content multi-modal and scene data into unified feature representation; dynamically determining multi-target weights such as click rate, retention rate and diversity according to real-time feedback based on a reinforcement learning model; three recommendation subsets are generated based on the user-content interaction relation, unified feature matching and diversity screening, and a final recommendation list is generated according to dynamic weight fusion. According to the method, multi-target adaptive collaborative optimization and multi-dimensional feature deep fusion are realized, and the recommendation accuracy and the user experience are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-objective content recommendation method, apparatus, computer device, and storage medium. Background Technology

[0002] With the explosive growth of internet content, recommendation systems have become a core tool for efficiently connecting users with massive amounts of content. However, existing content recommendation algorithms still have many limitations in practice. Most algorithms focus on optimizing single business objectives such as click-through rate or conversion rate. This single-objective optimization model easily leads to homogenized recommendation results, forming "information cocoons." In the long run, this will limit user interests and thus cause a decline in long-term user retention.

[0003] To alleviate this problem, some research has turned to multi-objective optimization, attempting to simultaneously consider multiple metrics such as click-through rate, user retention, and content diversity. However, in these multi-objective recommendation systems, the weights of each objective are usually pre-set based on human experience or statically configured, making it difficult to flexibly adjust according to real-time user interactions and differentiated application scenarios, such as the cold start phase for new users or specific promotional activities, resulting in insufficient overall system adaptability. Furthermore, whether single-objective or multi-objective algorithms, their mining of content and user features is often insufficient. Multimodal information such as text, video, and tags on the content side, as well as behavioral sequences and deep interest preferences on the user side, are not effectively integrated and utilized. The singularity of feature dimensions makes it difficult for the algorithm to truly support collaborative optimization of multiple objectives. On the other hand, the optimization iteration of existing solutions heavily relies on offline historical data for evaluation, lacking a feedback loop that links with real-time online user behavior, causing the algorithm update speed to lag behind the dynamic changes in user needs and failing to respond quickly to the current situation.

[0004] How to build a recommendation system that can effectively coordinate multiple business objectives, dynamically adapt to changes in scenarios and users, make full use of multi-dimensional information, and respond to feedback in a timely manner, so as to improve the accuracy of recommendations and user experience, has become an urgent technical problem to be solved. Summary of the Invention

[0005] This invention provides a multi-objective content recommendation method, apparatus, computer device, and storage medium. The technical problem it aims to solve is: how to build a recommendation system that can effectively coordinate multiple business objectives, dynamically adapt to changes in scenarios and users, fully utilize multi-dimensional information, and respond to feedback in a timely manner, so as to improve the accuracy of recommendations and user experience.

[0006] In a first aspect, embodiments of the present invention provide a multi-objective content recommendation method, comprising:

[0007] Collect and preprocess user data, multimodal content data, and scenario data;

[0008] Feature extraction is performed on the preprocessed user data, content multimodal data, and scene data to obtain user feature vectors, content feature vectors, and scene feature vectors. The user feature vectors, content feature vectors, and scene feature vectors are then fused into a unified feature representation through a cross-modal attention mechanism.

[0009] Based on the current business scenario and real-time feedback data, the weights of multiple recommendation objectives are dynamically determined through a reinforcement learning model. These multiple recommendation objectives include at least click-through rate, user retention rate, and content diversity.

[0010] A first recommendation subset is generated based on the interaction between users and content. A second recommendation subset is generated based on the correlation between content and user interests calculated using the unified feature representation. A third recommendation subset is generated by filtering candidate content whose similarity to the selected content is lower than a threshold. The first, second, and third recommendation subsets are then merged according to the dynamically determined weights to generate the final recommendation list.

[0011] Secondly, embodiments of the present invention also provide a multi-target content recommendation apparatus, which includes a unit for performing the above-described method.

[0012] Thirdly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0014] This invention provides a multi-objective content recommendation method, apparatus, computer device, and storage medium. The method includes: collecting and preprocessing user data, multimodal content data, and scene data; extracting features from the preprocessed user data, multimodal content data, and scene data to obtain user feature vectors, content feature vectors, and scene feature vectors; fusing the user feature vectors, content feature vectors, and scene feature vectors into a unified feature representation using a cross-modal attention mechanism; dynamically determining the weights of multiple recommendation objectives based on the current business scenario and real-time feedback data using a reinforcement learning model, wherein the multiple recommendation objectives include at least click-through rate, user retention rate, and content diversity; generating a first recommendation subset based on the interaction relationship between users and content; calculating the relevance between content and user interests based on the unified feature representation to generate a second recommendation subset; and selecting content from candidate content with a similarity lower than a threshold to generate a third recommendation subset; and fusing the first, second, and third recommendation subsets according to the dynamically determined weights to generate a final recommendation list. By introducing a reinforcement learning model to dynamically determine multi-objective weights, this method achieves collaborative optimization and adaptive balance of objectives such as click-through rate, retention rate, and diversity, overcoming the limitations of static weight configuration. Through a cross-modal attention mechanism, it deeply integrates multi-dimensional features of users, content, and scenarios, providing a unified and rich representational foundation for recommendation decisions. This method mechanistically constructs a recommendation system capable of collaboratively addressing multiple objectives, dynamically adapting to changes, and fully utilizing multi-dimensional information, thereby improving the overall accuracy of recommendations and user experience. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the multi-target content recommendation method provided in this embodiment of the invention;

[0017] Figure 2 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

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

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0023] Please see Figure 1 This invention provides a multi-target content recommendation method, which includes the following steps:

[0024] S1 collects and preprocesses user data, multimodal content data, and scene data.

[0025] In practice, raw user data, multimodal content data, and scenario data are collected from the business backend and client. User data includes user identifiers, attributes, and historical behavior sequences such as clicks, dwell times, and favorites sorted by timestamps. Multimodal content data covers text descriptions, tags, image frames, audio streams, and video files. Scenario data includes request time, user device type, operation, and network conditions. These raw data undergo preprocessing operations such as cleaning, deduplication, missing value imputation, and standardization to prepare a standardized data source for subsequent feature engineering.

[0026] S2, feature extraction is performed on the preprocessed user data, content multimodal data and scene data to obtain user feature vectors, content feature vectors and scene feature vectors, and the user feature vectors, content feature vectors and scene feature vectors are fused into a unified feature representation through a cross-modal attention mechanism.

[0027] In practice, the preprocessed user data is processed using sequence models such as Transformer to generate user feature vectors representing dynamic interests. For content data, convolutional neural networks are used to extract text and image features, while recurrent neural networks are used to extract temporal features from audio and video. These features are then linked to external knowledge graphs to enrich semantics, resulting in a comprehensive content feature vector. For scene data, encoding techniques are used to transform information such as time and device into scene feature vectors. The crucial step then involves using a cross-modal attention mechanism to fuse these heterogeneous user, content, and scene feature vectors. This mechanism dynamically focuses on the feature combination most relevant to the current recommendation task by calculating attention weights between different feature modalities, ultimately outputting a unified, deep-level joint feature representation that provides an information foundation for subsequent multi-objective decision-making.

[0028] In some preferred embodiments, feature extraction of the user data to obtain a user feature vector includes: arranging the user's historical behavior sequence in chronological order, the historical behavior sequence including the user's clicks, pauses, and favorites of content; inputting the historical behavior sequence into a Transformer model, calculating the dependencies between various behaviors in the sequence through the multi-head self-attention mechanism in the Transformer model; and extracting the mean or the vector of the last position of the sequence representation vector output by the Transformer model as the user feature vector.

[0029] In practice, the first step is to extract a strictly timestamped sequence of user historical behaviors from the preprocessed user data. Each item in this sequence is a behavior unit, recording the user's specific action on a particular content item, such as clicking, staying for more than a certain duration, saving, or sharing. Each behavior unit typically includes information such as content identifier, action type, and timestamp.

[0030] Furthermore, the user's historical behavior sequence is fed into a pre-built Transformer model for processing. The core component of the Transformer model is a multi-head self-attention mechanism. In this mechanism, each behavioral unit in the sequence is transformed into an embedding vector containing information such as content and operation type. The self-attention layer calculates the association weight between any two behavioral units in the sequence, enabling the model to identify behaviors that are semantically or causally closely related, even if they are geographically distant. For example, a user might save an article about a travel destination on Monday and click on a flight search page for that destination on Friday; the self-attention mechanism can capture this cross-temporal interest association. Through this mechanism, the model not only focuses on individual behaviors but, more importantly, models the complex dependencies and long-term patterns between behaviors.

[0031] Furthermore, after multiple layers of encoding by the Transformer model, the entire input sequence is transformed into a sequence of sequence representation vectors. To obtain a fixed-length vector representing the overall user interests, pooling is employed. Common pooling strategies include calculating the mean of the entire sequence representation vector sequence or directly extracting the output vector from the last position of the sequence (corresponding to the most recent behavior), i.e., the user feature vector. This user feature vector encapsulates the temporal patterns and interest preferences in the user's historical behavior sequence, representing a dynamic and context-sensitive user representation.

[0032] By using the Transformer model to process user historical behavior sequences to generate user feature vectors, significant superiority over traditional user modeling methods is achieved. Traditional methods, such as statistical labeling or simple recurrent neural networks, often struggle to effectively capture complex, non-local behavioral dependencies in long sequences. The multi-head self-attention mechanism in the Transformer model, however, has the advantage of parallel computation of the pairwise correlation strength between all elements in the sequence. This characteristic allows the model to directly model the leaping correlations and periodic patterns present in long-term user behavior.

[0033] For example, a user might have developed an interest in a technology topic weeks ago and read multiple articles, then their interest might have temporarily subsided, only to be rekindled upon encountering related news. Self-attention mechanisms can bridge this intermediate time interval, linking early deep reading with recent triggering behaviors, thus more accurately determining the strength of a user's sustained interest in a topic, rather than relying solely on recent behavior. This ability to capture long-term dependencies allows the generated user feature vectors to more comprehensively and profoundly reflect the evolution of user interests and stable preferences, avoiding the problems of one-sided or rapidly shifting interest representations caused by focusing only on short-term behaviors.

[0034] Therefore, the user feature vector generated in this embodiment is of higher quality and richer in information, providing a more solid and accurate user-side information foundation for subsequent cross-modal fusion and multi-objective recommendation decisions. This directly improves the accuracy of interest matching in subsequent recommendation steps, especially for scenarios that require understanding users' long-term interests to optimize retention rates and other objectives, contributing a key technical improvement.

[0035] In some preferred embodiments, feature extraction of the content multimodal data to obtain a content feature vector includes: for text data, extracting local semantics through a convolutional neural network and / or obtaining contextual semantic embeddings through a pre-trained language model to obtain text features; for image or video frame data, extracting visual features through a convolutional neural network; for audio or time-series video data, extracting time-series variation features through a recurrent neural network; concatenating or weighting the extracted text features, visual features, and time-series variation features, and associating them with corresponding entity information in a preset knowledge graph to generate the content feature vector.

[0036] In practice, for text data such as titles, descriptions, or comments, two deep learning methods are employed in parallel or selectively. One method uses convolutional neural networks to perform convolution operations on the segmented text to extract key phrases and local semantic patterns. The other method utilizes a pre-trained language model to obtain the contextual semantic embedding after inputting the entire text; this embedding captures the deeper meaning of words in a specific context.

[0037] Furthermore, for image data or keyframes in the content, convolutional neural networks are used for visual feature extraction. These networks are typically pre-trained on large-scale image datasets and are capable of extracting multi-level visual features, from edges and textures to high-level semantic objects. For audio data or video data exhibiting temporal changes, recurrent neural networks or their variants are used. These networks excel at modeling temporal dependencies and can extract spectral variation patterns in audio or evolutionary features of actions and scenes in video.

[0038] Furthermore, textual features, visual features, and temporal variation features extracted from different modalities of the same content item are integrated. The integration method can be simple vector concatenation, or weighted fusion based on prior knowledge or learned weights, thereby forming a preliminary multimodal content representation vector.

[0039] Furthermore, the aforementioned multimodal features are associated with a pre-defined knowledge graph. By identifying entities or topics in the content, they are linked to corresponding nodes in the knowledge graph, and the graph embedding vector of that node and its surrounding relationships is obtained. This graph embedding vector is then fused with the aforementioned multimodal feature vector to generate the final content feature vector. This step injects structured external knowledge into the content, for example, clarifying that "Interstellar" is a film directed by a specific director and belonging to the "hard science fiction" subgenre.

[0040] This embodiment significantly improves the representational power and semantic richness of content feature vectors by implementing multimodal deep feature extraction and combining it with external knowledge graphs, thereby producing clear technical effects. Traditional content feature engineering often relies on manually defined labels or shallow statistical features, which are difficult to fully mine the complex information in unstructured data, and the information of each modality is isolated. This solution first uses deep models optimized for specific modalities, such as convolutional neural networks, pre-trained language models, and recurrent neural networks, to automatically extract high-dimensional and deep features from text, visual, and temporal dimensions, respectively. This overcomes the limitations of manual features and ensures a more comprehensive and essential portrayal of the content itself.

[0041] Furthermore, fusing features from different modalities and linking them to a knowledge graph in the final stage achieves information complementarity and enhancement at the technical level. Multimodal fusion allows content features to simultaneously incorporate semantic, visual, and auditory cues. For example, a video about "cats" might have a feature vector containing the textual semantics of "pet" and "animal," the visual features of a cat, and possibly the audio pattern of a meow. Linking to a knowledge graph introduces a vast amount of structured common sense and domain knowledge, enabling the explicit identification of implicit connections between content, such as linking "Newton" with "physics" and "the law of universal gravitation." The feature vectors generated by this deep fusion not only represent the superficial attributes of the content but also contain its deeper semantic context and position within the world of knowledge.

[0042] Therefore, using such rich feature vectors for subsequent similarity calculations and recommendation matching enables more accurate and interpretable content understanding and alignment with user interests. It allows for the identification of deeper connections beyond surface keyword matching; for example, recommending a popular science article explaining "relativity" to a user who recently watched the movie "Interstellar" because the knowledge graph reveals a deep connection between the two concepts in "theoretical physics" and "spacetime." This directly improves the quality of content-based recommendations, which is crucial for optimizing user retention and content diversity.

[0043] In some preferred embodiments, feature extraction is performed on the scene data to obtain a scene feature vector, including: classifying the scene data into time-dimensional data, device-dimensional data, and network environment-dimensional data; for time-dimensional data, extracting the hour information, weekday information, and whether it is a holiday; mapping the hour information to sine and cosine codes with a 24-hour period, performing one-hot encoding on the weekday information, and converting the holiday information into Boolean encoding; for device-dimensional data, extracting the device type, screen size, and operating system version; performing one-hot encoding on the device type and operating system version, and dividing the screen size into predefined size ranges and performing range encoding; for network environment-dimensional data, extracting the network type and bandwidth level, and performing one-hot encoding; concatenating the above time-dimensional encoding, device-dimensional encoding, and network environment-dimensional encoding to form an initial scene encoding vector; inputting the initial scene encoding vector into a fully connected neural network for nonlinear transformation and dimensionality reduction, and outputting the scene feature vector.

[0044] In practice, the first step is to divide the collected raw scene data into three dimensions: time dimension data, device dimension data, and network environment dimension data.

[0045] Furthermore, fine-grained feature extraction and encoding are performed for each dimension. For the time dimension data, three key features are extracted: the current hour information, the day of the week information, and whether it is a holiday or public holiday information based on the calendar. The hour information uses sine and cosine encoding, that is, calculating sin(2π*hour / 24) and cos(2π*hour / 24) to map the cyclic time into a continuous vector to preserve the proximity between "23:00" and "01:00". The day of the week information uses one-hot encoding, mapping Monday to Sunday to seven binary bits. Holiday information is directly converted into a Boolean value (0 or 1). For the device dimension data, device type, screen size, and operating system version are extracted. Discrete categorical variables such as device type and operating system version use one-hot encoding. Screen size, as a continuous variable, is first divided into several predefined size ranges based on business experience, and then range encoding is performed. For the network environment dimension data, network type and bandwidth level are extracted and processed using one-hot encoding as well.

[0046] Furthermore, all the encoded feature vectors are concatenated in dimensional order to form a high-dimensional initial scene encoding vector that is a mixture of sparse and dense data. This vector contains the original discrete and continuous information of the scene.

[0047] Furthermore, this initial scene encoding vector is input into a fully connected neural network. This network typically consists of multiple hidden layers and includes non-linear activation functions. Its role is to perform non-linear transformations and feature dimensionality reduction. The network can learn the interaction relationships between features of different dimensions; for example, "weekend evening" and "using a tablet" may both point to a longer period of entertainment. Simultaneously, it compresses the high-dimensional, sparse initial encoding into a low-dimensional, dense scene feature vector. This vector is a distributed, abstract representation of the initial features, making it more suitable for subsequent model computation.

[0048] This embodiment effectively transforms the original scene description into a high-quality scene feature vector through structured dimensional partitioning, targeted feature engineering, and deep neural network transformation, resulting in a clear technical effect. Traditional methods may simply treat scene information as an independent feature or perform coarse encoding, failing to fully capture its complexity and non-linear impact on recommendation decisions. The technical effect of this solution is primarily reflected in the rationality and completeness of feature representation. By dividing the scene into three core dimensions—time, device, and network—and extracting the most discriminative features for each, comprehensive coverage of scene information is ensured. In particular, the use of cyclic encoding for time scientifically handles the periodicity of time, enabling the model to understand the differences in user intent at different times.

[0049] Furthermore, using a fully connected neural network to perform nonlinear transformations and dimensionality reduction on the concatenated initial encoding is key to achieving deep learning effects. While the initial encoded vector is comprehensive, its high dimensionality and potential feature sparsity make it inefficient for directly applying to complex models and hinder its ability to learn complex patterns. Fully connected networks, through their multi-layered nonlinear transformation capabilities, can automatically learn the complex interactions and combined meanings between different scene factors. For example, it can learn the combined features of "weekday lunch break" and "4G mobile network," which may correspond to a user's desire for short-term, fast-paced content consumption; while the combination of "weekend evening" and "home Wi-Fi large-screen TV" may correspond to a need for in-depth, long-form video entertainment. The low-dimensional, dense vectors output by the network are precisely the abstract and distributed representations of these complex scene patterns.

[0050] Therefore, the scene feature vector generated in this embodiment is no longer a simple set of isolated features, but a powerful feature that contains scene contextual semantics and is easy for the model to process. When this vector is input into subsequent cross-modal attention mechanisms or reinforcement learning models, it can provide profound contextual information about "the environment in which recommendations are made," thereby significantly enhancing the scene adaptability of recommendations. This enables recommendations to dynamically adjust recommendation strategies based on the user's specific spatiotemporal and device environment, such as recommending short videos during commuting breaks and recommending high-definition movies in a large-screen home environment, thereby improving the overall user experience and the achievement of multiple business objectives.

[0051] In some preferred embodiments, the method further includes: when the target is identified as a new user, obtaining the attribute information provided by the new user during registration; when the target is identified as new content, obtaining the metadata tag of the new content; matching the attribute information or the metadata tag with a pre-built knowledge graph to obtain the embedding vector of the corresponding entity in the knowledge graph; for new users, inputting their behavior sequence into a Transformer model to extract an initial behavior feature vector, and concatenating the initial behavior feature vector with the embedding vector obtained from the knowledge graph to form an enhanced user feature vector; for new content, using a convolutional neural network or a recurrent neural network to extract its multimodal original feature vector, and concatenating the original feature vector with the embedding vector obtained from the knowledge graph to form an enhanced content feature vector.

[0052] When a new user is identified as the target of a recommendation request, a processing flow for that new user is triggered. First, the basic attribute information provided by the new user during registration is retrieved. This information typically includes, but is not limited to, structured fields such as age, gender, region, and occupation. Simultaneously, when a newly published content item is identified as the target for recommendation, a processing flow for that new content is triggered, retrieving the metadata tags for that new content, such as author, category, and manually added topic tags.

[0053] Furthermore, the acquired attribute information or metadata tags are matched against a pre-built large-scale knowledge graph. The knowledge graph contains entities (such as "college student," "Beijing," and "science fiction movie") and the relationships between them. The matching process aims to find the knowledge graph entity node corresponding to the input information. Upon successful matching, the pre-learned embedding vector of that entity node in the knowledge graph is read. This embedding vector, learned through graph embedding techniques, represents the entity's structural and semantic information within the graph.

[0054] Furthermore, feature enhancement is performed. For new users: although the user lacks a historical behavior sequence, a lightweight or general-purpose Transformer model can still be invoked to attempt to process the very few initial behaviors the user might perform (such as the first click), extracting an initial behavioral feature vector, which may contain very little information. The key step is to concatenate this weak initial behavioral feature vector with an embedding vector obtained from the knowledge graph based on its attribute information. The concatenated vector forms an enhanced user feature vector, which contains not only a hint of behavioral clues but also prior knowledge from the knowledge graph representing the user's group or attribute preferences.

[0055] Furthermore, for new content: similarly, models such as convolutional neural networks or recurrent neural networks are used to process its original multimodal data, including text and images, to extract an original feature vector. Simultaneously, the embedding vector obtained by matching its metadata tags with a knowledge graph is concatenated with this original feature vector to form an enhanced content feature vector. This injects rich category and semantic prior information into new content that currently lacks popularity.

[0056] The cold start processing scheme defined in this embodiment creatively solves the long-standing problem of cold start for new users and new content in recommendation by combining the embedding of prior knowledge from the knowledge graph with feature extraction from the basic model, and produces significant technical effects. Traditional methods, when faced with the cold start problem, often suffer from feature sparsity and model failure due to a lack of interaction data, and can only perform random recommendations or coarse-grained matching based on extremely limited metadata, resulting in poor performance.

[0057] The technical effectiveness of this solution stems primarily from the introduction of knowledge graphs. As a structured external knowledge source, knowledge graphs contain rich semantic relationships and common-sense logic between entities in their embedding vectors. For new users, their attribute information is mapped to knowledge graph entity embeddings, essentially giving them a "group profile." For example, a new user whose registration information is "university student" and "located in Beijing" incorporates prior distributions of interests related to "student group" and "Beijing area" into their initial features through knowledge graph embedding. For new content, the embeddings obtained from tag mapping inject deep semantic features of its category and topic. This provides a solid information supplement for cold-start objects, filling the absolute gap in behavioral data.

[0058] Furthermore, the key technological advantage lies in the fact that the simple operation of "feature concatenation" organically combines prior knowledge with individual characteristics. For new users, the concatenated feature vector is no longer an invalid vector close to zero, but a vector with sufficient information and clear direction, enabling subsequent steps such as cross-modal fusion and similarity calculation to be performed effectively. This allows for relatively reasonable and somewhat personalized recommendations to be made when a new user uses the app for the first time, greatly improving the initial experience. For new content, the enhanced feature vector allows it to be better understood and indexed by the content matching model, thus having the opportunity to be recommended to users who may be interested in it, breaking the vicious cycle of "no recommendation without historical interaction".

[0059] Therefore, this embodiment effectively alleviates the initial performance bottleneck of recommendations caused by data sparsity, accelerates the capture of new users' interests and the discovery of the value of new content, and improves the coverage of recommendations and the overall health of the ecosystem.

[0060] S3. Based on the current business scenario and real-time feedback data, the weights of multiple recommendation objectives are dynamically determined through a reinforcement learning model. The multiple recommendation objectives include at least click-through rate, user retention rate, and content diversity.

[0061] In practice, based on the unified feature representation and real-time feedback loop described above, a reinforcement learning model is invoked to dynamically determine the weights of multiple recommendation objectives. This model takes the unified feature representation and historical states as input, and aims to optimize long-term user value (e.g., rewards synthesized from metrics such as click-through rate and retention rate). It learns online using algorithms such as temporal difference and outputs a set of dynamic weight values, such as click-through rate weight, user retention rate weight, and content diversity weight. These weights are not fixed but adaptively adjust based on real-time user interactions and changes in the context.

[0062] In some preferred embodiments, the step of dynamically determining the weights of multiple recommendation targets using a reinforcement learning model includes: using at least one of the user feature vector, content feature vector, and scene feature vector, along with corresponding historical weight values, as the state input of the reinforcement learning model; defining a reward function aimed at comprehensively measuring long-term user value, wherein the reward function is obtained by weighted summation of real-time metrics of the multiple recommendation targets; using a temporal difference algorithm, updating the Q-value function in the reinforcement learning model based on the current state, the action taken, and the immediate reward obtained in the next state, and outputting an adjustment action for the weights of the multiple recommendation targets based on the updated Q-value function, wherein the adjustment action is executed within a preset weight range.

[0063] In practice, the first step is to define the state of the reinforcement learning model. This state is a vector that contains at least one or more combinations of the current user feature vector, content feature vector, and scene feature vector to describe the current "environment." Furthermore, to reflect the continuity of decision-making, the state typically also includes the weight values ​​previously determined for each recommendation objective as historical state information.

[0064] Furthermore, a reward function aligned with long-term business objectives is defined. This function serves as the target for optimization of the reinforcement learning model. In this scheme, the reward function is designed to be obtained by a weighted sum of real-time metrics for multiple recommendation objectives.

[0065] Furthermore, the model utilizes a temporal difference algorithm for learning and decision-making. In each decision step, the model outputs an "action" At based on its internal policy function, according to the current state St. This action is the adjustment vector for the weights of each recommendation target. This action is applied to the current weights, ensuring the adjusted weights are within a preset reasonable range, for example, the click-through rate weight range is [0.2, 0.6]. After making recommendations, real-time user interaction data is collected, and an immediate reward Rt+1 is calculated accordingly. Simultaneously, the environment transitions to a new state St+1. The core of the temporal difference algorithm is updating the model's value function, typically the Q-value function. The Q-value is updated by calculating the difference between the predicted reward of the current state-action pair and the actual observed immediate reward plus the estimated reward for the next state, making the prediction more accurate. Based on the updated Q-value function, the model can optimize its policy, thereby making weight decisions that yield higher cumulative rewards in similar future states.

[0066] By employing a reinforcement learning model based on temporal difference algorithms to dynamically determine weights, a fundamental shift from static or rule-driven weight configuration to data-driven, adaptive weight decision-making is achieved, resulting in significant technical improvements. Traditional methods rely on pre-setting fixed weights based on human experience or simple heuristic rules, which cannot cope with dynamically changing user interests, different needs in various scenarios, and complex trade-offs between multiple objectives. The reinforcement learning framework constructed in this solution formalizes the weight decision-making process as a sequential decision problem, and its core technical effectiveness lies in "online learning" and "long-term optimization."

[0067] First, the model receives reward feedback through interaction with the environment, enabling it to learn weight adjustment strategies online and continuously. This means that weights are not static but are fine-tuned based on feedback from each user interaction and each recommendation result. For example, when the model discovers that increasing diversity weights slightly reduces the single click-through rate but significantly improves subsequent user dwell time and return rates, resulting in higher cumulative rewards, the model will gradually learn to favor increasing diversity weights in similar scenarios. This online optimization capability based on direct feedback makes weight allocation highly scenario-sensitive and user-personalized adaptable.

[0068] Furthermore, the model aims to maximize cumulative rewards (i.e., long-term user value) rather than simply optimizing immediate metrics. This is a core advantage of reinforcement learning algorithms such as temporal difference. It allows the model to consider the potential impact of current actions on the future when making weight decisions, thus engaging in a degree of "long-term planning." For example, when user interest reaches saturation, the model may proactively sacrifice some short-term click-through rates (by adjusting weights to reduce the priority of related content) to explore new points of interest (increasing diversity weights), aiming to obtain greater long-term rewards from longer-term user activity and retention. This optimization orientation towards long-term value is difficult to achieve with static weighting methods.

[0069] Therefore, this embodiment enables multi-objective recommendation to autonomously and intelligently balance different objectives, find the optimal trade-off point that is suitable for specific users and the current scenario, thereby improving the overall adaptability of recommendation and long-term business benefits.

[0070] In some preferred embodiments, the method further includes: pre-configuring weight adjustment rules associated with specific business scenarios, such as new user cold start or promotional activities; when a match is detected between the current business scenario and the specific business scenario, a one-time basic adjustment is made to the weights of the multiple recommendation targets according to the weight adjustment rules; the weights after the basic adjustment are used as the initial state of the reinforcement learning model, and then the reinforcement learning model is fine-tuned online according to subsequent real-time feedback data.

[0071] In practice, the step of pre-adjusting weights based on business rules before reinforcement learning decisions forms a hybrid decision-making mechanism. Specifically, administrators or business experts first need to pre-configure a set of weight adjustment rules explicitly associated with specific business scenarios. These rules are deterministic, domain-knowledge-based logical judgments. For example, a rule might be defined as follows: when the scenario is identified as "new user cold start," the initial weight of "content diversity" is increased by 50%, and the weight of "click-through rate" is correspondingly decreased; when the scenario is identified as "promotional period," the weight of "conversion rate" related objectives is significantly increased.

[0072] Furthermore, during operation, whenever weights need to be determined, a scenario detection step is executed first. This analyzes current user characteristics, request context, and other information to determine if they match a predefined specific business scenario. Once a match is detected, such as identifying a newly registered user with scarce behavioral data, or detecting that the user is currently in a global promotional activity period, the corresponding rules are triggered.

[0073] Furthermore, upon triggering, the reinforcement learning model is not immediately invoked. Instead, a one-time, deterministic adjustment is made to the weights of the multiple recommendation targets based on the matched rules. This adjustment directly applies the calculations or assignments defined in the rules, quickly setting the weights to a reasonable starting point that aligns with the prior knowledge of the business in this scenario.

[0074] After the rules are adjusted, the adjusted weights, along with the current state features, are used as the initial input state for the reinforcement learning model. The reinforcement learning model then takes over, using this "warmed-up" or "guided" initial weight state, combined with continuously flowing real-time feedback data, to perform online, data-driven fine-tuning. The reinforcement learning model can make more refined and personalized adjustments based on this foundational weights to adapt to the individual differences of different users in this specific scenario.

[0075] This embodiment effectively addresses the risks of high exploration costs, unstable initial performance, or convergence to suboptimal solutions that pure reinforcement learning methods may face in specific key scenarios by introducing a hybrid mechanism that combines business rule-based weight pre-adjustment with online fine-tuning of reinforcement learning. This results in a synergistic and effective technical outcome.

[0076] Pure reinforcement learning models explore optimal strategies from scratch, which can require extensive trial and error when the state space is large. In special but critical business scenarios such as new user cold starts or global promotional campaigns, this initial random exploration may lead to poor user experience or missed important business opportunities. This solution first utilizes pre-configured weight adjustment rules to provide valuable prior knowledge guidance. These rules embody business experts' deep understanding of optimal strategies in specific scenarios, enabling the rapid initialization of weights to a relatively reasonable and safe baseline. For example, forcibly increasing diversity weights during cold starts can immediately alleviate the problem of narrow recommendations for new users due to insufficient data, ensuring a basic user experience.

[0077] Furthermore, fixed rules lack flexibility and personalization. The further technical advantage of this solution lies in using the adjusted weights as the starting point for the reinforcement learning model. This is equivalent to setting a "high-value starting point" for the reinforcement learning exploration process, significantly shortening the exploration path and time required to converge to an excellent strategy. Based on this, the reinforcement learning model can fully leverage its data-driven and adaptive learning advantages, finely and personally adjusting the weights according to each new user's specific initial feedback (even if minimal) or different users' varying reactions to promotional activities. This collaborative model of "rules-first guidance, learning-followed optimization" ensures both rapid startup and basic effectiveness in key scenarios while retaining the ability to adapt to complex individual differences and dynamic changes over the long term, achieving a balance between robustness and adaptability.

[0078] S4. Generate a first recommendation subset based on the interaction between users and content; generate a second recommendation subset based on the correlation between content and user interests calculated using the unified feature representation; and select content from candidate content whose similarity to the selected content is lower than a threshold to generate a third recommendation subset; and merge the first recommendation subset, the second recommendation subset, and the third recommendation subset according to the dynamically determined weights to generate a final recommendation list.

[0079] In practice, three core recommendation strategies are executed in parallel to generate candidate subsets. The first strategy, based on the historical interaction between users and content, such as through collaborative filtering or graph neural networks to mine group preferences, generates a first recommendation subset aimed at optimizing click-through rates. The second strategy, based on the aforementioned unified feature representation, directly calculates the correlation between candidate content features and current user features, generating a second recommendation subset aimed at improving long-term user engagement. The third strategy focuses on list diversity, actively filtering candidate items that differ significantly from the already selected content by calculating the similarity between content and setting a threshold, generating a third recommendation subset aimed at ensuring exploratory nature. Finally, based on the target weights dynamically determined by the reinforcement learning model, the content in the three subsets is weighted, fused, and re-ranked to generate the final recommendation list, which is then returned to the user.

[0080] Furthermore, user interaction behaviors (such as clicks, skips, and prolonged stays) on this recommendation list are collected in real time, and this behavioral data is transformed into real-time metrics such as click-through rate and dwell time, serving as new feedback data input. This feedback data is used to immediately update the reinforcement learning model's strategy, and also to update the parameters of the feature extraction model and recommendation strategy model through incremental learning techniques, thereby achieving a continuous closed-loop iteration of "recommendation-feedback-optimization".

[0081] In some preferred embodiments, generating the first recommendation subset based on the interaction between users and content includes: constructing a heterogeneous graph with users and content as nodes, aggregating information of neighboring nodes using a graph neural network, updating node representations, and selecting the Top-K content items with the highest similarity as the first recommendation subset based on the inner product or cosine similarity between the updated user node and content node representations.

[0082] In practice, to generate the first recommendation subset, a heterogeneous graph with users and content as nodes is first constructed. In this graph, if there are historical interactions between users and content, an edge is established between the corresponding user node and content node. A graph neural network is then used to process this graph: through a message passing mechanism, each node aggregates the feature information of its direct neighbors. For example, a content node aggregates the features of all users it has interacted with, and a user node aggregates the features of all content it has interacted with. After multiple rounds of iterative aggregation, the representation vectors of all user nodes and content nodes are updated. When generating recommendations, for the current target user, the inner product or cosine similarity between its updated node representation and the representations of all candidate content nodes is calculated. The K (K is a positive integer) content items with the highest similarity are selected to form the first recommendation subset. This method focuses on mining the implicit associations in the user-content interaction graph.

[0083] Furthermore, the step of generating a second recommendation subset based on the correlation between the content and the user's interests calculated by the unified feature representation includes: calculating the similarity between the candidate content feature vector and the current user feature vector in the feature space, and optimizing the feature space by using a contrastive learning loss function, so that the feature distance of positive sample pairs decreases and the feature distance of negative sample pairs increases, and finally selecting the Top-K content with the highest similarity to the user feature vector as the second recommendation subset.

[0084] In practice, for generating the second recommendation subset, the unified feature representation of the current user is used as vector U, and the unified feature representation of each candidate content is used as vector Ci. Within the same feature space, the similarity between vector U and each Ci is calculated. To optimize this feature space and make the similarity calculation more reflective of true interest relevance, a contrastive learning loss function is introduced during model training. The design goal of this loss function is to minimize the distance in the feature space for content pairs that the user has historically interacted with (positive samples), and maximize the distance for randomly sampled non-interacted content pairs (negative samples). Finally, during the inference phase, the K candidate contents with the highest similarity to the current user's feature vector U are selected to form the second recommendation subset. This method directly matches users and content based on deep features.

[0085] Furthermore, the step of filtering content from candidate content with a similarity lower than a threshold to the selected content to generate a third recommendation subset includes: calculating the cosine similarity between candidate content and content already selected in the recommendation list; if the similarity between candidate content and any content in the selected list is lower than a preset threshold, then the candidate content is added to the candidate pool, and content that maximizes the diversity score of the overall recommendation list is selected from the candidate pool according to the maximum marginal utility algorithm to form the third recommendation subset.

[0086] In practice, the generation of the third recommendation subset focuses on diversity control. First, a set of content initially selected for the recommendation list is maintained. For each candidate content, the cosine similarity between its feature vector and the feature vector of every content in the selected list is calculated. If the similarity of the candidate content to all content in the selected list is below a preset threshold, it is considered sufficiently different from the existing list and placed in a candidate pool. Then, the maximum marginal utility algorithm is applied: this algorithm selects content from the candidate pool that maximizes the overall diversity utility of the recommendation list (e.g., the breadth of category coverage or the dispersion of feature vectors), adding them one by one until the third recommendation subset is formed. The core of this method is to actively filter low-similarity content to ensure diversity.

[0087] By specifically defining the generation algorithms for three recommendation subsets, candidate content generation based on different technical principles and optimization objectives was achieved. This provides high-quality intermediate results with complementary technical approaches and clear optimization objectives for subsequent weighted fusion, thus forming a solid foundation for achieving multi-objective collaborative optimization.

[0088] The first recommended subset, based on a graph neural network generation method, effectively leverages collective intelligence and implicit feedback. By modeling and aggregating information from user-content interaction graphs, it can discover complex relationships between users and content, and is particularly adept at uncovering collaborative patterns such as "users you like also like" or "users with similar interests like." This has a natural advantage for goals that rely on broad behavioral patterns, such as capturing trends, addressing data sparsity, and optimizing click-through rates.

[0089] The second recommendation subset, based on a unified feature representation and contrastive learning-based generation method, achieves accurate personalized deep matching. The training objective of contrastive learning directly optimizes the feature space, making positive and negative sample pairs of users and content clearly distinguishable in the vector space. This ensures that the calculated similarity more realistically reflects the depth of users' individual interests and preferences, rather than merely behavioral co-occurrence. This method is crucial for understanding users' unique interests, improving the relevance of recommended content, and enhancing long-term user satisfaction.

[0090] The third recommended subset, generated using a similarity threshold and maximum marginal utility algorithm, provides a proactive and controllable diversity guarantee mechanism. Instead of passively filtering from other results, it actively selects from a low-similarity candidate pool, aiming to optimize the overall diversity utility of the list. This ensures that regardless of the results of the first two subsets, there will ultimately be a dedicated source of content focused on breaking down information silos and increasing exploratory engagement.

[0091] These three methods generate candidate content from different perspectives, and their respective technical advantages are integrated in the subsequent weight fusion step, so that the final recommendation list can benefit from group behavior mining, individual deep interest matching and proactive diversity exploration at the same time. This is the key technical guarantee for achieving multi-objective balanced optimization.

[0092] In some preferred embodiments, the step of merging the first, second, and third recommendation subsets based on dynamically determined weights to generate a final recommendation list includes: allocating the dynamically determined click-through rate weight to the predicted scores of each content in the first recommendation subset; allocating the dynamically determined user retention rate weight to the predicted scores of each content in the second recommendation subset; allocating the dynamically determined content diversity weight to the predicted scores of each content in the third recommendation subset; sorting all content according to its weighted final score, and selecting the N highest-ranked content to generate the final recommendation list, where N is a preset positive integer.

[0093] In practice, a set of dynamically determined weight values ​​are obtained from the aforementioned reinforcement learning model or hybrid decision-making mechanism. Assume these three weights correspond to click-through rate, user retention rate, and content diversity objectives, respectively, denoted as W_ctr, W_retention, and W_diversity.

[0094] Furthermore, each content item in the first recommendation subset is assigned a base predicted score. This score is derived from the raw output of the model used to generate this subset (e.g., the neural network), such as the user-content node similarity score. Then, this base score is multiplied by a dynamically determined click-through rate weight W_ctr to obtain the weighted score of the content item on the first target dimension: Score1_i = BaseScore1_i * W_ctr.

[0095] Furthermore, for each content item in the second recommendation subset, its base score is derived from a similarity calculation based on a unified feature representation. This base score is multiplied by the dynamically determined user retention rate weight W_retention to obtain the weighted score of that content item on the second objective dimension: Score2_j = BaseScore2_j * W_retention.

[0096] Furthermore, for each content item in the third recommendation subset, its base score might be a score reflecting its contribution to improving list diversity. This base score is multiplied by the dynamically determined content diversity weight W_diversity to obtain the weighted score for that content item on the third objective dimension: Score3_k = BaseScore3_k * W_diversity.

[0097] It's important to note that a content item may appear in multiple subsets simultaneously. In this case, the content item will have multiple weighted scores. These weighted scores from different dimensions will be summed to obtain the final composite score for the content item. If a content item appears in only one subset, its composite score is the weighted score for that subset.

[0098] Furthermore, all content items to be recommended (usually the union of all subsets) are sorted from highest to lowest according to their calculated final composite scores. Starting from the top of the sorted list, the top N content items are selected sequentially, where N is a preset positive integer, such as 10 or 20. These N content items, in sorted order, constitute the final recommendation list presented to the user.

[0099] The linear weighted fusion and ranking method defined in this embodiment transforms the abstract dynamic multi-objective trade-off decision into a concrete, operable, and transparent score calculation process, thereby achieving effective integration of multi-strategy results and generation of the final recommendation list. The core effect of this method is to establish a clear bridge connecting "target weights" and "candidate content".

[0100] First, by directly associating each recommendation subset with a specific business objective weight and performing a multiplication operation, the importance of the objective directly regulates the recommendation results. For example, when the reinforcement learning model determines, based on real-time feedback, that more emphasis needs to be placed on user retention, it outputs a higher W_retention value. In the fusion step, this directly leads to an overall amplification of the weighted scores of each item in the second recommendation subset (based on deep interest matching, aimed at optimizing retention), making these items more likely to achieve a higher ranking in the final ranking. This association is direct and interpretable, ensuring that high-level weight decisions can unambiguously affect the underlying ranking results.

[0101] Furthermore, a weighted summation and re-sorting mechanism is employed to achieve a unified comparison of multi-dimensional scores. Each content item receives a score contribution from different optimization perspectives based on its appearance in different subsets. The final summed score comprehensively reflects the overall "value" of the content across multiple objectives. The sorting operation then globally selects the best candidate from all candidates based on this comprehensive value. This method is not only simple and efficient to operate, but it can also naturally handle situations where a content item simultaneously satisfies multiple objectives (e.g., a piece of content is popular with similar users, matches the current user's deep interests, and increases list diversity). Its final score will be the accumulation of multiple positive contributions, thus placing it reasonably higher in the list.

[0102] Therefore, this embodiment ensures that the final generated recommendation list is not a simple compilation, but an optimized result produced under a unified quantitative framework, comprehensively considering the dynamically changing business objective weights. This allows the composition of the recommendation list to accurately reflect the trade-off strategy for each objective at the current moment, a key step in realizing multi-objective recommendation from "parallel generation" to "integrated output".

[0103] In some preferred embodiments, the method further includes: collecting user interaction behaviors for each item in the final recommendation list in real time via a message queue, the interaction behaviors including clicking, dwell time, and swiping to skip; using a stream processing engine to statistically analyze the click-through rate, average dwell time, and information entropy based on content category of the final recommendation list in real time within a swiping time window; using the statistically obtained click-through rate, average dwell time, and information entropy as the real-time indicators to constitute training samples; employing an online gradient descent algorithm to incrementally update the network parameters of the reinforcement learning model using the training samples, with the goal of minimizing the loss of recommendation effect; and simultaneously using the training samples to incrementally update the parameters of the graph neural network model on which the first recommendation subset is based, the feature matching model on which the second recommendation subset is based, and the similarity calculation model on which the third recommendation subset is based.

[0104] In practice, user interaction behavior is first collected in real time through a message queue deployed on the server. Whenever a user clicks on content in the recommendation list, performs a swipe to skip, or the client reports the duration of time spent on a particular piece of content, these behavioral events are immediately encapsulated into messages and sent to the message queue. The message queue ensures asynchronous and reliable data reception under high concurrency.

[0105] Furthermore, a stream processing engine (such as Apache Flink) acts as a consumer, continuously pulling these interaction events from the message queue. The engine defines a sliding time window, for example, the past 5 minutes. Within this window, the engine aggregates and calculates key metrics in real time: it calculates the ratio of clicks to impressions for all content displayed in the window, obtaining the real-time click-through rate; it calculates the average dwell time for all clicked content; and simultaneously, it analyzes the category distribution of recommended and interacted content, calculating its information entropy as a diversity metric. These calculated metrics constitute a real-time metric set.

[0106] Furthermore, these real-time metrics are correlated with the contextual information used when the recommendation list was generated (such as user IDs, a list of content IDs, and the model version used) to form a training sample. This sample reflects the immediate reward obtained by adopting a specific recommendation strategy under specific conditions.

[0107] Furthermore, an online gradient descent algorithm is employed for updating the reinforcement learning model. This algorithm aims to minimize the loss in recommendation performance, and the loss function is typically designed based on temporal difference error. Using newly generated training samples, the algorithm calculates the gradient of the loss function with respect to the model parameters and immediately updates the network parameters of the reinforcement learning model with a small learning rate, fine-tuning its policy towards achieving higher rewards.

[0108] Furthermore, incremental updates are performed concurrently on the underlying recommendation generation model. Newly generated training samples are used again, but different incremental learning methods are employed depending on the model type. For the graph neural network model on which the first recommendation subset is generated, mini-batch gradient descent may be used to update the parameters of its graph encoder. For the feature matching model on which the second recommendation subset is generated, the parameters of its feature extraction network or similarity measurement network are updated using new data containing positive and negative sample pairs. For the similarity calculation model on which the third recommendation subset is generated, the parameters of its content feature encoder may be updated to calculate similarity more accurately. These updates are all incremental, designed to quickly incorporate the latest user feedback knowledge.

[0109] The real-time feedback and iteration technology solution defined in this embodiment constructs an efficient and low-latency closed loop from user behavior perception to model parameter update, completely transforming the recommended optimization mode from the traditional "offline batch processing" to "online continuous learning", thereby greatly improving agility and adaptability.

[0110] Traditional methods rely on accumulating massive amounts of historical data for periodic offline training and model releases, resulting in long iteration cycles and difficulty in capturing rapid changes in user interests and trends. The technical advantage of this solution is primarily reflected in its real-time performance. By combining message queues and stream processing engines, it can capture and process the latest user interaction signals with latency in the order of seconds or minutes, transforming them into quantifiable evaluation metrics. This means that the perception of recommendation effectiveness is almost instantaneous, providing a data foundation for rapid response.

[0111] Furthermore, the key technological advantage lies in the "incremental update" mechanism. By employing an online gradient descent algorithm to update model parameters on small batches or single samples, the enormous overhead and latency of retraining on full data are avoided. This allows the reinforcement learning model to instantly adjust its weight decision-making strategy based on the latest reward feedback, while the underlying graph neural network and feature matching model can quickly encode the latest user preference information into its parameters. For example, when a new trending topic begins to receive a large number of clicks and dwell times, the stream processing engine quickly identifies its high click-through rate and long dwell time, forming positive feedback samples. The incremental learning algorithm can rapidly use these samples to update the model, increasing the recommendation weight and understanding ability for similar content in a very short time, thereby quickly capturing trending topics.

[0112] Therefore, this embodiment makes the entire recommendation process a living, continuously evolving organism. It can adapt to subtle changes in user behavior and dynamic evolution of market content at near real-time speeds, significantly shortening the feedback loop from observing changes to adjusting strategies. This ensures that the recommendation results remain timely, accurate, and satisfying to users, which is a key advantage in maintaining recommendation competitiveness in today's rapidly changing internet environment.

[0113] Technical effects of the present invention

[0114] The method defined in this invention achieves significant technical results by constructing a complete technical solution integrating multimodal feature fusion, dynamic weight decision-making, multi-strategy generation, and a real-time feedback closed loop. First, this method breaks through the limitations of traditional recommendation systems that rely on a single objective or static weight combination. By introducing a reinforcement learning model to dynamically determine the weights of multiple objectives such as click-through rate, user retention rate, and content diversity, it can automatically adjust the optimization direction based on real-time feedback and scenario information. For example, when it detects that a user is rapidly skipping similar content consecutively, the diversity weight can be automatically increased, guiding subsequent recommended content to be more diverse, thereby effectively alleviating the information cocoon problem and improving the adaptability to user interest drift and diverse scenarios.

[0115] Furthermore, this invention integrates heterogeneous features of users, content, and scenarios into a unified representation through a cross-modal attention mechanism, and executes three recommendation strategies with different focuses in parallel based on this deep representation. This design enables the simultaneous utilization of group behavior patterns, individual deep interest features, and diversity protection mechanisms. The fusion of these strategy results based on dynamic weights means that the final recommendation list is not the product of a single strategy, but rather the result of collaborative optimization of multiple business objectives. This mechanism ensures that the recommendation results not only guarantee a basic click-through rate, but also improve user satisfaction and long-term retention through deep interest matching, and maintain an ecological health through diversity control, achieving an effective balance and comprehensive improvement among multiple objectives.

[0116] Furthermore, this invention tightly integrates real-time feedback data into the model update loop, forming a highly efficient online learning closed loop. Each user interaction is rapidly converted into a signal for model optimization, enabling the recommendation strategy to adapt to changes in user needs at near real-time speed. This rapid iteration capability significantly shortens the algorithm response cycle compared to traditional offline training and periodic deployment models, allowing recommendations to more quickly capture trending topics and subtle changes in user behavior patterns, thereby continuously improving the timeliness and accuracy of recommendations. In summary, the method defined in this embodiment, through its unique combination of technical features, achieves a synergistic improvement in the adaptability, overall performance, and iterative efficiency of recommendations.

[0117] Please see Figure 2 , Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0118] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0119] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it causes the processor 502 to perform a multi-objective content recommendation method.

[0120] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0121] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a multi-target content recommendation method.

[0122] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0123] The processor 502 is used to run a computer program 5032 stored in the memory to implement a multi-target content recommendation method.

[0124] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0125] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0126] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform steps of a multi-target content recommendation method.

[0127] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0128] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0129] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0130] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0131] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0132] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0133] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-objective content recommendation method, characterized in that, include: Collect and preprocess user data, multimodal content data, and scenario data; Feature extraction is performed on the preprocessed user data, content multimodal data, and scene data to obtain user feature vectors, content feature vectors, and scene feature vectors. The user feature vectors, content feature vectors, and scene feature vectors are then fused into a unified feature representation through a cross-modal attention mechanism. Based on the current business scenario and real-time feedback data, the weights of multiple recommendation objectives are dynamically determined through a reinforcement learning model. These multiple recommendation objectives include at least click-through rate, user retention rate, and content diversity. A first recommendation subset is generated based on the interaction between users and content; a second recommendation subset is generated based on the correlation between content and user interests calculated based on the unified feature representation; and a third recommendation subset is generated by filtering candidate content whose similarity to the selected content is lower than a threshold. The first recommendation subset, the second recommendation subset, and the third recommendation subset are then merged based on the dynamically determined weights to generate the final recommendation list.

2. The multi-objective content recommendation method according to claim 1, characterized in that, Feature extraction is performed on the user data to obtain a user feature vector, including: The user's historical behavior sequence is arranged in chronological order, and the historical behavior sequence includes the user's clicks, pauses, and favorites of content; The historical behavior sequence is input into the Transformer model, and the multi-head self-attention mechanism in the Transformer model is used to calculate the dependencies between the behaviors in the sequence. Extract the mean or the vector at the last position of the sequence representation vector output by the Transformer model as the user feature vector; Feature extraction is performed on the content multimodal data to obtain content feature vectors, including: For text data, local semantics are extracted through convolutional neural networks, and / or contextual semantic embeddings are obtained through pre-trained language models to obtain text features; For image or video frame data, visual features are extracted using a convolutional neural network; For audio or time-series video data, temporal variation features are extracted using a recurrent neural network; The extracted text features, visual features, and temporal change features are concatenated or weighted and fused, and associated with the corresponding entity information in the preset knowledge graph to generate the content feature vector; Feature extraction is performed on the scene data to obtain a scene feature vector, including: The scenario data is categorized into time-dimensional data, device-dimensional data, and network environment-dimensional data. For time-dimensional data, extract the hour information, weekday information, and whether it is a holiday information at the current moment; map the hour information into sine and cosine codes with a 24-hour period, perform one-hot encoding on the weekday information, and convert the holiday information into Boolean value codes; For device-level data, extract device type, screen size, and operating system version; perform one-hot encoding on device type and operating system version, and divide screen size into predefined size ranges and perform range encoding; For network environment dimension data, extract network type and bandwidth level, and perform one-hot encoding; The time dimension encoding, device dimension encoding, and network environment dimension encoding mentioned above are concatenated to form the initial scene encoding vector; The initial scene encoding vector is input into a fully connected neural network for nonlinear transformation and dimensionality reduction, and the scene feature vector is output.

3. The multi-objective content recommendation method according to claim 1, characterized in that, The step of dynamically determining the weights of multiple recommendation targets using a reinforcement learning model includes: At least one of the user feature vector, content feature vector, and scene feature vector, along with the corresponding historical weight values, are used as the state input of the reinforcement learning model. Define a reward function that aims to comprehensively measure long-term user value. The reward function is obtained by weighted summation of real-time metrics of the multiple recommendation objectives. Using the temporal difference algorithm, the Q-value function in the reinforcement learning model is updated based on the current state, the action taken, and the immediate reward obtained in the next state. The updated Q-value function is then used to output an adjustment action for the weights of the multiple recommendation targets, which is performed within a preset weight range.

4. The multi-objective content recommendation method according to claim 3, characterized in that, Also includes: Pre-configure weight adjustment rules associated with specific business scenarios, such as new user cold start or promotional activities; When the current business scenario is detected to match the specific business scenario, the weights of the multiple recommendation targets are first adjusted in a one-time basic adjustment according to the weight adjustment rules; The weights after the basic adjustment are used as the initial state of the reinforcement learning model, and then the reinforcement learning model is fine-tuned online based on subsequent real-time feedback data.

5. The multi-objective content recommendation method according to claim 1, characterized in that: The generation of the first recommendation subset based on the interaction relationship between users and content includes: constructing a heterogeneous graph with users and content as nodes, using a graph neural network to aggregate information of neighboring nodes, updating node representations, and selecting the Top-K content with the highest similarity as the first recommendation subset based on the inner product or cosine similarity between the updated user node and content node representations. The step of generating a second recommendation subset based on the correlation between content and user interests calculated by the unified feature representation includes: calculating the similarity between the feature vector of candidate content and the feature vector of the current user in the feature space, and using a contrastive learning loss function to optimize the feature space so that the feature distance of positive sample pairs decreases and the feature distance of negative sample pairs increases, and finally selecting the Top-K content with the highest similarity to the user feature vector as the second recommendation subset; The step of filtering candidate content with similarity to selected content below a threshold to generate a third recommendation subset includes: calculating the cosine similarity between candidate content and content already selected in the recommendation list; if the similarity between candidate content and any content in the selected list is below a preset threshold, then adding the candidate content to the candidate pool; and selecting content from the candidate pool that maximizes the diversity score of the overall recommendation list according to the maximum marginal utility algorithm to form the third recommendation subset.

6. The multi-objective content recommendation method according to claim 5, characterized in that, The process of fusing the first, second, and third recommendation subsets based on dynamically determined weights to generate a final recommendation list includes: The dynamically determined click-through rate weights are assigned to the predicted scores of each item in the first recommendation subset; The dynamically determined user retention rate weights are assigned to the predicted scores of each item in the second recommendation subset; The dynamically determined content diversity weights are assigned to the predicted scores of each content in the third recommendation subset; All content is sorted according to its weighted final score, and the top N content items are selected to generate the final recommendation list, where N is a preset positive integer.

7. The multi-objective content recommendation method according to claim 1, characterized in that, The method further includes: The system collects user interaction behavior for each item in the final recommendation list in real time via a message queue. The interaction behavior includes clicking, dwell time, and swiping to skip. Using a stream processing engine, the click-through rate, average dwell time, and information entropy based on content category of the final recommendation list are statistically analyzed in real time within a sliding time window. The click-through rate, average dwell time, and information entropy obtained from the statistics are used as the real-time indicators to form the training sample. An online gradient descent algorithm is used to minimize the loss in recommendation performance, and the network parameters of the reinforcement learning model are updated using the incremental training samples. Simultaneously, the parameters of the graph neural network model on which the first recommendation subset is based, the feature matching model on which the second recommendation subset is based, and the similarity calculation model on which the third recommendation subset is based are updated in an incremental learning manner using the training samples.

8. The multi-objective content recommendation method according to claim 2, characterized in that, The method further includes: When the target is identified as a new user, obtain the attribute information provided by the new user during registration; when the target is identified as new content, obtain the metadata tags of the new content. The attribute information or metadata tags are matched with the pre-built knowledge graph to obtain the embedding vector of the corresponding entity in the knowledge graph; For new users, their behavior sequence is input into the Transformer model to extract an initial behavior feature vector, and this initial behavior feature vector is concatenated with the embedding vector obtained from the knowledge graph to form an enhanced user feature vector; For new content, a multimodal original feature vector is extracted using a convolutional neural network or a recurrent neural network, and this original feature vector is concatenated with the embedding vector obtained from the knowledge graph to form an enhanced content feature vector.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the multi-objective content recommendation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the multi-objective content recommendation method as described in any one of claims 1-7.

Images