A time-aware hybrid expert routing method and system for breaking news distribution

Abstract

The application discloses a time-sensitive hybrid expert routing method and system for breaking news distribution. The method comprises the following steps: extracting the timeliness features of candidate news, including the time length after publication and the touch rate change features; inputting the timeliness features into a timeliness gating network to output a hard routing decision; distributing the candidate news to a target expert network in a heterogeneous expert network according to the hard routing decision, including lightweight and heavy expert networks with different computational complexities, and calling the target expert network to output a prediction score; wherein the timeliness gating network is trained based on a joint loss function including a prediction loss and a computational complexity constraint loss term. The application drives the on-demand offloading of computing resources through timeliness features, combines an asynchronous speculation mechanism to realize inference parallelization, significantly reduces the delay while ensuring accuracy, avoids cold start overfitting, and is widely applicable to news, short video and other strong timeliness recommendation scenarios.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and recommendation systems, specifically to a timeliness-aware hybrid expert routing method and system for breaking news distribution. Background Technology

[0002] Recommendation systems, as the core engine of modern information distribution platforms, have been widely used in news, short video streams, and social media scenarios. In recent years, complex deep learning ranking models, represented by Deep Interest Networks (DIN) and Deep Cross Networks (DCN), have become industry standards. These models can achieve extremely high accuracy in predicting click-through rate (CTR) and conversion rate (CVR) by introducing high-order feature crossover and long sequence attention mechanisms. However, with the exponential growth of the candidate set of recommendation systems, the existing "isomorphic reasoning paradigm" (i.e., using a unified heavy network to score all candidate content) has exposed the following irreconcilable technical bottlenecks when facing the "high timeliness and high concurrency" of sudden news distribution scenarios: (1) Inverted computing power and extreme waste of computing resources: In real large-scale industrial recommendation pipelines, a single user request often requires fine-tuning and scoring hundreds to thousands of candidate news articles, accumulating an extremely large number of floating-point operations (FLOPs). Existing isomorphic deep models lack the ability to perceive the life cycle of news. Especially for newly added breaking news or cold start news, the audience and content value are still in an unverified state (high uncertainty), but the system still forcibly calls heavy networks with large parameters to calculate complex feature interactions. This "one-size-fits-all" homogeneous computing allocation method has led to a serious "computing power inversion" phenomenon, resulting in a huge waste of valuable online GPU resources; (2) Missed golden distribution window caused by computing delay: The core value of breaking news lies in "timeliness", and its best distribution window is often limited to 1 to 2 hours after publication. The high complexity of heavy network inference will directly prolong the system's response time, and may even cause serious request queuing and degradation during peak traffic hours. This extremely high computing delay makes it impossible for the system to digest and distribute the surge of breaking news in time, causing the news to decline in popularity while waiting in the queue, which seriously damages the reach of breaking events and the media influence of the platform; (3) Parameter redundancy and overfitting trap during the cold start period: Newly published news often lacks sufficient user history interaction behavior (such as clicks, dwell time, etc.), resulting in extremely sparse behavioral sequence features of the input model. With such sparse data distribution, forcibly using heavy networks (such as deep interest networks / deep cross networks) with tens or even hundreds of millions of parameters for inference can easily lead to the model fitting into local noise, causing serious overfitting problems. In this case, the prediction accuracy of the heavy model is even worse than that of a lightweight dual-tower model or a simple heuristic recall strategy.

[0003] To alleviate the aforementioned problems, existing technologies in the industry have proposed several optimization solutions, but all have limitations. First, existing hybrid expert architectures rely heavily on "user profile features" or "multi-task attributes (such as clicks and likes)" for their routing gating signals, rather than "item lifecycles." Such architectures cannot dynamically offload computation based on the timeliness of news and often employ "soft routing" (i.e., weighted fusion of multiple experts), which essentially does not reduce the total computational cost (FLOPs) per request. Second, while some technologies attempt to incorporate "publication time features" of news into the model input or introduce "time decay penalties" into the loss function, this only remains at a superficial level of "feature weighting." The underlying model structure remains statically isomorphic, failing to fundamentally achieve elastic offloading of computational resources. Furthermore, some systems have hard-coded static rule bypasses for cold-start content, but these solutions lack end-to-end joint optimization mechanisms and cannot adapt to the entire lifecycle evolution of news from "cold start" to "hot surge" and then to "decline," easily leading to drastic fluctuations in predicted scores during stage transitions. In summary, the industry urgently needs a new routing architecture that can sense the dynamic timeliness of content and accordingly achieve elastic allocation of computing resources at the underlying structure, so as to achieve the optimal balance between the efficiency of breaking news recommendation and the accuracy of prediction under strict inference latency constraints. Summary of the Invention

[0004] This invention provides a timeliness-aware hybrid expert routing method and system for breaking news distribution, aiming to solve the technical problems of existing isomorphic recommendation models when processing breaking news, such as serious waste of computing resources, missed golden distribution windows due to computational delays, and overfitting caused by sparse cold-start data.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides a timeliness-aware hybrid expert routing method for breaking news distribution, comprising:

[0007] Step 1: Obtain the candidate news set output by the upstream coarse ranking module, and extract the timeliness features of each candidate news item in the candidate news set; wherein, the timeliness features include at least the duration after publication and the change in reach rate.

[0008] Step 2: Input the timeliness features into a pre-trained timeliness gating network and output a hard routing decision for candidate news; wherein, the hard routing decision is used to instruct the target expert network to uniquely assign the candidate news to a target expert network among multiple heterogeneous expert networks; the multiple heterogeneous expert networks include at least lightweight expert networks and heavy expert networks with different computational complexities.

[0009] Step 3: Based on the hard routing decision, distribute candidate news to the corresponding target expert network;

[0010] Step 4: Call the target expert network to perform forward reasoning on the candidate news, output the predicted score, and pass the predicted score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy.

[0011] The time-sensitive gating network is trained end-to-end based on a joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource overhead.

[0012] Furthermore, in step 1, the timeliness characteristic specifically refers to:

[0013] The duration after publication is ,in, This is the hourly difference between the time the news was published and the time of the current request.

[0014] The reach rate change feature is the first derivative of the reach rate of candidate news within a preset sliding time window.

[0015] Furthermore, the timeliness feature also includes the content semantic vector of the candidate news, which is a high-dimensional dense feature obtained by encoding the textual metadata of the news through a pre-trained deep semantic model.

[0016] Furthermore, the specific structure of the time-sensitive gating network includes a cascaded input layer, a multilayer perceptron, and a Softmax activation layer; wherein, the time-sensitive gating network outputs the routing probability distribution of each expert network through the Softmax activation layer.

[0017] Furthermore, in step 2, the hard routing decision is specifically as follows:

[0018]

[0019] in, For the selected target expert network index, For the first The probability value output by the timeliness gating network of an expert network. ,in, To determine the number of expert networks, through The algorithm forces candidate news to activate only the single expert network path.

[0020] Furthermore, in step 2, the lightweight expert network is a dual-tower model structure based on vector dot product or cosine similarity; the heavy expert network is a deep interactive model with feature cross layers or attention mechanisms, and the number of parameters and computational complexity (FLOPs) of the heavy expert network are higher than those of the lightweight expert network.

[0021] Furthermore, the complexity constraint loss term is specifically set as follows:

[0022]

[0023] in, This represents the proportion of samples in the current training batch that are routed to the heavyweight expert network. The maximum percentage of heavy-duty experts is set as preset. This is a preset penalty coefficient.

[0024] Furthermore, step 1 also includes smoothing the timeliness characteristics of very recent news:

[0025] When the time elapsed since the publication of a candidate news item is less than a preset time threshold, the initial characteristics of the candidate news item are filled or weighted smoothed using the average reach rate change characteristics of existing news items in the same category.

[0026] Furthermore, asynchronous speculative execution optimization is also included between steps 2 and 4:

[0027] While the timeliness-gated network calculates the routing probabilities of each expert network, the lightweight expert network is asynchronously invoked to perform inference on candidate news and obtain the inference prediction score; if the hard routing decision points to the lightweight expert network, the inference prediction score is directly used as the final result.

[0028] Among them, asynchronous speculative execution optimization is only executed when preset triggering conditions are met;

[0029] The preset triggering conditions are: the duration after the candidate news is published is less than the first preset time threshold, or the duration after the candidate news is published is greater than the second preset time threshold and its reach rate change characteristics are lower than the preset activity threshold.

[0030] Secondly, the present invention also provides a timeliness-aware hybrid expert routing system for the distribution of breaking news, comprising: a timeliness feature extraction module for extracting the characteristics of changes in duration and reach after publication, a timeliness gating module for outputting hard routing decisions, a heterogeneous expert network module including a lightweight expert network and a heavy expert network, and a prediction scoring module for outputting the estimated score.

[0031] in,

[0032] The timeliness feature extraction module is used to obtain the candidate news set output by the upstream coarse ranking module and extract the timeliness features of each candidate news in the candidate news set. The timeliness features include at least the duration after publication and the change in reach.

[0033] The timeliness gating module is used to input timeliness features into a pre-trained timeliness gating network and output a hard routing decision for candidate news; wherein, the hard routing decision is used to instruct the target expert network to uniquely assign the candidate news to multiple heterogeneous expert networks.

[0034] The heterogeneous expert routing module is used to distribute candidate news to the corresponding target expert network based on hard routing decisions. The multiple heterogeneous expert networks include at least lightweight expert networks and heavyweight expert networks with different computational complexities.

[0035] The prediction scoring module is used to call the target expert network to perform forward reasoning on candidate news, output the prediction score, and pass the prediction score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy.

[0036] The model training module is used to train the time-sensitive gated network based on the joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource overhead.

[0037] Furthermore, the timeliness feature extraction module is specifically used for:

[0038] The program employs a logarithmic scale to calculate the post-publication duration; calculates the first derivative of the reach rate of candidate news within a preset sliding time window as a feature of reach rate change; and extracts the content semantic vector of candidate news through a pre-trained deep semantic model.

[0039] Furthermore, the timeliness gating module executes... The computation implements hard routing decisions to force candidate news to activate only a single expert network path; the lightweight expert network is a dual-tower model structure, while the heavy expert network is a deep interactive model with feature cross layers or attention mechanisms, and the heavy expert network has a higher number of parameters and computational complexity than the lightweight expert network.

[0040] Furthermore, the timeliness-aware hybrid expert routing system for breaking news distribution provided by this invention also includes:

[0041] The feature smoothing submodule is used to fill or weight smooth the initial features of candidate news by using the average reach rate change feature of existing news in the category when the time since the publication of candidate news is less than a preset time threshold.

[0042] The speculative execution submodule is used to start the inference link of the lightweight expert network in parallel while the timeliness gating module is operating, and decide whether to directly adopt the output of speculative inference based on the routing decision results.

[0043] The technical solution provided by this invention brings at least the following beneficial effects:

[0044] (1) Hard routing mechanism breaks the homogeneous computing power inversion and achieves extreme inference latency: Unlike the soft weighted fusion of traditional hybrid expert routing, this invention innovatively introduces The hard routing mechanism forces each candidate news item to penetrate only a single expert network. By diverting a massive amount of news in the "cold start" or "long tail decay" phase to a lightweight expert network, the redundant floating-point operations are offloaded from the underlying structure, significantly reducing the overall inference latency of the system and ensuring that breaking news can be quickly calculated and displayed within the golden distribution window.

[0045] (2) Lifecycle awareness and heterogeneous network decoupling effectively avoid cold start overfitting: For the first time, the dynamic timeliness characteristics of news (publication duration, reach rate derivative) are used as the core gating signal for routing. When news is in the cold start period with extremely little behavioral sequence, the system automatically routes it to the lightweight dual-tower model that is not prone to overfitting; when news enters the rapid dissemination period and accumulates rich interactive features, it is then routed to the heavy network with a complex attention mechanism to mine deep interests. This heterogeneous computing power matching completely solves the overfitting trap caused by data sparsity.

[0046] (3) Computational complexity constraint training to achieve the optimal dynamic balance between efficiency and effectiveness: Abandoning rigid hard-coded manual rules, this invention directly introduces a computational complexity constraint loss term during the training stage of the time-gated network. The system can adaptively learn routing strategies that meet the actual online hardware capacity limitations through gradient descent (such as strictly constraining the proportion of heavy expert calls to 30% to 50%), maximizing the utilization of the online GPU cluster without harming or even improving the overall click-through rate.

[0047] (4) Industrial-grade smoothing and speculative fallback design to ensure stability in extreme scenarios: To address the statistical variance problem in the early stages of a sudden event, a smoothing rollback mechanism for similar categories was designed; through an asynchronous speculative execution strategy, the parallelization of decision-making and reasoning was achieved, significantly improving the system's throughput limit during peak periods. These engineering closed-loop mechanisms ensure that the ranking score remains stable when the system faces extreme sudden traffic surges, and there will be no service interruption or severe fluctuations. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1A flowchart illustrating the steps of a timeliness-aware hybrid expert routing method for breaking news distribution provided in an embodiment of the present invention;

[0050] Figure 2 A schematic diagram illustrating the application architecture and insertion position of a timeliness-aware hybrid expert routing system for breaking news distribution provided in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the time-sensitive gated network and heterogeneous expert routing mechanism provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0053] This invention provides a timeliness-aware hybrid expert routing method for breaking news distribution, such as... Figure 1 As shown, it specifically includes:

[0054] Step S1: Obtain the candidate news set output by the upstream coarse ranking module and extract failure features including post-publication duration and reach rate changes;

[0055] That is, the timeliness features of each candidate news item in the candidate news set are extracted through step S1;

[0056] Step S2 (Main Link): Input the features into the timeliness-gated network, calculate the routing probability of each expert network, and output the hard routing decision;

[0057] Hard routing decisions are used to instruct the target expert network to uniquely assign candidate news to multiple heterogeneous expert networks.

[0058] Step S3: Determine whether the hard routing decision points to the lightweight expert network. If yes, execute the zero-latency return in step S4; otherwise, execute the normal inference in step S4.

[0059] Step S4: This includes zero-latency return and normal inference, as follows:

[0060] Zero-latency return: Directly adopt the inferred prediction score, and output the final prediction score in combination with the disaster recovery backup mechanism;

[0061] Normal reasoning: Discard the speculative results, call the heavy expert network for reasoning, and output the final predicted score with a fallback mechanism.

[0062] In the application process, step S3 distributes candidate news to the corresponding target expert network based on hard routing decisions; then in step S4, the target expert network is invoked to perform forward reasoning on the candidate news, output a predicted score, and pass the predicted score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy.

[0063] The time-sensitive gating network is trained end-to-end based on a joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource overhead.

[0064] Furthermore, when the preset triggering conditions are met, the process between steps S2 and S4 includes asynchronous inference execution optimization, i.e., parallel inference link: when the triggering conditions are met, a lightweight expert network is asynchronously invoked to perform inference reasoning and obtain the inference prediction score, i.e., the inference score is prepared in advance. The preset triggering conditions are: the post-publication duration of the candidate news is less than a first preset time threshold, or the post-publication duration of the candidate news is greater than a second preset time threshold and its reach rate change characteristics are lower than a preset activity threshold.

[0065] Specifically, the post-publication duration is represented using a logarithmic scale; the reach rate change feature is the first derivative of the reach rate of the candidate news within a preset sliding time window. Furthermore, the timeliness feature also includes the content semantic vector of the candidate news extracted through a pre-trained deep semantic model. Hard routing decisions are executed by applying the probability distribution output by the timeliness-gated network. The computation is implemented to force candidate news to activate only a single expert network path. The lightweight expert network is a dual-tower model structure based on vector dot product or cosine similarity; the heavyweight expert network is a deep interaction model with feature cross layers or attention mechanisms (such as deep interest networks or deep cross networks).

[0066] This invention also provides a timeliness-aware hybrid expert routing system for breaking news distribution, comprising: a timeliness feature extraction module for extracting post-publication duration and reach rate change features; a timeliness gating module for outputting hard routing decisions; a heterogeneous expert network module including a lightweight expert network and a heavyweight expert network; and a prediction scoring module for outputting estimated scores; wherein, the timeliness feature extraction module is used to obtain a candidate news set output by the upstream coarse ranking module and extract the timeliness features of each candidate news item in the candidate news set, the timeliness features including at least post-publication duration and reach rate change features. The system includes the following components: a timeliness gating module, which inputs timeliness features into a pre-trained timeliness gating network and outputs a hard routing decision for candidate news; the hard routing decision instructs the candidate news to be uniquely assigned to a target expert network among multiple heterogeneous expert networks; a heterogeneous expert routing module, which distributes candidate news to the corresponding target expert network according to the hard routing decision, wherein the multiple heterogeneous expert networks include at least lightweight expert networks and heavyweight expert networks with different computational complexities; a prediction scoring module, which calls the target expert network to perform forward inference on the candidate news, outputs a prediction score, and passes the prediction score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy; and a model training module, which trains the timeliness gating network based on a joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource consumption.

[0067] Specifically, the timeliness feature extraction module is further used to: calculate the post-publication duration using a logarithmic scale; calculate the first derivative of the reach rate of candidate news within a preset sliding time window as a reach rate change feature; and extract the content semantic vector of candidate news through a pre-trained deep semantic model. The computation implements hard routing decisions to force candidate news to activate only a single expert network path; the lightweight expert network is a dual-tower model structure, while the heavy expert network is a deep interactive model with feature cross layers or attention mechanisms, and the heavy expert network has a higher number of parameters and computational complexity than the lightweight expert network.

[0068] Furthermore, the timeliness-aware hybrid expert routing system for breaking news distribution provided in this embodiment of the invention further includes: a feature smoothing submodule, used to fill or weight smooth the initial features of the candidate news by using the average reach rate change features of existing news in the relevant category when the time since the release of the candidate news is less than a preset time threshold; and a speculative execution submodule, used to start the inference link of the lightweight expert network in parallel while the timeliness gating module is operating, and decide whether to directly adopt the output of speculative inference based on the routing decision result.

[0069] The timeliness-aware hybrid expert routing method and system for breaking news distribution provided in this invention can be widely applied to news and information apps, short video social platforms, and various information flow systems with strong timeliness distribution requirements. See also... Figure 2 This information flow system, acting as the computational engine for the recommendation fine-ranking layer, receives candidate sets and real-time features from the upstream recall and coarse-ranking modules, and outputs the calculated predicted scores to the downstream re-ranking and distribution modules. The system includes: a recall module, a timeliness feature extraction module, a timeliness gating module, an expert routing switch, a lightweight expert network, a heavy expert network, and a final ranking module (i.e., a predicted score module). The expert routing switch, lightweight expert network, and heavy expert network constitute a heterogeneous expert network module. At the discrete time step of the initiated user request (recommendation request), the recall module outputs the corresponding candidate news set and, through timeliness... The feature extraction module extracts the timeliness feature vector of each news item in the candidate news set and inputs it into the timeliness gating module. The timeliness gating module outputs a hard routing decision for the candidate news item through the timeliness gating network. That is, the hard routing decision argmax operation forces the candidate news item to activate only a single expert network path. The result of the hard routing decision argmax operation determines the expert routing switch to distribute the candidate news item to the corresponding lightweight expert network or heavy expert network. The predicted score is obtained based on the predicted score output by the lightweight expert network and the heavy expert network. Finally, the final ranking module generates a recommendation list and returns it to the corresponding user request. The core logic of this system is: by extracting the dynamic timeliness features of candidate news in real time, a hard routing gating network is driven to allocate news at different life cycle stages to expert networks with matching computational complexity for processing, thereby maximizing system throughput and reducing inference latency while ensuring recommendation accuracy.

[0070] See Figure 3 The specific steps of a timeliness-aware hybrid expert routing system for breaking news distribution provided in this application embodiment include:

[0071] Step 1: Real-time extraction and transformation of multidimensional time-sensitive features

[0072] At discrete time steps when a user initiates a recommendation request The system targets the candidate newsletter set issued by the recall module. Every news item Extract its dynamic timeliness feature vector. :

[0073] (1) Duration after publication: Calculate the time difference between the news publication timestamp and the current system time. (Hours). To compress the long-tailed distribution and improve the model's sensitivity to newly released news, a logarithmic scaling transformation is used: ,in, Duration after publication.

[0074] (2) Reach Slope: Extracts the reach slope of the news within the most recent preset sliding time window. Specifically, the first derivative of the reach slope is calculated to characterize the current trend of the news's popularity.

[0075] (3) Semantic content vector: Extracting high-dimensional vector representations of news titles and summaries It is used to capture the category attribute to which the news belongs.

[0076] The extracted multidimensional features are then concatenated after normalization to form a timeliness feature vector. Its mathematical expression is: .in, , The first Characteristics of changes in the duration and reach of news articles after their publication.

[0077] For very recent news with a publication duration of less than a preset threshold (e.g., 10 minutes), due to insufficient statistical samples of its reach rate, the system performs feature smoothing processing: it fills in the gaps by using the average reach rate variation characteristics of existing news in the category to which the news belongs, in order to eliminate statistical noise in the early stage of cold start.

[0078] The timeliness feature vector (including post-publication duration, reach rate changes, and semantic content vectors) is directly obtained from the upstream recommendation support system via a standard data interface. As a mature feature supply center, the upstream recommendation support system is capable of providing the aforementioned high-dimensional and dynamic features, as detailed below:

[0079] (1) Acquisition of semantic content vector: This feature is provided by the content understanding engine of the upstream recommendation support system. During the candidate news entry stage, the upstream module has extracted local implicit semantics, including position embeddings and fragment embeddings, by calling pre-trained language models such as BERT (Bidirectional Encoder Representation Transformer), and aggregated them into a highly condensed fixed-length news-level semantic vector. Before performing gated routing, this system only needs to directly pull this dimensionality-reduced dense representation through the feature library interface.

[0080] (2) Direct reading of post-publication duration: Post-publication duration is the standard time baseline for upstream modules to address the cold start challenge. The upstream recommendation support system writes a timestamp accurate to milliseconds into the metadata table the moment the news is entered into the database. This system can obtain this feature by directly reading this timestamp, subtracting it from the absolute time of the current request, and then performing a logarithmic scaling transformation.

[0081] (3) Streaming subscription for reach rate change characteristics: For the dynamic characteristic of reach rate change, the upstream recommendation support system (such as the system based on Monolith architecture) has been configured with a real-time feature repository. The upstream stream processing framework aggregates click and exposure logs in real time and calculates the instantaneous exposure growth rate (i.e., the derivative within the sliding window) that reflects the propagation momentum. In addition, the upstream system ensures conflict-free storage of features in a high-concurrency environment through Cuckoo collision-free hash table (Cuckoo HashMap) technology.

[0082] Step 2: Decision allocation based on hard-router gated network

[0083] The system will use timeliness feature vectors The input is fed into a timeliness-gated network. In a preferred embodiment, the dimension of the timeliness feature input vector is typically set to 64 to 128 dimensions.

[0084] This gated network consists of a multi-layer sensing architecture. The first layer has 64 fully connected sensing nodes, and the second layer has 32 nodes. The routing probability distribution of each expert network is output through the Softmax activation layer. (In this embodiment) .

[0085] This invention explicitly employs a hard routing mechanism. By executing... The algorithm performs calculations and selects the expert path with the highest probability as the target expert network. : .

[0086] Unlike the weighted summation of traditional hybrid expert models, this mechanism forces a single candidate news item to penetrate only a single expert network path, thereby physically offloading all computational overhead of non-target paths during the inference phase.

[0087] Step 3: Computational power matching and scoring of heterogeneous expert networks

[0088] Based on hard routing decisions, the system distributes candidate news items to one of the following heterogeneous experts. It should be noted that the input to the activated target expert network includes not only the aforementioned timeliness features, but also all the features required for regular recommendation (including but not limited to user profile features, user historical behavior sequence features, contextual features, and basic news attribute features) to ensure that the expert network has sufficient information for accurate click-through rate (CTR) prediction.

[0089] (1) Lightweight expert network: It adopts a vector dual-tower structure with extremely low computational complexity. To ensure extremely fast dot product calculation, the embedding vector dimensions of the news side and user side of the lightweight expert network are preferably 32 to 64 dimensions. The expert is mainly responsible for processing: very new news in the early stage of publication (features are sparse and suitable for simple models) or long-tail news at the end of its life cycle (lower value and no need for heavy computation).

[0090] Although the lightweight expert network adopts a dual-tower model structure similar to the upstream coarse-ranking module, their technical functions and feature domains differ fundamentally in this system. First, in the feature perception domain, the upstream coarse-ranking dual-tower model is limited by the latency constraints of full-database retrieval and can only process low-dimensional static features; while the lightweight expert network of this invention is deployed in the fine-ranking layer, and its input fully integrates multi-dimensional dynamic time-sensitive features (such as the first derivative of reach rate) transmitted by the timeliness-gated network, achieving refined shallow ranking. Second, in terms of data distribution and optimization objectives, the lightweight expert network specializes in processing specific sub-distributed data (i.e., sparse news with features in the cold start or long-tail decay phase) diverted by hard routing decisions. Compared to the upstream coarse-ranking model that fits the global distribution, this lightweight expert network converges independently on specific sparse distributions, thus becoming a dedicated network for solving the "overfitting of new news cold start" problem, forming a perfect heterogeneous computational complement to the heavyweight expert network.

[0091] (2) Heavyweight expert network: This type of network employs a heavyweight network with complex feature crossover or attention mechanisms (such as deep interest network or deep crossover network). To preserve sufficient feature interaction space, the dimension of its underlying feature embedding vector is typically set to 128 to 256 dimensions. This heavyweight expert network is mainly responsible for processing breaking news during periods of high popularity or high-value dissemination, accurately fitting user interests through high-intensity computing power.

[0092] Step 4: Training under the constraint of joint loss function

[0093] To significantly reduce the computational overhead during model training and prevent catastrophic forgetting during the training of expert networks, this embodiment freezes the parameters of the lightweight expert network and the heavy expert network during the training phase, and only updates the parameters of the time-sensitive gating network.

[0094] The system collects historical distribution data from real online recommendation stream logs to construct a training set, with each sample... Defined as a set of features and labels: .feature Includes the timeliness feature vector of candidate news at the time of its exposure. (i.e., post-publication duration, reach changes, and semantic content vectors) and other necessary contextual features; tags Then, construct a sample based on real user feedback (e.g., clicks or valid dwell times are set as positive samples). Only exposures without clicks are set as negative samples. In the data preprocessing stage, to address potential outliers (i.e., extreme outliers in the long-tail distribution) in features such as reach rate, this system does not employ a global threshold truncation method that disrupts the original information structure. Instead, it introduces an empirical cumulative distribution mapping mechanism based on a dynamic feature statistical window. Specifically, within a rolling feature statistical window (e.g., setting the window to cover two hours before sample exposure), the system calculates the relative ranking of each news feature value in the current sample set and maps it to quantiles, replacing the original absolute values ​​that are prone to gradient explosion. This processing mechanism removes the absolute destructive power of extreme abnormal noise while losslessly preserving the relative popularity ranking of outliers (i.e., avoiding erasing the "sudden" attribute of breaking news itself), and strictly maintains the uniform distribution of preprocessed timeliness features within the overall time window. This mechanism effectively avoids skewed training set distribution, preventing the gating network from overfitting or routing biased to a very small number of extremely popular news items during the learning phase. In addition, when assembling training batches, the system uses a time-series sliding sampling mechanism to extract samples, strictly following the temporal causal order to avoid the "time travel" phenomenon of exposing future information to historical data in advance, and to ensure that the gated network learns the real dynamic routing strategy.

[0095] To ensure that the routing strategy achieves Pareto optimality between "efficiency" and "effectiveness," this invention introduces a joint loss function during the training phase of the time-sensitive gated network. :

[0096]

[0097] in, Cross-entropy loss for predicting click-through rate; Set a preset upper limit for heavy expert usage (e.g., 40%). For indicator functions; The optimal range for the value of the balance coefficient is [missing value]. Through this constraint, the gating network can adaptively learn the optimal traffic splitting strategy under hardware capacity limitations.

[0098] To address the technical challenges of expert network freezing and gradient truncation caused by hard routing (argmax) operations, this system employs a multiplicative differentiable mechanism based on softmax probabilities. Specifically, during forward propagation, the system activates only a single target expert network based on the hard routing decision; however, when constructing the computation graph and outputting the final prediction score, the system performs a scalar multiplication of the output value of this target expert network (treated as a constant node) with the softmax routing probability value of the corresponding branch of the gating network.

[0099] During the backpropagation phase, since the expert network parameters have been frozen, the gradient of the joint loss function will directly pass through the output node of the expert network and be losslessly backpropagated to the time-sensitive gating network. This allows the gating network to adaptively learn the optimal feature splitting strategy that satisfies the computing power constraints with extremely low computational resource cost.

[0100] Step 5: Optimization of speculative execution during the inference phase

[0101] Considering that the computational complexity of lightweight expert networks (such as the dual-tower model) is much lower than that of heavyweight expert networks, this embodiment introduces an asynchronous speculative execution mechanism to pursue the ultimate low latency in inference. Specifically, while starting the gating network computation in step 2, the system simultaneously starts a speculative inference link, directly calling the lightweight expert network to score candidate news. Since lightweight expert computation is extremely fast, the speculative prediction score is usually completed when the gating decision is produced. At this time, if the hard routing decision produced in step 2 points to the "lightweight expert network", the system does not need to call the model again, but directly retrieves the speculative prediction score result from memory and returns it; if the decision points to the "heavyweight expert network", the system discards the speculative prediction score and starts the heavyweight expert inference normally. This strategy utilizes the multi-threaded parallel capabilities of the hardware to further reduce the end-to-end latency of the lightweight routing branch.

[0102] Step 6: Degradation and Disaster Recovery Mechanisms for Feature Missing Issues

[0103] In real-world distributed systems, delays in the upstream real-time computing engine may occur, leading to failures in acquiring reach rate change features. To address such anomalies, this system is configured with a disaster recovery mechanism: when the monitoring detects missing or empty timeliness features, the gated network will be physically bypassed, and the system will by default force-route affected candidate news to the lightweight expert network. This degradation mechanism ensures that even under extreme system failures, the recommendation chain still possesses basic distribution capabilities, preventing a service cascade.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the foregoing embodiments have described the present invention in detail, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the embodiments of the present invention.

[0105] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A timeliness-aware hybrid expert routing method for breaking news distribution, characterized in that, include: Step 1: Obtain the candidate news set output by the upstream coarse ranking module, and extract the timeliness features of each candidate news item in the candidate news set; wherein, the timeliness features include at least the duration after publication and the change in reach rate. Step 2: Input the timeliness features into a pre-trained timeliness gating network and output a hard routing decision for candidate news; wherein, the hard routing decision is used to instruct the target expert network to uniquely assign the candidate news to a target expert network among multiple heterogeneous expert networks; the multiple heterogeneous expert networks include at least lightweight expert networks and heavy expert networks with different computational complexities. Step 3: Based on the hard routing decision, distribute candidate news to the corresponding target expert network; Step 4: Call the target expert network to perform forward reasoning on the candidate news, output the predicted score, and pass the predicted score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy. The time-sensitive gating network is trained end-to-end based on a joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource overhead.

2. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, In step 1, the timeliness characteristic is specifically as follows: The duration after publication is ,in, This is the hourly difference between the time the news was published and the time of the current request. The reach rate change feature is the first derivative of the reach rate of candidate news within a preset sliding time window.

3. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, The timeliness feature also includes the content semantic vector of the candidate news, which is a high-dimensional dense feature obtained by encoding the textual metadata of the news through a pre-trained deep semantic model.

4. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, In step 2, the hard routing decision is specifically as follows: ； in, For the selected target expert network index, For the first The probability value output by the timeliness gating network of an expert network. ,in, To determine the number of expert networks, through The algorithm forces candidate news to activate only the single expert network path.

5. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, In step 2, the lightweight expert network is a dual-tower model structure based on vector dot product or cosine similarity; the heavy expert network is a deep interactive model with feature cross layers or attention mechanisms, and the heavy expert network has a higher number of parameters and computational complexity than the lightweight expert network.

6. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, The complexity constraint loss term is specifically set as follows: ； in, This represents the proportion of samples in the current training batch that are routed to the heavyweight expert network. The maximum percentage of heavy-duty experts is set as preset. This is a preset penalty coefficient.

7. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, Step 1 also includes smoothing for the timeliness characteristics of breaking news: When the time elapsed since the publication of a candidate news item is less than a preset time threshold, the initial characteristics of the candidate news item are filled or weighted smoothed using the average reach rate change characteristics of existing news items in the same category.

8. The timeliness-aware hybrid expert routing method for breaking news distribution as described in claim 1, characterized in that, Steps 2 through 4 also include asynchronous speculative execution optimization: While the timeliness-gated network calculates the routing probabilities of each expert network, the lightweight expert network is asynchronously invoked to perform inference on candidate news and obtain the inference prediction score; if the hard routing decision points to the lightweight expert network, the inference prediction score is directly used as the final result. Among them, asynchronous speculative execution optimization is only executed when preset triggering conditions are met; The preset triggering conditions are: the duration after the candidate news is published is less than the first preset time threshold, or the duration after the candidate news is published is greater than the second preset time threshold and its reach rate change characteristics are lower than the preset activity threshold.

9. A timeliness-aware hybrid expert routing system for breaking news distribution, characterized in that, include: A timeliness feature extraction module for extracting post-release duration and reach characteristics, a timeliness gating module for outputting hard routing decisions, a heterogeneous expert network module including lightweight expert networks and heavy expert networks, and a prediction scoring module for outputting estimated scores. in, The timeliness feature extraction module is used to obtain the candidate news set output by the upstream coarse ranking module and extract the timeliness features of each candidate news in the candidate news set. The timeliness features include at least the duration after publication and the change in reach. The timeliness gating module is used to input timeliness features into a pre-trained timeliness gating network and output a hard routing decision for candidate news; wherein, the hard routing decision is used to instruct the target expert network to uniquely assign the candidate news to multiple heterogeneous expert networks. The heterogeneous expert routing module is used to distribute candidate news to the corresponding target expert network based on hard routing decisions. The multiple heterogeneous expert networks include at least lightweight expert networks and heavyweight expert networks with different computational complexities. The prediction scoring module is used to call the target expert network to perform forward reasoning on the candidate news, output the prediction score, and pass the prediction score to the downstream reordering and distribution module for it to execute the final recommendation list generation strategy. The model training module is used to train the time-sensitive gated network based on a joint loss function, which includes a prediction loss term to ensure prediction accuracy and a computational complexity constraint loss term to control computational resource overhead.

10. A timeliness-aware hybrid expert routing system for breaking news distribution as described in claim 9, characterized in that, Also includes: The feature smoothing submodule is used to fill or weight smooth the initial features of the candidate news by using the average reach rate change feature of the existing news in the category when the time after the candidate news is published is less than a preset time threshold. The speculative execution submodule is used to start the inference link of the lightweight expert network in parallel while the timeliness gating module is operating, and decide whether to directly adopt the output of speculative inference based on the routing decision results.

Images