Short video recommendation method and system

Abstract

The invention relates to the technical field of computers, and discloses a short video recommendation method and system. The short video recommendation method comprises the steps of performing candidate recall in an index database based on a query vector to obtain a candidate set; executing sparse gating on the content factor vector under the constraint of gating budget to generate a fusion vector; calculating an energy type score based on the user state and the fusion vector, sorting the candidate sets, performing exposure offset correction, and outputting a short video recommendation result; and restraining the evolution direction of the continuous time user state according to the energy type scoring gradient. The method can effectively correct the exposure bias caused by the display position and the flow strategy, and solves the problem of bias mismatch between training and online, so that the sorting result of the candidate set is more reasonable, the purpose of improving the short video recommendation effect is achieved, meanwhile, the user intention drift is reduced, and the user experience is improved. Therefore, the purpose of short video recommendation more conforming to the long-term behavior or interest change of the user is achieved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a short video recommendation method and system. Background Technology

[0002] Short videos, with their fragmented and entertaining characteristics, have become a mainstream information dissemination medium. The core requirement of their recommendation systems is to accurately capture users' dynamic preferences and quickly match them with high-quality content. However, the unique nature of the short video scenario presents multiple challenges to recommendation technology:

[0003] First, user behavior log data is affected by display position (such as the top and bottom of the homepage) and platform traffic strategy (such as prioritizing the exposure of popular content), resulting in serious exposure bias. The log data used in the offline training phase differs from the actual traffic distribution and interface environment after online deployment, causing the trained model to mismatch with the actual application scenario, further reducing the recommendation effect.

[0004] Secondly, online recommendations need to meet millisecond-level latency requirements. Traditional recommendation solutions mostly rely on offline statistical features or single-path similarity calculations (such as collaborative filtering based solely on user click history). They are ineffective in scenarios such as cold starts (new users or new content with no historical data), sudden changes in user interests (such as users suddenly paying attention to content in a new field), and complex interfaces (multiple modules coexisting). User intent drifts, leading to deviations in the recommended short videos. Summary of the Invention

[0005] To address at least some of the aforementioned technical problems, this invention provides a short video recommendation method and system that effectively corrects exposure bias caused by display location and traffic strategies, solves the bias mismatch problem between training and deployment, thereby making the ranking results of the candidate set more reasonable, achieving the goal of improving the short video recommendation effect, while reducing user intent drift, and achieving the goal of short video recommendations that are more in line with users' long-term behavior or interest changes.

[0006] This invention provides a short video recommendation method, including:

[0007] Received short video content data and user behavior log data;

[0008] Based on the short video content data, video features, audio features and text features are extracted, and a content factor vector is generated through factorized projection, wherein the content factor vector includes semantic factors.

[0009] Based on environmental tags, the semantic factors are distributed and aligned to generate aligned semantic factors, and the user behavior log data is organized into a user event stream.

[0010] Based on the user event stream, a continuous-time user state is constructed and a query vector is generated to determine the gating budget;

[0011] Based on the query vector, candidate retrieval is performed in an index database built from aligned semantic factors to obtain a candidate set;

[0012] Sparse gating is performed on the content factor vector under the constraint of the gating budget to generate a fusion vector;

[0013] Based on the continuous time user state and the fusion vector, an energy-based score is calculated and the candidate set is sorted. Exposure bias correction is performed when sorting the candidate set, and short video recommendation results are output.

[0014] The evolution direction of the continuous-time user state is constrained by the gradient of the energy-based scoring.

[0015] This invention provides a short video recommendation system, applicable to the short video recommendation method provided in this invention, comprising:

[0016] The acquisition and alignment module is used to receive short video content data and user behavior log data; based on the short video content data, it extracts video features, audio features and text features, and generates a content factor vector through factorized projection, wherein the content factor vector includes semantic factors; it performs distribution alignment on the semantic factors based on environmental labels to generate aligned semantic factors, and organizes the user behavior log data into a user event stream;

[0017] The state fusion module is used to construct continuous-time user states based on the user event stream and generate query vectors, and determine the gating budget; based on the query vectors, perform candidate recall in an index library built from aligned semantic factors to obtain a candidate set; and perform sparse gating on the content factor vectors under the constraints of the gating budget to generate a fusion vector.

[0018] The scoring and ranking module is used to calculate an energy-based score based on the continuous-time user state and the fusion vector, and to rank the candidate set. During the ranking of the candidate set, exposure bias correction is performed, and short video recommendation results are output. The evolution direction of the continuous-time user state is constrained according to the gradient of the energy-based score.

[0019] The embodiments of the present invention have the following technical effects:

[0020] Exposure bias correction enhances model robustness: Exposure bias correction is performed when sorting the candidate set, which effectively corrects the exposure bias caused by display position and traffic strategy, solves the bias mismatch problem between training and deployment, improves the robustness of the recommendation system, and achieves exposure bias correction and fairer ranking results.

[0021] Continuous-time user state modeling to adapt to real-time intent: By modeling the continuous-time user state based on user event streams, real-time response to changes in interest and conversation rhythm is achieved, improving the timeliness of recommendations.

[0022] User state evolution direction constraint reduces intent drift: By using a scoring gradient to constrain the consistency of user state direction, the training objective and online evolution are integrated, reducing intent drift and improving recommendation performance in cold start and interest mutation scenarios. After outputting short video recommendation results, it ensures that user states (such as interests and needs) evolve smoothly and reasonably over time, achieving the goal of short video recommendations that better match users' long-term behavior or interest changes. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a short video recommendation method provided in an embodiment of the present invention;

[0025] Figure 2 This is a structural block diagram of a short video recommendation system provided in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] Figure 1 This is a flowchart illustrating a short video recommendation method provided in an embodiment of the present invention. Figure 1 As shown, the short video recommendation method includes steps S110 to S180:

[0029] S110: Receive short video content data and user behavior log data.

[0030] Each user behavior log entry should include user ID, video ID, timestamp, behavior type, etc., forming a traceable "user-video" event.

[0031] S120. Based on short video content data, extract video features, audio features, and text features, and generate content factor vectors through factorized projection, wherein the content factor vectors include semantic factors.

[0032] S130. Based on the environment label, the semantic factors are distributed and aligned to generate aligned semantic factors, and the user behavior log data is organized into a user event stream.

[0033] Perform reception and consistency checks on short video content data and user behavior log data. The server receives video files, audio tracks, and text information based on the video's unique identifier, and receives user behavior log data based on the user's unique identifier. Then, the video files are decoded, representative frame sequences from several time slices are extracted, the audio tracks are uniformly sampled and amplitude normalized, and the text information is segmented and cleaned. User behavior log data is deduplicated in ascending chronological order, retaining key events such as exposure, clicks, dwell time, fast forward, rewatch, and sharing, and recording the event timestamps and duration or intensity.

[0034] Multimodal feature extraction is performed. On the visual side, video features are extracted using a convolutional network or a temporal visual encoder; on the audio side, audio features are extracted using an acoustic encoder; and on the text side, text features are extracted using a pre-trained language model. To ensure cross-modal temporal consistency, representative frame time slices are aligned with audio segments and subtitle time slices on the same timeline to avoid subsequent semantic confusion. The output of feature extraction is a fixed-length vector for easy unified projection.

[0035] Factorization projection is performed. Video, audio, and text features are concatenated and input into the factorization projection network to obtain content factor vectors. These vectors contain four sub-vectors: semantic factors, style factors, creator factors, and social factors. Semantic factors carry stable semantics such as themes and actions; style factors carry expressive styles such as filters, layouts, and background music; creator factors carry stable creator preferences and common themes; and social factors carry historical interaction strength and weak social proof signals. To reduce coupling between sub-factors, the projection network adopts a shared trunk and branch head structure, and decorrelation regularization and adversarial discriminators are introduced during training to make it difficult for semantic factors to be identified by style or creator origin.

[0036] The core computational expression for the above steps is factorization mapping:

[0037]

[0038] For content factor vectors, For factorized projection functions, For video feature vectors, For audio feature vectors, For text feature vectors, For semantic factors, For style factors, For the creator factor, Social factors.

[0039] Environment tags are constructed and distribution alignment is performed. Environment tags indicate external factors influencing the distribution, i.e., the environment, such as creator domain, posting time period, topic domain, and traffic level. Environment tags are discrete combinations and can be represented by enumeration, multi-hot encoding, embedding vectors, etc. To reduce the interference of environmental differences on semantics, semantic factors in different environments are mapped to a reference environment to obtain aligned semantic factors. The reference environment is a preset baseline prototype (such as the global main environment or the maximum coverage environment), not a single external factor. Alignment adopts a distribution alignment method with entropy regularization (see below for details).

[0040] User behavior log data is organized into user event streams. For each user, events such as exposure, clicks, dwell time, fast-forwarding, rewatching, and sharing are organized into time-ordered triple sequences, each containing event time, event type, and intensity. Intensity values ​​are mapped by business rules; for example, dwell time is converted to a logarithmic scale, and fast-forwarding is calculated based on the skip ratio. The user event stream is stored in a randomly accessible data structure, supporting both offline training data loading and incremental updates for online inference.

[0041] In terms of system architecture, this step can be implemented collaboratively in the offline feature pipeline and the online service. The offline pipeline completes video decoding, modal feature extraction, factorization projection, environment alignment, and event normalization, and periodically produces content factor vectors, aligned semantic factors, and snapshots of the user event stream. When the online service receives a request (online recommendation call (entry / refresh / page turning); carrying user ID, environment tag, session context, and page number / quantity), it reads the corresponding content factor vector and aligned semantic factors, and simultaneously supplements the recent session events from the real-time log (appended to the end of the user event stream for state updates), forming a complete input for subsequent steps.

[0042] For example, a short video dataset is selected, and video decoding and sampling are performed to extract video, audio, and text features. Content factor vectors are obtained through factorized projection, and a distribution alignment task is constructed using context labels to produce aligned semantic factors. User behavior logs are aggregated and normalized into a time-ordered user event stream. Validation results show that, under the same recall and ranking models, using aligned semantic factors can improve the candidate hit rate across topic time periods, and using normalized user event streams can improve the identifiability of user states over continuous time. This embodiment illustrates that the principle and implementation method of this step can be implemented in an engineering system and has a direct benefit to subsequent modeling.

[0043] Specifically, in some implementations, based on short video content data, video features, audio features, and text features are extracted and content factor vectors are generated through factorized projection. This includes: obtaining multimodal features based on short video content data through a visual encoder, an audio encoder, and a text encoder, respectively; mapping the multimodal features into content factor vectors through linear projection and a branching structure. The content factor vectors also include style factors, creator factors, and social factors.

[0044] Data access and preprocessing follow a unified standard. Video data is decoded to obtain frame sequences, and several segments are uniformly extracted along the timeline. Each segment is cropped to a fixed size and pixel normalized. Audio data is resampled to a uniform sampling rate, and the Mel spectrum is extracted and amplitude normalized. Text data undergoes Chinese subtotaling and cleaning, with unified capitalization and punctuation, and removal of invalid markers. Unified capitalization and punctuation refer to the standardization of Chinese punctuation and the unification of full-width and half-width characters; invalid markers include, but are not limited to, blank tags, garbled / invisible characters, excessively long repeating characters, and low-confidence ASR segments. To ensure semantic consistency, video segments, audio segments, and subtitles are aligned on the same timeline, and missing modalities are replaced with zero vectors or learnable placeholders. A missing modality refers to a sample lacking one of the following: audio / subtitle / image. This is caused by silence, lack of subtitles, decoding failure, copyright obscuration, or abnormal frame extraction.

[0045] Multimodal feature extraction utilizes mature encoders. The visual encoder extracts spatiotemporal features from video segments, outputting a fixed-length vector. The audio encoder extracts temporal acoustic features from the Mel spectrum, outputting a fixed-length vector. The text encoder takes a sequence of words as input and extracts sentence-level representations as fixed-length vectors. The encoders can be pre-trained offline and fine-tuned on this task to balance representational power and online efficiency.

[0046] Factorized projection employs a shared trunk plus branch structure. The shared trunk fuses features from three modalities, outputting a fused representation; the four branches correspond to semantic factors, style factors, creator factors, and social factors, respectively. During training, the semantic branch learns stable semantic representations through weak supervision of topics or labels; the style branch uses statistical features of visual style, color scheme, or audio rhythm as auxiliary signals; the creator branch uses aggregated features of creator identifiers as constraints, emphasizing consistency across works (stable representations of the same creator across videos (convergence of creator factor aggregation)); the social branch compresses historical interaction statistics to avoid scale bias caused by directly using raw counts. To reduce information leakage between branches, decorrelation regularization and adversarial discriminators are set between the shared trunk output and each branch to suppress semantic information spillover to the style or creator branches, thereby improving the interpretability of each branch.

[0047] The core computational expression in this step is the factorization mapping: The training process consists of three parts. First, cross-modal alignment pre-training uses a contrastive objective (positive sample pairs from the same video are brought closer together, while negative sample pairs from different videos are brought further apart, achieving alignment in a shared semantic space) to bring video and text, and audio and text, closer together in the same semantic space, thereby enhancing semantic consistency. Second, factorized supervised training, where each of the four branches is connected to a corresponding weakly supervised signal (e.g., the semantic branch uses topics / tags; the style branch uses visual style clusters or color statistics; the creator branch uses author IDs; and the social branch uses interactive graph community IDs) to optimize classification or regression objectives (classification predicts topics / authors / style clusters; regression predicts continuous indicators such as completion rate, dwell time, and sharing rate), while applying decorrelation regularization within the batch to improve branch independence. Third, online distillation and delayed updates (online, the teacher model uses soft labels to calibrate students, and parameters are updated in small steps at low frequency to avoid jitter), offline periodic refresh of factorized projection parameters, and online only forward computation (only encoder and projection, gating fusion, and scoring are performed, without backward gradient and weight updates) to ensure stable service latency.

[0048] In terms of system architecture, the offline pipeline completes decoding, time alignment, multimodal feature extraction, and factorized training, producing factorized projection parameters and encoder weights. The online service loads these weights, performs forward inference on newly added videos, generates content factor vectors, and writes them to the feature storage for direct reading during recall and ranking. To facilitate backtracking and hot updates, the content factor vectors are managed using version numbers and timestamps, and rollback to the most recent stable version is performed when abnormal drift is detected.

[0049] Quality control employs a combination of automated and manual methods. The automated system calculates the correlation coefficient between branches for each batch of output (Pearson correlation (or cosine similarity mean) of the output vectors of two branches within the same batch of samples). If the coefficient exceeds a threshold (indicating branch information leakage / insufficient decoupling, requiring the addition of correlation regularization or retraining), an alarm is triggered. The system also calculates the deviation between the semantic factors of new videos and the mean vectors of similar videos (primarily based on semantic tagging systems (topics / domains), combined with semantic clusters obtained through clustering; similar videos are equivalent to those with the same tag or cluster). Excessive deviation triggers a recalculation. The manual system periodically checks interpretable examples of semantic and style factors to verify the clarity of branch division of labor.

[0050] For example, in actual business data, 100,000 short videos are selected as the training set and 10,000 short videos as the validation set. After preprocessing and time alignment, a multimodal encoder is trained and factorized projection is performed. After going live, feature extraction and factorization mapping are performed on newly added videos to generate content factor vectors, which are then written into the feature library. Compared with the single fusion representation without factorization, the system using content factor vectors can more accurately retrieve based on semantic nearest neighbors in the downstream recall stage, and can enable or suppress style and social information as needed in the ranking stage, thereby improving relevance stability and cross-scenario generalization. The above embodiments show that the technical means of this step are clear, feasible, and can achieve the expected results in the engineering system.

[0051] In some implementations, semantic factors are distributed and aligned based on environment tags to generate aligned semantic factors, and user behavior log data is organized into user event streams. This includes: using a distributed mapping method to align semantic factors between different environments corresponding to environment tags to obtain aligned semantic factors; and grouping user behavior log data into user event streams that include exposure, clicks, dwell time, fast forward, rewatch, and sharing in chronological order.

[0052] The alignment process consists of four stages: environment segmentation, reference environment determination, distribution mapping solution, and alignment vector generation. In the environment segmentation stage, semantic factors are divided into several subsets based on environment labels. Subsets with insufficient sample size and high noise levels are removed, or merged into adjacent subsets. In the reference environment determination stage, reference environments are selected from subsets with stable sample size and balanced topic distribution, avoiding subsets with abnormal traffic or extreme topics as references. In the distribution mapping solution stage, the mapping weights from different environments to the reference environment are calculated, ensuring that samples from two sets are paired and close in the semantic space. To facilitate engineering implementation and ensure stability, this invention employs an alignment objective with entropy regularization to obtain the transmission weights, which are then used to generate the aligned semantic factors. The core calculation expression is as follows:

[0053] Solving for the optimal transmission weight, assuming the source environment is... The reference environment is The set of semantic factors of the source environment is as follows:

[0054]

[0055] The reference environmental semantic factor set is as follows:

[0056]

[0057] Then the optimal transmission weight matrix It can be obtained from the following formula:

[0058]

[0059] Aligned semantic factor generation (baric projection). For the source environment... Given 1 semantic factor, generate its aligned semantic factor:

[0060]

[0061] Source environment sample Aligned semantic factors; Source Environment The first in One semantic factor; Reference Environment The first in One semantic factor; Source Environment The corresponding set of semantic factors; Reference Environment The corresponding set of semantic factors, Source Environment The Middle semantic factors To reference environment The Middle semantic factors The assigned transmission weights, where , ; : Optimal transmission weight matrix; Optimal transmission weight matrix The first in One element; Entropy regularity coefficient; : L2 norm; Natural logarithm; Source Environment semantic factor set The number of elements; Reference Environment semantic factor set The number of elements.

[0062] To ensure feasibility, the solution process employs an iterative update approach, limiting the number of iterations and numerical precision thresholds to control computational costs. Small batches of samples are constructed for each environment and the reference environment, and the transfer matrix and weights are calculated. After obtaining the transfer weights, centroid projection is performed to generate aligned semantic factors. The aligned product and the original semantic factors are stored together, managed using version numbers for easy rollback and online consistency checks. Alignment quality is monitored using two types of metrics: the reduction ratio of the mean difference and covariance difference between environments, and the nearest neighbor stability during cross-environment retrieval. If either metric falls below a threshold, a rollback to the previous version is initiated, and an alarm is recorded.

[0063] Among them, the mean / covariance difference reduction ratio: comparison before and after alignment. The percentage decrease, of which To align with the source environment semantic factor set The mean vector, To align with the source environment semantic factor set The covariance matrix, To align the source environment The mean vector of the set of aligned semantic factors. To align the source environment The covariance matrix of the set of aligned semantic factors. Represents the L2 norm, Represents the Frobenius norm; Nearest Neighbor Stability: The Jaccard / overlap rate of the same query retrieving TopKKK in different environments. If it falls below the threshold, a rollback is triggered.

[0064] The purpose of organizing user behavior log data into user event streams is to unify behavior records from different sources onto the same timeline and intensity scale for subsequent use in analyzing user status over continuous time. Specific steps include time cleaning, event extraction, intensity mapping, and sequence generation. Time cleaning converts all timestamps to a unified time zone and removes duplicate records. Event extraction retains only six key event categories: exposure, click, dwell time, fast forward, rewatch, and share. Other auxiliary fields are used for subsequent analysis but are not included in the event stream. Intensity mapping generates standardized intensities for each event: dwell time is mapped based on viewing duration, fast forward time based on skip ratio, and rewatch time based on repetition count. Exposure, click, and share use unit intensity (normalizing discrete events to a scalar with a unified dimension; clicks / shares are counted as 1; exposures are measured using "effective exposures"). Sequence generation arranges events in ascending chronological order, adds session boundary markers and content identifiers, and forms a read-only log partitioned by user, supporting incremental writing and random access.

[0065] In terms of system architecture, alignment and normalization are handled in the offline computing chain. Offline tasks scan newly added semantic factors and new behavior logs daily, generating aligned semantic factors and user event stream snapshots, which are then written to feature storage. Online services only read the aligned semantic factors during recall and sorting, without participating in alignment calculations, thus ensuring stable request latency. Event streams are incrementally merged into the daily snapshot after a user generates a new behavior, allowing them to be read by subsequent steps in the next request.

[0066] The key concepts introduced in the above steps of this invention are environment tags and distribution mapping. Environment tags consist of engineering-obtainable metadata, avoiding reliance on difficult-to-observe latent variables. Distribution mapping aligns the environment to a reference environment through a weight matrix, and centroid projection ensures that the aligned semantic factors remain within the content space, facilitating direct use by the vector indexing and ranking modules. The effects of these techniques are improved transferability of cross-environment retrieval results, reduced sensitivity to environmental changes during the ranking stage, and a fairer recall opportunity for cold-start videos and videos in long-tail environments after alignment.

[0067] For example, newly added video data from the past week is selected, and environmental tags are constructed based on creator domains and publication time periods. After extracting semantic factors, alignment is performed. Weekday daytime samples are used as the reference environment for alignment of other time periods. After alignment, cross-time period nearest neighbor retrieval for the same topic term shows improved nearest neighbor consistency (the set of similar videos retrieved for the same topic term at different time periods is more consistent, specifically manifested in a higher overlap rate of the first few nearest neighbor results and smaller ranking fluctuations), and reduced score fluctuations in cross-time period rankings. Behavioral logs are grouped into event streams by user, and continuous user status can promptly reflect interest shifts after reading the event stream. This embodiment demonstrates that the distribution alignment and event stream regularization in this step can be implemented in real business scenarios and generate stable benefits.

[0068] S140. Construct continuous-time user states based on user event streams and generate query vectors to determine the gating budget.

[0069] S150. Based on the query vector, perform candidate recall in the index database built from aligned semantic factors to obtain a candidate set.

[0070] S160. Perform sparse gating on the content factor vector under the constraint of the gating budget to generate the fused vector.

[0071] The user event stream consists of exposure, click, dwell, fast forward, rewatch, and share. On the online platform, the most recent session window is scrolled and aggregated at fixed time steps. Each event is mapped to a sparse vector and superimposed with time decay to obtain an event-driven signal, which is then input into the continuous-time user state model. For example, multiple events (exposure, click, dwell, fast forward, rewatch, and share) within the user's most recent session window are taken. Each event is first converted to a uniform intensity and then decayed and accumulated according to the principle of "the more recent, the greater the weight," resulting in an "event-driven signal." This event-driven signal is then used to update the user's continuous interest state, allowing it to quickly shift towards new interests based on recent interactions. To balance feasibility and stability, a discretized continuous-time update is adopted. The core calculation expression is:

[0072]

[0073] A continuous-time user state vector. For time step, For state update functions, This is an event-driven signal. The state update function can use either a feedforward network with residuals or a gated recurrent unit, which is computed in batches by the online inference engine to ensure controllable request latency.

[0074] The query vector is obtained by linearly mapping continuous-time user states and is used to retrieve the candidate set in the semantic space. The core calculation expression is: , For query vector, The query mapping matrix is ​​used. Offline, a vector index library is built using aligned semantic factors. Inner product or cosine similarity is used as the retrieval metric. An inverted index and product quantization encoding are constructed using an approximate nearest neighbor library. Online, a candidate set is obtained by retrieving the query vector, and recently overexposed content is removed according to business rules.

[0075] The gating budget refers to the upper limit of the number / combination of factors allowed to be activated in a single request; it is generated by mapping online latency targets, concurrency pressure, user segmentation (cold start / stable / topic switching), and business strategies, and is released daily with each version update and is rollback-capable. The gating budget controls the number of factors that can be activated during this round of fusion, avoiding overfitting on non-critical modalities. The gating budget is obtained by mapping continuous-time user states, specifically based on statistics such as user topic diversity, recent interest switching, and session stability. The mapping function is a learnable monotonic function, outputting non-negative integers. Subsequently, factor-level gating is performed on each content in the candidate set. The core calculation expression is:

[0076]

[0077]

[0078] For the first The gating score of each factor. For logical functions, For gating weights, Connect vectors. The first content factor vector One portion, This is the fusion vector. To satisfy the gating budget, the gating score is obtained through the projection operator. The implementation involves selecting the first few factors equal to the gating budget in descending order of scores across all factors and setting them to one, while setting the rest to zero. Alternatively, a temperature-controlled hardening approximation can be used to facilitate end-to-end training. The fused vector is written to the feature bus for the scoring and ranking modules to read.

[0079] The system architecture comprises three clear data pathways. The event pathway subscribes to user events in real time from the log platform and generates event-driven signals. The state pathway calls the continuous-time user state model to generate query vectors and gating budgets. The content pathway reads aligned semantic factors from the feature storage. The retrieval service accesses the vector index library with the query vectors, returns a candidate set, and then enters the gating service along with the gating budget. The gating service produces a fused vector and writes it back to the cache for easy sharing in subsequent multi-way sorting.

[0080] To ensure feasibility and maintainability, the state model and gating model are deployed in a canary release manner. Only the forward graph is loaded online (only the inference computation path (encoding / projection / gating / fusion / scoring) is loaded online, excluding backward gradients and parameter updates). The query mapping matrix and gating weights are updated daily with version numbers, allowing for rollback in case of anomalies (parameters are deployed in a canary release with version numbers, and rollback is possible in case of anomalies). The retrieval side is configured with multiple replica indexes for different user groups to support high concurrency. The gating service and retrieval service are decoupled through a memory queue to implement timeout isolation and degradation strategies (gating and retrieval are isolated by a queue, so a slowdown on either side will not disrupt the entire chain, and timeout degradation is possible).

[0081] For example, taking a group of highly active users as the target, the user event stream of the most recent session is input into the continuous-time user state model to obtain the continuous-time user state and query vector, and a candidate set is obtained through online retrieval. Factor-level gating is performed on the content factor vector of each candidate, and a fusion vector is generated under the gating budget constraint. Then, the ranking module completes the scoring and output (for each candidate, the gating is first calculated (to determine which factors are enabled), and then the enabled factors are weighted and fused into a content representation, and a similarity score is calculated with the user query representation. After adding the necessary debiasing and timeliness correction, the data is ranked). Compared to full fusion without gating budget, the gating and fusion approach implemented in this step results in higher relevance stability when switching topics within the same session (recommendation results are less prone to significant fluctuations when a user switches from one topic to another during continuous browsing; they still keep up with new topics while reducing "fluctuations" caused by style or social noise). Sensitivity to noise factors such as short-term social spikes is reduced, leading to improved conversion from recall to ranking (for example, if a user frequently stays in and revisits "food" in recent sessions, event-driven signals shift their interest towards food; when the gating budget only allows two types of factors, the system prioritizes "semantic + style" and suppresses social noise, thus resulting in more stable ranking when switching topics within the same session). The above examples verify the operability and benefits of this step in a real-world system.

[0082] In some implementations, constructing a continuous-time user state based on the user event stream and generating a query vector to determine the gating budget includes: representing the user event stream as a time-driven signal; modeling the continuous-time user state using a continuous-time model based on the time-driven signal; performing a state update based on aligned semantic factors when fast-forwarding, rewatching, or long-term dwell is detected; generating a query vector from the continuous-time user state and determining the gating budget based on control rules.

[0083] Preferably, constructing continuous-time user states and generating query vectors based on user event streams includes: representing user event streams as time-driven signals; modeling continuous-time user states using a continuous-time model (converting exposure, clicks, dwell time, fast-forwarding, rewatching, and sharing within a session into "statistical features with time decay," inputting into a continuous-time model, and allowing the continuous-time model to continuously output a "current interest profile"); when fast-forwarding, rewatching, or prolonged dwell time is detected, performing state updates based on aligned semantic factors (updating to the user's current interest profile and query vector; the purpose is to allow the profile to change with recent behavior, reflect interest migration in a timely manner, and avoid using expired preferences for recommendations); generating query vectors from continuous-time user states; and determining the gating budget based on control rules.

[0084] The construction of the user event stream follows the steps outlined above. This step first maps the event stream to time-driven signals. In the engineering implementation, recent sessions are aggregated using a sliding window with a fixed time step. Each event is mapped to a sparse vector based on its type and intensity, and time decay is superimposed to form the driving input for the current moment. For example, if a user repeatedly pauses and rewatches a "food" video or frequently fast-forwards a "game" video within the recent window, the system accumulates pauses / rewatches as positive intensity and fast-forwards as negative intensity, and reduces the weight of earlier events to obtain the driving input of "increased food preference and decreased game preference." The driving input retains only statistics related to individual short-term interests, avoiding the introduction of redundant features unrelated to ranking. The driving input retains only information directly related to short-term preferences, such as "behavior type + intensity + time decay," and does not mix non-ranking key fields such as device model, network status, and log noise into the state update, thereby reducing interference from irrelevant features.

[0085] The continuous-time user state is updated using discretization to meet online latency constraints. The core calculation expression is:

[0086]

[0087] A continuous-time user state vector. For time step, For state update functions, The signal is time-driven. The state update function can be implemented by a feedforward network with residuals or a gated recurrent unit. The parameters are loaded into the online inference engine after offline training, and only forward computation is performed to ensure stable service latency.

[0088] To address sudden shifts in interest, this step injects state upon detecting a strong trigger event. Strong trigger events include fast-forwarding, rewatching, and prolonged viewing. Fast-forwarding indicates a rapid rejection of the current content, while rewatching and prolonged viewing indicate a strong preference for it. The core calculation expression for state injection is:

[0089]

[0090] For the moment when a strong trigger event occurs, For injection coefficient, For semantic injection matrix, The aligned semantic factors of the short video that triggered the event. and These represent the user states before and after injection, respectively, over consecutive time. Injection coefficients are calculated based on event type and intensity: smaller or negative injections are applied to fast-forwarding, while larger positive injections are applied to rewatching and prolonged periods of time spent on the page. The semantic injection matrix is ​​trained offline by minimizing the prediction error of subsequent clicks within the session, ensuring that the user states after injection more closely reflect real interests.

[0091] The query vector is obtained by linearly mapping continuous-time user states and is used to enter the semantic recall space. The core calculation expression is: , For query vector, The query mapping matrix is ​​used. The online retrieval service accesses a vector index library built on aligned semantic factors using query vectors, and returns a candidate set using inner product or cosine similarity. To avoid feedback bias caused by overexposure, the retrieval results undergo simple deduplication and frequency restriction before being sorted (first, a query vector is generated from the user's current interest profile; the query vector is used to perform a nearest neighbor search in the "semantic factor vector index library" to return candidates; before the candidates are sorted, deduplication (only one record is kept for the same author / video / material) and frequency control (limiting the number of times the same topic or author appears in the results) are performed, and then handed over to downstream fusion and sorting).

[0092] The gating budget is used to limit the number of factors that can be activated simultaneously during subsequent factor-level fusion, controlling the model's dependence on non-target channels such as style, creator, and social interaction. The gating budget is obtained from a continuous-time user state mapping, and its core calculation expression is:

[0093]

[0094] For the gating budget, As the baseline constant, It is a scaling constant. For logical functions, This is the budget mapping vector. In the offline phase, the budget mapping vector is trained with the objective function "minimizing the usage of non-semantic channels while ensuring ranking relevance." Online, the gating budget is read by the downstream fusion module (referring to the feature fusion and scoring module on the ranking side; it reads candidate multi-class content factors (semantic / style / creator / social) and the gating budget, selects a small number of factors to participate in the fusion, generates the final content representation used for scoring and ranking, and outputs the score). This is used to truncate or harden the gating scores of each factor channel, ensuring that only a limited number of channels relevant to the current intent are activated.

[0095] In terms of system architecture, this step is implemented collaboratively by three types of microservices. The state service is responsible for generating time-driven signals and updating user states over continuous time, and provides interfaces for query vectors and gating budgets. The retrieval service maintains aligned semantic factor indexes and provides approximate nearest neighbor retrieval. The gating service performs factor-level gating and fusion after reading the candidate set and gating budget, outputting a fused vector for scoring and ranking. Vector data is transmitted between services via a lightweight protocol. The vectors use a batch interface and support timeout degradation to ensure stability under high concurrency scenarios.

[0096] To ensure feasibility and maintainability, core parameters employ a layered version management system: state update functions, semantic injection matrices, and query mapping matrices are updated daily on a rolling basis, while gating budget mapping parameters are hot-updated hourly. The online inference engine only loads versions that have passed offline acceptance and online gray-scale verification. The monitoring side records the norm of user states over continuous time, the distribution of query vectors, and the utilization rate of gating budgets at the session granularity, triggering a rollback when abnormal drift occurs.

[0097] For example, a group of active users is selected, and a one-minute time step is set. The user event stream over the past thirty minutes is continuously updated to obtain the continuous-time user state. State injection is performed on detected review events to generate new continuous-time user states and query vectors, which are then used to obtain a candidate set for retrieval. Under gated budget constraints, only semantic and a small number of creator factors are activated, generating a fusion vector and entering the ranking process. Compared to a baseline system that does not use state injection and gated budget, this embodiment shows improvements in click-through rate and dwell time during topic mutations, while the proportion of non-semantic channels activated decreases, indicating that continuous-time user states and budget control effectively improve relevance and robustness. The above process can be directly deployed within existing recommendation engineering systems, with a clear implementation path and observable benefits.

[0098] In some implementations, candidate recall is performed based on the query vector in an index library built from aligned semantic factors, including: mapping the query vector to a semantic retrieval space; and obtaining a candidate set based on the query vector using an approximate nearest neighbor retrieval method in a vector index library built from aligned semantic factors.

[0099] In the offline phase, a semantic retrieval space and vector index library are established. First, the aligned semantic factors are standardized and whitened to remove outlier and duplicate samples. Then, a retrieval mapping and quantizer are fitted to the training set to complete the index structure. In practice, either inverted index with quantization or graph structure indexes can be used. Inverted index with quantization is suitable for large-scale, high-concurrency scenarios. The construction process includes clustering to obtain several centers, building an inverted list for each center, multiplying and quantizing the residual vectors, and storing short codes and content identifiers. Graph structure indexes are suitable for low-latency scenarios with sufficient memory and can supplement hot sets. The index library is deployed in shards and with multiple replicas, supporting bucketing by topic and region to reduce cross-bucket interference.

[0100] In the online phase, the query vector is mapped to the semantic retrieval space and similarity retrieval is performed. The core calculation expression is:

[0101]

[0102]

[0103] For the retrieval vector, To retrieve the mapping matrix, For query vector, For index vectors, For similarity, both retrieval vectors and index vectors are normalized to stabilize the units across different sessions and time periods. For inverted and quantized indexes, firstly, several closest centers are selected based on the retrieval vector, limiting the number of items in the search list (the index is first clustered / bucketed, with each bucket having a center vector; online, the similarity between the retrieval vector and each center is calculated, and the buckets with the highest similarity are selected as candidate buckets. Searching is only performed in the inverted lists corresponding to these candidate buckets, and a fixed return limit is set for each bucket; simultaneously, a global limit is set (e.g., the total number of candidates does not exceed a certain number), and items are truncated if this limit is exceeded to ensure controllable latency and cost). Then, similarity is calculated using quantization approximation in the corresponding inverted lists, and several candidate identifiers are returned. For graph-structured indexes, starting from several in-points, a bounded-width search is performed along highly similar edges, and candidates are collected within the access budget.

[0104] To ensure usability of the results, a series of engineered constraints need to be implemented on the recall results. These include deduplication based on user history, imposing frequency limits on recently exposed content, and removing samples that hit blacklists or do not meet regional and copyright constraints. For content with a cold start or missing semantic factors, a fallback channel based on the similarity between title keywords and creators is used to avoid recall interruptions.

[0105] The retrieval mapping matrix is ​​trained offline, aiming to maximize the similarity interval between clicked or dwelled samples. It is calibrated using the most recent data and undergoes gray-scale validation before deployment. Incremental index updates follow a process of adding the training set first, then the index, avoiding direct exposure of unverified anomalous vectors to the online system. Full reconstruction is performed during off-peak hours. After reconstruction, consistency metrics are compared between the new and old indexes, and traffic is switched accordingly (a small percentage of traffic is initially diverted for validation when a new index or mapping matrix is ​​deployed; if metrics are normal, the volume is gradually increased; if anomalies occur, the system reverts to the old version. This reduces release risk and ensures online stability).

[0106] In terms of system architecture, the retrieval service and the state service are decoupled. The state service outputs a query vector, and the retrieval service receives the query vector and performs an approximate nearest neighbor search in local memory or cache, returning a candidate set and the original similarity. To reduce latency, the retrieval service supports batch request aggregation and asynchronous backfilling; to improve stability, a search budget and timeout threshold are set, and if the threshold is exceeded, it falls back to lightweight rule recall.

[0107] In terms of quality and monitoring, online systems continuously monitor similarity distribution, center hit distribution, recall coverage, and candidate diversity. When distribution drift or coverage decline occurs, the system automatically downgrades by increasing the number of search lists or switching to redundant replica indexes. Offline systems use manual sampling and visual alignment to ensure consistency of neighboring samples for the same topic across different environments.

[0108] For example, in real business traffic, the retrieval mapping matrix and quantizer are updated daily, and popular content is simultaneously written to both the inverted quantized index and the graph structure index (or graph index). Each online request generates a query vector from the state service, which is then mapped to obtain a retrieval vector. Several centers are selected in the inverted and quantized indexes for searching, and the popular samples returned by the graph structure index are merged. After applying deduplication and exposure constraints, a candidate set is formed. Compared to the baseline without aligned semantic factors, the stability of nearest neighbors across time periods and across creators is improved, the coverage of subdivided topics is increased, and the overall retrieval latency remains within a predetermined threshold, providing higher-quality candidate input for subsequent fusion and ranking.

[0109] In some implementations, sparse gating is performed on the content factor vector under the constraint of a gating budget to generate a fusion vector, including: setting gating variables for semantic factors, style factors, creator factors, and social factors respectively; using a binary gating function to achieve sparse selection; and performing a weighted summation of the passed factors under the constraint of a gating budget to generate a fusion vector.

[0110] The online process is as follows: the state service outputs continuous-time user states and a gating budget; the retrieval service returns candidate content and its content factor vectors; the gating service calculates gating scores for four factors for each candidate, performs binarization selection under the gating budget constraint, and finally performs a weighted summation of the selected factors to obtain a fusion vector, which is written to the memory cache for ranking and retrieval. To ensure latency and stability, the gating calculation uses a pure forward graph, and all parameters are generated offline and updated hot-by-version.

[0111] The gating score employs a linear scoring method based on feature concatenation, mapped to a range of 0 to 1 via a logistic function. The core calculation expression is:

[0112]

[0113]

[0114]

[0115] For the first The gating score of each factor. For logical functions, For the first The scoring weights of each factor For vector concatenation, A continuous-time user state vector. The first candidate content A factor vector, For binary gated variables, This indicates selecting the highest score. One factor, For the gating budget, For the fusion vector, The fusion weights of the gated factors, The number of factors.

[0116] In practical applications, the following points should be noted: First, gating binarization uses a direct-pass approximation during the training phase, and then applies a threshold or forward pass during the forward pass. First, rounding down the nominal value and approximating the gradient with continuous scores in the reverse process balances trainability and inference consistency. Second, the fusion weights employ a normalization strategy to improve numerical stability: the gated factors are assigned according to gating scores or fixed weights. and to Unitize to prevent differences This leads to dimensional drift. Third, if the gating budget is zero or the effective scores are insufficient, it falls back to enabling only semantic factors to ensure system output. Fourth, the input to the gating score only includes the user's state over continuous time and the factor itself, avoiding the use of features from subsequent scoring to prevent information leakage. Fifth, a "strong constraint list" can be configured for different business lines, such as forcibly closing certain factor channels under copyright or brand security requirements. This allows for the forced closure or restriction of certain factor channels (e.g., social, creator) under compliance / security / business strategy requirements, avoiding the introduction of sensitive associations and facilitating unified operation and auditing.

[0117] Offline training employs a multi-task approach linked to the ranking objective: first, a ranking loss based on historical clicks or duration; second, gating sparsity regularization to ensure the average number of activation factors approaches the gating budget; and third, stability constraints to reduce the jitter of gating patterns between adjacent requests. The gating weights and fusion weights share a portion of the underlying representation with the ranking header (the shared portion only carries general semantic representations, without including style / creator / social-specific information; it also maintains consistent dimensionality and numerical scale to facilitate reuse and stable deployment of gating and ranking), reducing redundant parameters. Training data covers multiple time periods, topics, and creator scenarios to ensure the gating strategy generalizes across different environments.

[0118] In terms of system architecture, the gating service is deployed statelessly, receiving candidate sets and continuous-time user states via a message bus, calculating fusion vectors, and then sending them back in batches. To ensure high concurrency, gating score calculation, TopK selection, and fusion weight assignment are split into three operators, each vectorized. When the candidate set is large, truncation is prioritized at the score stage to reduce the load on subsequent operators. The monitoring side tracks the average budget utilization, the activation ratio of each factor, and the norm distribution of the fusion vector for each session, and triggers rollback for abnormal drift.

[0119] For example, in a real-world online gray-scale test, a user group with relatively stable user states over a continuous period is selected, and the gating budget is set to a moderate level output by the state model. For the same candidate set, the gating service typically enables semantic and creator factors while suppressing style and social factors. After sorting the fused vector inputs, compared to full fusion without gating, click-through rate and dwell time steadily increase, and there is no sudden drop in scores when switching topics. This result demonstrates that binary gating with gating budget constraints and weighted fusion can reduce the impact of noisy channels while ensuring relevance, achieving more robust cross-scenario performance.

[0120] S170. Calculate energy-based scores based on continuous-time user states and fusion vectors, sort the candidate sets, perform exposure bias correction when sorting the candidate sets, and output short video recommendation results.

[0121] S180. The evolution direction of the continuous-time user state is constrained by the gradient of the energy-based scoring.

[0122] The energy-based scoring system employs a dual-path structure that can operate stably in online scenarios: one path is a bilinear interaction between continuous-time user state and fusion vector, and the other is a nonlinear parallel interaction path. The final score is obtained by weighted summation of the two paths. The core calculation expression is:

[0123]

[0124] in, Indicates user Candidate Videos Energy-based scoring function A continuous-time user state vector. For the fusion vector, It is a bilinear weight matrix. It is a feedforward network. To output the weight vector, This is an element-wise multiplication.

[0125] The sorting phase performs list-based normalization on the candidate set to determine the order in which they are displayed. The core calculation expression is:

[0126]

[0127] Indicates that for a given user Under the condition of, candidate set Chinese candidate video (Candidate Video) For candidate set The conditional probability of any candidate video in the dataset being selected as a positive sample is given by the energy-based scoring function. Obtained by softmax normalization, For temperature parameters, Let be the candidate set. Online inference outputs the top few items in descending order of score as the recommendation results; offline training minimizes the negative log-likelihood generated by the above formula using cross-entropy or contrastive loss.

[0128] To ensure that the update direction of continuous-time user states aligns with the direction of increasing positive sample scores, this step introduces a directional constraint: the gradient of the energy-based score with respect to continuous-time user states is calculated, requiring that the evolution direction of the continuous-time user states be consistent with the direction of this gradient. The core calculation expression is:

[0129]

[0130] For loss of directional consistency, For the direction of the scoring gradient, As a positive sample identifier, The output direction of the state update function is used, and the direction consistency loss constrains the consistency of the two directions through cosine similarity. The signal is time-driven. During training, the orientation consistency loss and ranking loss are weighted and summed together for optimization, while online, only the forward computation and score output are retained.

[0131] In terms of implementation, scoring and ranking are performed in batches within the same inference service. First, bilinear interaction terms are calculated for the candidate set, then nonlinear interaction paths are added, and finally, energy-based scores are synthesized and list-based normalization is completed. To reduce latency, bilinear interaction terms are pre-computed using dense matrix multiplication, and nonlinear interaction paths utilize lightweight feedforward networks with half-precision inference enabled. Necessary rule constraints and deduplication are performed before ranking output to ensure compliance and diversity. The server monitors the score range distribution, triggering automatic calibration of temperature parameters and the normalization layer when distribution drift occurs.

[0132] The training of directional constraints is conducted offline: training batches are constructed using positive and negative samples observed in the logs, and energy-based scores are obtained by forward propagation. The gradients of the energy-based scores with respect to the user's state over continuous time are preserved in the computation graph. The directional consistency loss is calculated, and the state update function and score parameters are updated in reverse. To ensure stability, gradients are pruned during training, and the distribution of directional angles is recorded as a convergence metric. Specifically, the gradients obtained from each training iteration are truncated to an upper norm limit; gradients exceeding this limit are scaled down proportionally to prevent excessively large update steps from causing divergence. The distribution of directional angles is statistically analyzed (a more concentrated distribution indicates greater stability) to determine training stability, whether the learning rate needs to be reduced / regularized, and as a risk signal before large-scale deployment. In the online phase, gradients are not calculated; only the trained parameters are used for inference.

[0133] In terms of system architecture, this step involves collaboration between the rating and ranking service and the feature service. The feature service provides continuous-time user status and fused vectors, which the rating and ranking service receives and performs score calculation and ranking. The two services exchange batch vectors via an in-memory queue, supporting timeout isolation and degradation. When the rating and ranking service times out or encounters an anomaly, the system degrades to a lightweight ranking based on historical click-through rates to ensure business continuity.

[0134] For example, a group of users is selected from actual traffic, and the fusion vectors of continuous-time user states and corresponding candidates are input into the rating and ranking service to obtain energy-based scores and output recommendation results. After adding directional consistency loss during offline training, compared with the baseline using only ranking loss, the score fluctuation during topic switching in online sessions is reduced, and both click-through rate and dwell time are improved; at the same time, the norm and directional distribution of continuous-time user states are more stable, indicating that the evolution direction of continuous-time user states has aligned with the direction of score improvement. This embodiment shows that the calculation process and system decomposition of this step can be directly implemented in existing recommendation architectures and achieve more stable ranking performance in complex session scenarios.

[0135] In some implementations, calculating an energy-based score based on continuous-time user states and fusion vectors, and ranking the candidate set, includes: calculating bilinear and nonlinear interaction terms between continuous-time user states and fusion vectors and weighting them to form an energy-based score; and generating short video recommendation results based on the energy-based score using a score-based list ranking method.

[0136] The input consists of continuous-time user states and a fused vector. Before scoring, both are normalized and dimensionally aligned to ensure consistent dimensions and facilitate batch parallelism. The scoring model employs a dual-path structure: one path is a bilinear interaction to capture similarity along the main direction; the other is a nonlinear interaction to express nonlinear preferences and interaction terms. Ranking uses a score-based list method, calculating the relative probabilities of the candidate sets and determining the order accordingly.

[0137] Online inference outputs the top few items in descending order of energy-based scores; offline training minimizes the negative log-likelihood of the above formula using cross-entropy or contrastive loss, constructs positive and negative samples using logs and performs hard sample sampling to ensure that the ranking has sufficient discriminative power for similar content.

[0138] To ensure online feasibility, the model is designed as follows: First, batch normalization and lightweight linear transformations are used in the vector preprocessing stage to avoid costly normalization operations penetrating the entire graph. Second, bilinear interaction terms are computed in parallel using dense matrix multiplication, and nonlinear interaction paths employ a two-layer feedforward network with half-precision inference enabled to ensure latency stability. Third, score calibration is performed online using the mean and scale of sliding statistics to prevent list jitter caused by distribution drift. Fourth, display rules and diversity constraints are applied after sorting, including isolation from creators, isolation between topics, and near-duplicate removal, to ensure a good user experience.

[0139] The training and deployment process follows a strategy of "offline training, canary-scale validation, and online forward pass". In the offline phase, samples are extracted from recent logs, candidate sets are constructed by session, and parameter training and distillation are completed. In the canary-scale phase, score distribution, click-through rate, and dwell time are compared on small-scale traffic. Once a threshold is reached, the entire dataset is pushed out (first, small-scale canary-scale comparisons are performed with online metrics (click-through rate, dwell time, score distribution stability, etc.), and traffic is gradually increased until the entire dataset is reached after meeting the preset threshold). In the online phase, only the forward graph is loaded; no backward calculation is performed, and all weights are accompanied by version numbers for rollback. To enhance stability, gradients are pruned during training, weight decay is introduced for parameters, and consistency metrics between training and online distributions are monitored.

[0140] The system architecture consists of a feature service and a rating and ranking service. The feature service provides continuous-time user status and fusion vectors. Upon receiving these, the rating and ranking service performs energy-based scoring and list sorting, and writes the results along with necessary explanatory information (including candidate sources (recall channels / indexes), the set of factors with gating enabled and their weight overview, key scoring components (semantic / style / social contribution summaries), the triggering reasons for deduplication and frequency control, the final ranking score, and the version number) to the exposure log, supporting subsequent evaluation and retraining. The two services communicate via a batch interface with timeout and degradation strategies. When the rating and ranking service times out, the system reverts to the click-through rate-based alternative ranking to ensure continuity.

[0141] For example, a batch of typical session traffic (i.e., a representative sample set covering high-frequency scenarios, such as stable browsing, quick skipping, topic switching, strong social interaction, cold-start new users, etc., and stratified sampling by time period / device / author domain to approximate online distribution) is selected. The fusion vector of continuous time user status and corresponding candidates is input into the rating and ranking service to calculate the energy-based score and generate a list. While maintaining the same recall, compared with the single-path dot product model, the dual-path energy-based scoring in this step improves the click-through rate and dwell time in the topic switching scenario, reduces the compression ratio of similar content in the list, and improves the diversity of display. Online monitoring shows that the score distribution is stable, the temperature parameter does not need to be adjusted frequently, and the latency and throughput requirements are met. The above results show that this rating and ranking design can be directly deployed in existing recommendation systems and achieve stable benefits.

[0142] In some implementations, constraining the evolution direction of the continuous-time user state based on the gradient of the energy score includes: calculating the gradient of the energy score relative to the continuous-time user state; decomposing the gradient into conservative components containing directional information; and aligning the evolution direction of the continuous-time user state with the direction of the conservative components.

[0143] The energy-based score is calculated from the continuous-time user state and the fused vector. During training, a computational graph is constructed for each user and its positive candidate samples to obtain the gradient of the energy-based score with respect to the continuous-time user state. The core calculation expression is:

[0144]

[0145] The gradient vector, A continuous-time user state vector. For positive samples, the energy-based score is used. This is a positive sample identifier.

[0146] To extract the stable and guideable conservative component from the gradient, this invention introduces a scalar potential function network, which outputs a scalar potential value for the continuous-time user state, and obtains the conservative component by differentiating this potential value with respect to the continuous-time user state. The core calculation expression is as follows:

[0147]

[0148] For conservative weighting, The scalar potential function is used. During training, the mean squared error between the conservative component and the original gradient is minimized using a batch of samples. At the same time, smoothing regularization and norm constraints are applied to the conservative component to prevent noise direction from being amplified. Since the conservative component is obtained by the derivative of the scalar potential function with respect to the continuous-time user state, it naturally possesses directional consistency and path independence, making it easy to reuse in different session stages.

[0149] Conservative components are used to constrain the evolution direction of continuous-time user states. The basic update of the continuous-time user state is given by the state update function and the time-driven signal. This invention aligns the basic update direction with the conservative components using a direction consistency loss; the core calculation expression is:

[0150]

[0151] For loss of directional consistency, This indicates the output direction of the state update function. The signal is time-driven. The total training loss is:

[0152]

[0153]

[0154]

[0155] in, Total training loss; For the candidate set; For temperature parameters; This represents the direction of the scoring gradient. The output direction of the state update function is used, and the direction consistency loss constrains the consistency of the two directions through cosine similarity. Indicates that for a given user Under the condition of, candidate set Chinese candidate video The conditional probability of being selected as a positive sample is given by the energy-based scoring function. Obtained by softmax normalization; Indicates that for a given user Under these conditions, positive sample identification The conditional probability of the corresponding positive sample, that is... Substitution The obtained probability value; This represents the ranking loss, used to measure how well the model fits the ranking results of positive samples.

[0156] The total loss is calculated using backpropagation. The gradients of the scoring model, state update function, and scalar potential function network parameters are jointly optimized and updated using gradient descent. Training details include: pruning the L2 norm of the gradients of each network parameter to limit gradient explosion; using exponential moving average for batch gradient estimation to suppress jitter; and applying orientation consistency loss. Cosine similarity is used to achieve orientation alignment and reduce the impact of dimensional differences.

[0157] Backward gradients are not calculated during online computation. For continuous-time user states, a base direction is first determined by the state update function. Then, a conservative component is calculated by a scalar potential function network as a reference. A direction mixing strategy is introduced to synthesize the two into the final update direction, ensuring a controlled update step size. To avoid introducing additional instability, the weights for direction mixing are given by the stability and error statistics of the most recent session, with a safety threshold set. When the conservative component's amplitude is too small or its difference from the base direction is too large, the process reverts to the base direction. Specifically, a fast, low-cost update direction is first provided by the state update function; then, a more conservative and smoother reference direction is provided by the trend network. The two are weighted and synthesized according to the rule of "higher session stability leads to greater trust in the reference direction, lower stability leads to greater trust in the fast direction," with a limit on the maximum update step size. When the difference between the two directions is too large or the reference direction is too weak, the process directly reverts to the fast direction. This strategy guarantees controllable latency and stability in online computation.

[0158] In engineering implementation, gradient acquisition during training relies on an automatic differentiation framework. For energy-based scoring, only gradient backpropagation for continuous-time user states is retained, avoiding the propagation of constraints to irrelevant early features. The scalar potential function network employs a small feedforward structure for easy online differentiation. To improve numerical stability, conservative components are normalized and gently smoothed before use. The orientation consistency loss and ranking loss share the same batch of data, reducing data transfer overhead. The online service deploys the scalar potential function network, scoring model, and state model together, all loaded in read-only mode, supporting canary deployments and version rollbacks. The monitoring system records the orientation angle distribution, continuous-time user state norm, and list score variance, promptly downgrading upon anomaly detection.

[0159] Compared to directly using the original gradient, the conservative component is derived from the derivative of the scalar potential function, avoiding unstable directions caused by intra-batch noise, random negative sampling, and unobservable factors. Compared to the baseline system without directional constraints, this invention can keep the update direction consistent with the direction of score improvement under abrupt changes in interest or complex sessions, reducing invalid swings in user state over continuous time, and making the next recall and ranking closer to the user's true intention.

[0160] For example, in real business traffic, continuous-time user states and corresponding positive and negative sample logs (based on exposure, positive samples are those "clicked or effectively stayed / viewed / shared" within the same session window; negative samples are those "quickly skipped after exposure, with a very short stay and no subsequent interaction"; and, if necessary, non-clicked candidates under the same exposure conditions are added as hard negative samples) are used for training. Energy-based scores are calculated offline to obtain the gradient with respect to continuous-time user states. A scalar potential function network is trained so that its conservative components approximate this gradient, and the directional consistency loss and ranking loss are jointly optimized. After deployment, scoring and directional constraints only perform forward inference. Evaluation shows that in sessions with rapid topic switching, adding directional constraints reduces the variance of list scores, increases click-through rate and stay metrics, and makes the directional changes of continuous-time user states smoother, proving that this step has clear feasibility and benefits in a real-world system.

[0161] In some alternative implementations, the evolution direction of continuous-time user states can be constrained by pruning the gradient of the energy-based score. The core principle is to limit the maximum norm of the gradient to prevent abnormal gradients (such as gradient abrupt changes caused by accidental extreme user behavior) from causing drastic fluctuations in user states, thus ensuring the stability of state evolution.

[0162] For example, constraining the evolution direction of the continuous-time user state based on the gradient of the energy score may include: setting a maximum norm threshold for the gradient; calculating the Euclidean norm of the gradient of the energy score; if the Euclidean norm is greater than the maximum norm threshold, then pruning the gradient; updating the user state based on the pruned gradient; if the Euclidean norm does not exceed the maximum norm threshold, then keeping the gradient unchanged.

[0163] In some implementations, the exposure bias correction method includes: training a bias rating model that includes display location identifiers, interface template identifiers, and time period identifiers; calculating the correction gain using a double robust estimation method based on the estimated gain and the exposure probability output by the bias rating model; and using the correction gain for ranking weighting to reduce the impact of exposure bias.

[0164] Logs containing exposure, clicks, and playback duration are collected to construct sample key-value pairs (user, short video). Display location identifiers, interface template identifiers, and time period identifiers are recorded. Additionally, available contextual features such as device type, network status, and session length are supplemented as contextual conditions for bias correction and robustness control (distinguishing between short playback durations caused by network lag and exposure deviations from different entry points). This prevents non-interesting factors from being mistakenly learned as preference signals, thereby improving online stability and interpretability. One-hot encoding is used for display locations and interface templates, and time periods are discretized and binned, with time zones unified and duplicates removed.

[0165] A propensity score model is used. A lightweight binary classification model is employed to fit the probabilities of being displayed, and posterior calibration is performed to obtain an output that can be used as a probability. Implementation options include logistic regression or gradient boosting trees, with a distilled linear model loaded online to meet time delay constraints. The core expression is:

[0166]

[0167] For exposure probability, For logical functions, For model parameters, This is a feature vector containing display location identifiers, interface template identifiers, and time period identifiers. After training, [the following is applied]... Perform monotonic calibration and clip to interval Avoid extreme weighting.

[0168] The resulting model uses clicks or valid plays as labels, takes continuous-time user states and fusion vectors as input, and outputs estimated revenue. This model shares the underlying representation with the ranking model, ensuring consistency in objectives. It employs a double robust correction mechanism. A correction gain is calculated for each training sample, which serves as the sample weight for ranking learning.

[0169]

[0170] To adjust the earnings, To observe the benefits (such as whether there are clicks or whether the effective playback exceeds the threshold). For the estimated returns of the outcome model, This represents the exposure probability of the biased scoring model. To improve robustness, [the following is added]: After truncation and quantile normalization, non-negative weights are obtained. .

[0171] Ranking weighting and training. This maps the adjusted gains to sample weights in the list loss, or directly affects the training objective of the energy-based scoring. A typical approach is to weight the list-based cross-entropy, or multiply the point / pair loss by... At the same time, for extremely small A lower bound is applied to the samples to avoid training instability. To ensure online consistency, a weight mapping function is learned offline synchronously. and Mapped to sorting temperature and fractional bias, it can be used online with only forward computation required.

[0172] Online application. Real-time sorting does not depend on real-world data. The system obtains data from a propensity score model. Obtain from the result model Using offline learned monotonic mappings Obtain weighting factors and perform temperature calibration. ,in, Based on the fundamental temperature constant, For the propensity score model output value, It is a monotonically adjusting function. The final score is synthesized using a weighted method controllable by the engineering process:

[0173]

[0174] The weighted score. It is an energy-based scoring system. These are the weighting coefficients. This is a range pruning function. This approach does not require real labels online and adheres to computational budget constraints.

[0175] System architecture. The offline training pipeline retrains the propensity score model and the result model daily, mapping output parameters to calibration; the ranking service loads read-only weights online and performs batch inference after receiving the candidate set. and ,generate and ,synthesis And complete the list sorting. The monitoring side tracks weight distribution, Distribute and calibrate residuals, and roll back to the previous version in case of anomalies.

[0176] It should be noted that the display location and interface template are frozen in buckets to avoid code drift during the training-to-deployment period; Set a lower bound, commonly using quantiles or a fixed constant; self-normalize the weights to prevent specific traffic segments from dominating the loss. To avoid "reverse amplification," only a moderate amount is allowed online. Adjust the temperature and compare the changes in click-through rate and dwell time before and after sorting during the grayscale period.

[0177] For example, select one week's logs, train a propensity scoring model based on display location identifier, interface template identifier, and time period identifier, and output the results from the model. Based on this, the corrected revenue is calculated, and the ranking model is trained by weighting the list loss. This is then implemented in a real-world, phased-out test. and Weighted composite score Compared with the baseline without bias correction, the advantage of the first screen position was suppressed, the effective playback rate of content in the middle and later positions was improved, and the overall click-through rate and viewing time steadily increased, proving that exposure bias correction can be implemented in the engineering system and produce positive effects.

[0178] In some alternative implementations, exposure bias correction methods also include bias score matching correction. The core principle is to group candidate videos according to their bias scores, rank them within each group, eliminate the influence of exposure differences between groups, and ensure that videos with different exposure probabilities are compared from the same starting point.

[0179] For example, the tendency scores of all videos in the candidate set are binned (e.g., equal-frequency binning) to divide them into multiple groups with similar exposure probabilities; within each group, local ranking is performed based on energy-based scores; the ranking results of each group are integrated, and the overall user matching weight of the group is combined to generate the final recommendation sequence.

[0180] In summary, the short video recommendation method provided in this application has the following beneficial effects:

[0181] Deep fusion of multimodal features for accurate preference characterization: By extracting multimodal features from video, audio, and text, the limitations of single-modal information are avoided, and content attributes and user preferences are comprehensively captured to improve recommendation accuracy.

[0182] Distribution alignment solves cross-scene drift: Semantic distribution alignment driven by environment tags achieves unified representation across creators and time periods, improves the consistency of cross-scene recommendations, and reduces ranking instability caused by environment drift.

[0183] Continuous-time user state modeling to adapt to real-time intent: By modeling the continuous-time user state based on user event streams, real-time response to changes in interest and conversation rhythm is achieved, improving the timeliness of recommendations.

[0184] Sparse gating balances quality and efficiency: By employing gating budget and factor-level sparse gating, non-target channels (style noise, social noise) are suppressed, improving interpretability and robustness. Simultaneously, computational complexity is significantly reduced, meeting millisecond-level latency requirements for online recommendations while balancing recommendation quality and engineering feasibility.

[0185] Exposure bias correction enhances model robustness: Exposure bias correction is performed when sorting the candidate set, which effectively corrects the exposure bias caused by display position and traffic strategy, solves the bias mismatch problem between training and deployment, improves the robustness of the recommendation system, and achieves exposure bias correction and fairer ranking results.

[0186] User state evolution direction constraint reduces intent drift: By using a scoring gradient to constrain the consistency of user state direction, the training objective and online evolution are integrated, reducing intent drift and improving recommendation performance in cold start and interest mutation scenarios. After outputting short video recommendation results, it ensures that user states (such as interests and needs) evolve smoothly and reasonably over time, achieving the goal of short video recommendations that better match users' long-term behavior or interest changes.

[0187] Figure 2 This is a structural block diagram of a short video recommendation system provided in an embodiment of the present invention. This short video recommendation system is used to execute the short video recommendation method provided in the above embodiment. Figure 2 As shown, the short video recommendation system 200 may include:

[0188] The acquisition and alignment module 201 is used to receive short video content data and user behavior log data; based on the short video content data, it extracts video features, audio features and text features, and generates content factor vectors through factorization projection, wherein the content factor vectors include semantic factors; it performs distribution alignment of semantic factors based on environmental labels to generate aligned semantic factors, and organizes user behavior log data into user event streams;

[0189] The state fusion module 202 is used to construct continuous-time user states based on user event streams and generate query vectors to determine the gating budget; to perform candidate recall in an index database built from aligned semantic factors based on the query vectors to obtain a candidate set; and to perform sparse gating on the content factor vectors under the constraint of the gating budget to generate a fusion vector.

[0190] The scoring and ranking module 203 is used to calculate the energy-based score based on the continuous-time user state and the fusion vector, and to rank the candidate set. During the ranking of the candidate set, exposure bias correction is performed, and short video recommendation results are output. The gradient of the energy-based score constrains the evolution direction of the continuous-time user state.

[0191] The acquisition and alignment module 201 is deployed on the integrated access and feature computation node. A typical configuration includes dual general-purpose processors, ≥256GB of memory, a PCIe Gen4 / Gen5 bus, dual-port 100GbE network cards, and a local cache pool consisting of multiple NVMe SSDs. The node's built-in data plane GPU is used for video hardware decoding and multimodal feature inference (NVDEC / video codec unit + Tensor cores), while lightweight audio and text processing is handled by CPU vector instructions. PTP time synchronization is used to ensure cross-machine alignment; environment tag alignment and factorization projection are batch-processed on the GPU, with results written to object storage and memory key-value cache to support subsequent low-latency reads. The inbound port supports programmable switches or SmartNICs for rate limiting and packet capture, ensuring peak stability and data integrity.

[0192] For the state fusion module 202, online session computation and vector retrieval are deployed in a layered manner. Session computation nodes utilize a high-frequency CPU + large memory configuration to handle user event stream aggregation, continuous-time user state updates, and gating budget generation. To reduce tail latency, the network interface card supports RoCEv2 / RDMA and enables NUMA affinity. The vector retrieval cluster consists of GPU vector index nodes and large-memory CPU index nodes: hot indexes reside in GPU memory for ANN retrieval, while cold indexes reside in DDR + NVMe layered storage. These two types of nodes are interconnected via a 100 / 200GbE leaf-spine network. Gating computation and factor-level sparse selection are deployed on the same machine as retrieval to minimize cross-machine copying. The entire link utilizes observability probe cards, latency sampling, and power redundancy to improve stability.

[0193] For the scoring and ranking module 203, the online inference cluster adopts a hybrid configuration of "CPU as the primary and GPU as the secondary": a high-frequency multi-core CPU is responsible for batch bilinear interaction and list normalization, while a small-scale inference GPU carries the nonlinear interaction feedforward network. The two share intermediate features via PCIe zero-copy, and the pre-TLS / HTTP offloading is completed by a SmartNIC to free up CPU resources. Model weights are protected by both in-memory KV cache and NVMe local image, with automatic rollback in case of anomalies. The offline cluster used for gradient direction constraints and periodic training uses multi-GPU NVLink servers, distributed high-speed storage, and 400GbE interconnect. After training, the model is distributed to the online cluster through a model repository and a canary release system to ensure low-latency inference and rollbackable operation and maintenance.

[0194] The short video recommendation system provided in this embodiment of the invention is used to execute the short video recommendation method provided in the above embodiments, and has the same or corresponding technical features and effects. Contents not described in detail in this embodiment can be referred to the above embodiments, and will not be repeated here.

[0195] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Figure 3 As shown, the electronic device 400 includes one or more processors 401 and memory 402.

[0196] The processor 401 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 400 to perform desired functions.

[0197] The memory 402 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 401 may execute the program instructions to implement the short video recommendation method of any embodiment of the present invention described above and / or other desired functions. Various contents such as initial extrinsic parameters and thresholds may also be stored in the computer-readable storage medium.

[0198] In one example, the electronic device 400 may further include an input device 403 and an output device 404, these components being interconnected via a bus system and / or other forms of connection mechanisms (not shown). The input device 403 may include, for example, a keyboard, a mouse, etc. The output device 404 may output various information to the outside, including warning messages, braking force, etc. The output device 404 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0199] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device 400 relevant to the present invention are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 400 may include any other suitable components depending on the specific application.

[0200] In addition to the methods and devices described above, embodiments of the present invention may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps of the short video recommendation method provided in any embodiment of the present invention.

[0201] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of the present invention. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0202] Furthermore, embodiments of the present invention may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps of the short video recommendation method provided in any embodiment of the present invention.

[0203] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0204] It should be noted that the terminology used in this invention is for describing specific embodiments only and is not intended to limit the scope of this application. As shown in this specification, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may include the plural. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0205] It should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," "linked," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components. For those skilled in the art, the specific meaning of the above terms in the present invention can be understood according to the specific circumstances.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A short video recommendation method, characterized in that, include: Receive short video content data and user behavior log data; Based on the short video content data, video features, audio features and text features are extracted, and a content factor vector is generated through factorized projection, wherein the content factor vector includes semantic factors. Based on environmental tags, the semantic factors are distributed and aligned to generate aligned semantic factors, and the user behavior log data is organized into a user event stream. Based on the user event stream, a continuous-time user state is constructed and a query vector is generated to determine the gating budget; Based on the query vector, candidate retrieval is performed in an index database built from aligned semantic factors to obtain a candidate set; Sparse gating is performed on the content factor vector under the constraint of the gating budget to generate a fusion vector; Based on the continuous time user state and the fusion vector, an energy-based score is calculated and the candidate set is sorted. Exposure bias correction is performed when sorting the candidate set, and short video recommendation results are output. The evolution direction of the continuous-time user state is constrained by the gradient of the energy-based scoring.

2. The short video recommendation method according to claim 1, characterized in that, The exposure bias correction method includes: The training includes a propensity rating model that displays location identifiers, interface template identifiers, and time period identifiers; Based on the estimated revenue and the exposure probability output by the propensity scoring model, a double robust estimation method is used to calculate the corrected revenue. The adjusted returns will be used for ranking weighting.

3. The short video recommendation method according to claim 1, characterized in that, The evolution direction of the continuous-time user state is constrained by the gradient of the energy-based scoring, including: Calculate the gradient of the energy-based score with respect to the continuous-time user state; The gradient is decomposed into conservative components that contain directional information; The evolution direction of the continuous-time user state is aligned with the direction of the conservative component.

4. The short video recommendation method according to claim 1, characterized in that, Based on the short video content data, video features, audio features, and text features are extracted, and content factor vectors are generated through factorized projection, including: Based on the short video content data, multimodal features are obtained through a visual encoder, an audio encoder, and a text encoder, respectively. Multimodal features are mapped to content factor vectors through linear projection and branching structures. The content factor vectors also include style factors, creator factors, and social factors.

5. The short video recommendation method according to claim 4, characterized in that, Based on environmental tags, the semantic factors are distributed and aligned to generate aligned semantic factors. The user behavior log data is then organized into a user event stream, including: The semantic factors are aligned between different environments corresponding to the environment label using a distributed mapping method to obtain the aligned semantic factors; User behavior log data is grouped into user event streams in chronological order, including exposure, clicks, dwell time, fast forward, rewatch, and sharing.

6. The short video recommendation method according to claim 5, characterized in that, Based on the user event stream, a continuous-time user state is constructed and a query vector is generated to determine the gating budget, including: The user event stream is represented as a time-driven signal; Based on the time-driven signal, a continuous-time model is used to model the continuous-time user state; When fast forward, rewatch, or long pause is detected, the state is updated based on the aligned semantic factors. A query vector is generated from the continuous-time user state, and a gating budget is determined based on the control rules.

7. The short video recommendation method according to claim 1, characterized in that, Based on the query vector, candidate retrieval is performed in an index database built from aligned semantic factors, including: Map the query vector to the semantic retrieval space; In a vector index library constructed with aligned semantic factors, the candidate set is obtained by approximate nearest neighbor retrieval based on the query vector.

8. The short video recommendation method according to claim 6, characterized in that, Sparse gating is performed on the content factor vector under the constraint of the gating budget to generate a fused vector, including: Gating variables are set for the semantic factor, the style factor, the creator factor, and the social factor, respectively. Sparse selection is achieved using a binary gating function; Under the constraints of the gating budget, the passed factors are weighted and summed to generate a fusion vector.

9. The short video recommendation method according to claim 1, characterized in that, Calculating an energy-based score based on the continuous-time user state and the fusion vector, and ranking the candidate set, includes: The bilinear and nonlinear interaction terms between the continuous-time user state and the fusion vector are calculated and weighted to form an energy-based score. Based on the energy-based score, the short video recommendation results are generated using a score-based list sorting method.

10. A short video recommendation system, characterized in that, include: The acquisition and alignment module is used to receive short video content data and user behavior log data; based on the short video content data, it extracts video features, audio features and text features, and generates a content factor vector through factorized projection, wherein the content factor vector includes semantic factors; it performs distribution alignment on the semantic factors based on environmental labels to generate aligned semantic factors, and organizes the user behavior log data into a user event stream; The state fusion module is used to construct continuous-time user states based on the user event stream and generate query vectors, and determine the gating budget; based on the query vectors, perform candidate recall in an index library built from aligned semantic factors to obtain a candidate set; and perform sparse gating on the content factor vectors under the constraints of the gating budget to generate a fusion vector. The scoring and ranking module is used to calculate energy-based scores based on the continuous-time user status and the fusion vector, rank the candidate set, perform exposure bias correction when ranking the candidate set, and output short video recommendation results. The evolution direction of the continuous-time user state is constrained by the gradient of the energy-based scoring.

Images