Multi-scene oriented multi-task joint learning recommendation system
By dynamically managing sample states and parameter updates, the problems of false negatives and negative transfer in multi-scenario, multi-task recommendation are solved, improving the training stability and recommendation accuracy of the model.
Patent Information
- Application Number
- CN202611461518.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2026-08-12
- Filing Date
- 2026-09-20
- Publication Date
- 2026-11-03
AI Technical Summary
Existing multi-scenario, multi-task recommendation methods are prone to generating false negative samples and negative transfer when dealing with differences in feedback speed across different recommendation scenarios and tasks. Furthermore, they lack dynamic processing mechanisms, leading to model training bias and decreased prediction performance.
A multi-task joint learning recommendation system for multiple scenarios is adopted. Through sample construction module, feedback maturity time determination module, sample state determination module, joint learning module, update path control module and feedback update module, the system dynamically manages sample state and parameter updates to ensure the accuracy and consistency of training samples.
It effectively reduces the impact of false negative samples, lowers training label bias, reduces negative transfer, improves the model's adaptability to changes in user behavior, and enhances the training stability and accuracy of recommendation results for multi-scenario recommendation models.
Smart Images
Figure CN123030812A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent recommendation, and in particular to a multi-task joint learning recommendation system for multiple scenarios. Background Technology
[0002] As internet platform services become increasingly diverse, recommendation systems typically need to simultaneously serve multiple recommendation scenarios, such as homepage recommendations, search recommendations, short video recommendations, and product recommendations, and jointly predict multiple recommendation tasks, including clicks, dwell time, favorites, completions, and conversions. Existing multi-scenario, multi-task recommendation methods usually utilize shared networks to extract common features across different scenarios and tasks, and then obtain corresponding prediction results through task output layers to reduce redundant model training and improve data utilization.
[0003] However, the feedback generation speed varies significantly across different recommendation scenarios and tasks. For example, click feedback typically occurs shortly after exposure, while feedback such as collection, completion, or conversion may have a longer delay. Existing methods often use a uniform observation window or a fixed time window to determine sample labels, which can easily lead to false negatives by prematurely labeling samples that may generate feedback later, resulting in model training bias. Extending the observation window uniformly, on the other hand, increases the waiting time for training samples, reducing the timeliness of model updates.
[0004] Furthermore, while there is information that can be shared between different scenarios and tasks, their data distribution, user behavior patterns, and optimization directions are not entirely consistent. Existing joint learning networks typically employ a fixed parameter-sharing structure, failing to dynamically adjust the sharing relationship based on the parameter update direction of different scenario-task combinations during training. When the update directions of some scenarios or tasks conflict, negative transfer can easily occur, leading to a decline in the prediction performance of some recommendation tasks. Simultaneously, existing methods lack a complete dynamic handling mechanism for subsequent state updates of unconfirmed samples, late feedback replenishment, and cold starts in new scenarios.
[0005] To address this, a multi-task joint learning recommendation system for multiple scenarios is proposed. Summary of the Invention
[0006] The main objective of this invention is to provide a multi-task joint learning recommendation system for multiple scenarios, which can effectively solve the problems in the background technology.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A multi-task joint learning recommendation system for multiple scenarios includes an interaction data acquisition interface, a memory, a processor, and a recommendation result output interface. The memory stores historical interaction samples from multiple recommendation scenarios and network parameters of the joint learning network. The processor is configured with:
[0009] The sample construction module is used to acquire historical interaction data in multiple recommendation scenarios through the interaction data acquisition interface, associate the exposure event of the recommended object with multiple task feedback events that occur after the exposure event, and construct historical interaction samples corresponding to each recommendation scenario and each recommendation task.
[0010] The feedback maturity duration determination module is used to determine the feedback maturity duration for each scenario-task combination based on the feedback delay from the exposure event to the corresponding task feedback event in the historical interaction samples of the completed observation period.
[0011] The sample state determination module is used to determine the feedback state of historical interaction samples for each recommendation task based on the current training time, the exposure time of historical interaction samples, and the feedback maturity time of the corresponding scene-task combination. Specifically, historical interaction samples that have generated corresponding task feedback events are determined as mature positive samples, historical interaction samples that have not generated corresponding task feedback events but have reached the corresponding feedback maturity time after exposure are determined as mature negative samples, and historical interaction samples that have not generated corresponding task feedback events but have not reached the corresponding feedback maturity time after exposure are determined as samples to be confirmed. The mature positive samples and the mature negative samples are combined into a valid training sample set.
[0012] The joint learning module is used to construct a joint learning network that includes a shared parameter set and a dedicated parameter set corresponding to each scenario-task combination. It calculates the task loss based on the effective training sample set of each scenario-task combination and obtains the parameter update vector of each scenario-task combination for the shared parameter set based on the task loss.
[0013] The update path control module is used to calculate the update consistency between parameter update vectors of different scenario-task combinations, determine the parameter update path of the corresponding scenario-task combination based on the update consistency, use the parameter update vectors that meet the sharing conditions to update the shared parameter group, and use the task loss corresponding to the scenario-task combination that meets the conflict isolation conditions to update the corresponding proprietary parameter group.
[0014] The recommendation output module is used to obtain user features, candidate recommendation object features and scene context features in the target recommendation scenario, obtain prediction results of multiple recommendation tasks through the updated joint learning network, determine the recommendation score of the candidate recommendation object based on the prediction results of the multiple recommendation tasks, generate a recommendation list according to the recommendation score, and output the recommendation list through the recommendation result output interface.
[0015] The feedback update module is used to obtain the actual task feedback events generated after the recommendation list is output, update the samples to be confirmed and the historical interaction samples with confirmed feedback status according to the actual task feedback events, and redetermine the feedback maturity time of each scenario-task combination according to the preset update cycle.
[0016] Furthermore, the historical interaction data includes at least user characteristics, recommendation object identifier, recommendation object characteristics, scene identifier, scene context characteristics, exposure time, task feedback event type, and task feedback event occurrence time;
[0017] The sample construction module is used to associate task feedback events that occur within a preset associated time period with the corresponding exposure events based on the exposure event identifier, user identifier, and recommended object identifier, forming a chain of scene interaction events arranged according to the time of occurrence.
[0018] The recommended task includes at least two of the following: click prediction task, effective stay prediction task, favorite prediction task, completion prediction task, and conversion prediction task. The historical interaction samples that have completed the observation period are historical interaction samples whose time distance from the exposure time to the sample statistics time reaches the preset maximum observation time.
[0019] Furthermore, the feedback maturity time determination module is used to determine the scenario. Recommended tasks Based on historical interaction samples that have completed the observation period and generated corresponding task feedback events, a feedback delay set is determined, and a cumulative distribution of feedback delays is established based on the feedback delay set. The feedback maturation time is determined according to the following formula:
[0020]
[0021] in, Representing a scene Recommended Tasks Feedback maturity time This indicates a delay in feedback. This indicates that the feedback delay is no greater than The proportion of the number of task feedback events to the total number of corresponding task feedback events. This indicates the preset feedback coverage ratio, and ;
[0022] When the scene Recommended Tasks When the number of corresponding task feedback events is less than a preset sample size threshold, the recommendation task is used. The cumulative distribution of feedback delays in other recommendation scenarios is used to determine the task-level feedback maturity time of recommendation task k, and the task-level feedback maturity time is used as the scenario. Recommended Tasks Feedback maturation time.
[0023] Furthermore, the sample state determination module is used to determine the first sample state according to the following formula. A historical interaction sample for the scenario Recommended Tasks Feedback maturity:
[0024]
[0025] in, Indicates the first A historical interaction sample for the scenario Recommended Tasks Feedback maturity Indicates the current training moment. Indicates the first A historical interaction sample in the scene The exposure time below;
[0026] For historical interaction samples that did not generate corresponding task feedback events, when When, the historical interaction samples are identified as mature negative samples, when When the historical interaction sample is confirmed, it is identified as a sample to be confirmed; for historical interaction samples that have generated corresponding task feedback events, the historical interaction sample is identified as a mature positive sample, which is not limited by the feedback maturity.
[0027] Furthermore, the joint learning network includes a feature encoding layer, a shared representation network, multiple scene-task specific correction networks, and multiple task output layers. The feature encoding layer is used to convert user features, recommendation object features, and scene context features into input feature vectors. The shared representation network is used to generate shared features shared by multiple recommendation scenarios and multiple recommendation tasks based on the input feature vectors. Each scene-task specific correction network corresponds to a scene-task combination and is used to generate specific correction features for the corresponding scene-task combination based on the input feature vectors. Each task output layer corresponds to a recommendation task and is used to output the prediction result of the recommendation task based on the shared features and the specific correction features of the corresponding recommendation task in the current recommendation scenario. The network parameters of the shared representation network form the shared parameter set, and the network parameters of each scene-task specific correction network form the specific parameter set for the corresponding scene-task combination.
[0028] Furthermore, the joint learning module is used to perform the following processing for each scenario-task combination: read mature positive samples and mature negative samples from the corresponding effective training sample set, set the task label of the mature positive samples to 1, set the task label of the mature negative samples to 0, and exclude the samples to be confirmed from the task loss calculation of the corresponding recommendation task; calculate the task loss based on the task label and the prediction result output by the corresponding task output layer, and normalize the task loss according to the number of samples in the effective training sample set; calculate the gradient of the normalized task loss relative to each shared parameter group, and use the gradient as the parameter update for the corresponding scenario-task combination for the corresponding shared parameter group. When the number of effective training samples for a scenario-task combination is less than a preset effective sample count threshold, the use of the scenario-task combination's parameter update vector to update the shared parameter group is paused. For scenario-task combinations that already have a parameter update path, the parameter update path from the previous training cycle is maintained. For new scenario-task combinations that do not have a parameter update path from the previous training cycle, before the number of effective training samples reaches the preset effective sample count threshold, the output of the corresponding scenario-task-specific correction network is initialized to zero, the shared representation network is called to generate the task prediction result, and the use of the new scenario-task combination to update the shared parameter group and the corresponding specific parameter group is paused.
[0029] Furthermore, the update path control module is used for shared parameter groups. Calculate the scenario-task combination according to the following formula. Scene-Task Combination Update consistency:
[0030]
[0031] in, Representing a scenario-task combination Scene-Task Combination For shared parameter groups Update consistency, and These represent the scene-task combination respectively. Scene-Task Combination For shared parameter groups The parameter update vector, Indicates a preset positive number;
[0032] The update path control module is also used to determine the stable update consistency based on the update consistency of multiple consecutive training batches, and to weight the stable update consistency between the current scenario-task combination and other scenario-task combinations based on the number of samples in each effective training sample set, so as to obtain the comprehensive update consistency of the current scenario-task combination.
[0033] When the overall update consistency is not less than the shared threshold, the current scenario-task combination is added to the shared update set, and the corresponding shared parameter group is updated using the parameter update vectors of each scenario-task combination in the shared update set; when the overall update consistency is not greater than the conflict threshold, the update of the corresponding shared parameter group using the parameter update vectors of the current scenario-task combination is stopped, and the corresponding proprietary parameter group is updated using the task loss of the current scenario-task combination; when the overall update consistency is between the conflict threshold and the shared threshold, the parameter update path of the previous training cycle is maintained for scenario-task combinations that already have parameter update paths, and the update of the corresponding shared parameter group is stopped for new scenario-task combinations that do not have parameter update paths in the previous training cycle, and the corresponding proprietary parameter group is updated using its task loss; wherein, the conflict threshold is less than the shared threshold.
[0034] Furthermore, the recommendation output module is used to combine the shared features output by the shared representation network with the specific modified features corresponding to each recommendation task in the target recommendation scenario, obtain the prediction results of each recommendation task through the corresponding task output layer, fuse the prediction results of each recommendation task according to the task fusion weight corresponding to the target recommendation scenario, obtain the recommendation score of the candidate recommendation object, and select a preset number of candidate recommendation objects with the highest recommendation scores to generate a recommendation list.
[0035] The feedback update module is used to update the sample to be confirmed as a mature positive sample when a task feedback event is generated within the corresponding feedback maturity period of the sample to be confirmed; and to update the sample to be confirmed as a mature negative sample when no task feedback event is generated after the corresponding feedback maturity period has been reached. For historical interaction samples that have been identified as mature negative samples but generated task feedback events within the preset maximum correction period, they are corrected to mature positive samples.
[0036] The feedback update module is also used to write historical interaction samples whose feedback status has changed into the back-up training buffer. When the number of samples in the back-up training buffer reaches the preset back-up batch number or reaches the preset incremental training time, replay samples are extracted from the effective training sample set of each scenario-task combination, and the replay samples and historical interaction samples in the back-up training buffer are used together for the incremental training of the joint learning network.
[0037] The feedback update module is also used to redetermine the feedback maturity time based on the updated historical interaction samples within a preset rolling time window. When the difference between the redetermined feedback maturity time and the current feedback maturity time reaches a preset update threshold, the feedback maturity time of the corresponding scenario-task combination is updated, the feedback status of the historical interaction samples within the preset rolling time window is re-determined, and the incremental training of the joint learning network is triggered.
[0038] A multi-task joint learning recommendation method for multiple scenarios includes the following steps:
[0039] S1. Obtain historical interaction data under multiple recommendation scenarios, associate the exposure event of the recommended object with multiple task feedback events that occur after the exposure event, and construct historical interaction samples corresponding to each recommendation scenario and each recommendation task;
[0040] S2. For each scenario-task combination, determine the feedback maturity time of each scenario-task combination based on the feedback delay from the exposure event to the corresponding task feedback event in the historical interaction samples of the completed observation period.
[0041] S3. Based on the current training time, the exposure time of historical interaction samples, and the feedback maturity time of the corresponding scene-task combination, the historical interaction samples that have generated corresponding task feedback events are determined as mature positive samples, the historical interaction samples that have not generated corresponding task feedback events but have reached the corresponding feedback maturity time after exposure are determined as mature negative samples, and the historical interaction samples that have not generated corresponding task feedback events but have not reached the corresponding feedback maturity time after exposure are determined as samples to be confirmed. The mature positive samples and the mature negative samples are combined into an effective training sample set.
[0042] S4. Construct a joint learning network including a shared parameter set and a dedicated parameter set corresponding to each scenario-task combination. Calculate the task loss based on the effective training sample set of each scenario-task combination, and obtain the parameter update vector of each scenario-task combination for the shared parameter set based on the task loss.
[0043] S5. Calculate the update consistency between parameter update vectors of different scenario-task combinations, determine the parameter update path of the corresponding scenario-task combination based on the update consistency, use the parameter update vector that meets the sharing condition to update the shared parameter group, and use the task loss corresponding to the scenario-task combination that meets the conflict isolation condition to update the corresponding proprietary parameter group.
[0044] S6. Obtain user features, candidate recommendation object features, and scene context features in the target recommendation scenario. Obtain prediction results for multiple recommendation tasks through the updated joint learning network. Determine the recommendation score of the candidate recommendation object based on the prediction results of the multiple recommendation tasks. Generate a recommendation list according to the recommendation score and output the recommendation list.
[0045] S7. Obtain the actual task feedback events generated after the recommendation list is output. When a task feedback event occurs within the feedback maturity time corresponding to the sample to be confirmed, update the sample to be confirmed as a mature positive sample. If no task feedback event occurs after the corresponding feedback maturity time has been reached, update the sample to be confirmed as a mature negative sample. For historical interaction samples that have been determined to be mature negative samples but have generated task feedback events within the preset maximum correction time, correct them to mature positive samples. Use the historical interaction samples whose feedback status has changed for incremental training of the joint learning network, and redetermine the feedback maturity time of each scenario-task combination according to the preset update cycle.
[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] 1. This invention determines the feedback maturity time for different scenarios and task combinations, and classifies mature positive samples, mature negative samples and unconfirmed samples according to the time elapsed after sample exposure. Unconfirmed samples are temporarily excluded from task loss calculation, which can reduce false negative samples caused by delayed feedback and reduce training label bias.
[0049] 2. This invention dynamically determines the update path of shared parameter groups or proprietary parameter groups by calculating the update consistency between parameter update vectors of different scenario-task combinations, so that scenarios and tasks with consistent update directions can share training information, and scenarios and tasks with conflicting update directions can be isolated, which helps to reduce negative transfer in multi-task joint training.
[0050] 3. This invention can dynamically update the sample status based on actual task feedback events, and implement incremental training by replenishing the training buffer and replaying samples, so that late feedback can re-participate in model training. This avoids frequent full retraining while improving the adaptability of the joint learning network to changes in user behavior.
[0051] 4. This invention obtains and fuses the prediction results of multiple recommendation tasks in the target recommendation scenario through a joint learning network, which can take into account the prediction objectives of different recommendation tasks and improve the training stability and accuracy of recommendation results of multi-scenario recommendation models. Attached Figure Description
[0052] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0053] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0055] Example 1
[0056] like Figure 1-2 As shown, a multi-task joint learning recommendation system for multiple scenarios is deployed in a recommendation server or a cluster of recommendation servers. The recommendation server includes an interactive data acquisition interface, a memory, a processor, and a recommendation result output interface. The memory may include a database, a distributed file storage, and a cache, used to store historical interaction data, sample states, feedback maturity times, network parameters of the joint learning network, and parameter update paths. The processor may include a central processing unit and a graphics processing unit, used to execute processing logic such as sample construction, feedback maturity time calculation, sample state determination, network training, update path control, and recommendation result generation.
[0057] The interactive data acquisition interface can use a message queue interface, log acquisition interface, or application programming interface (API) to connect to multiple business terminals and business servers. The recommendation result output interface can use an API to send the generated recommendation list to the client or business server corresponding to the target recommendation scenario.
[0058] This embodiment uses homepage information feed, search results, and detail page related recommendations as three recommendation scenarios, and click prediction, effective dwell time prediction, and conversion prediction as three recommendation tasks. The effective dwell time prediction task uses reaching a preset effective dwell time threshold as the task feedback event, while the conversion prediction task uses placing an order, making a payment, or completing a preset business operation as the task feedback event. In other embodiments, additional tasks such as collection prediction, completion prediction, or sharing prediction can be added based on actual business needs.
[0059] During system operation, the system collects exposure events and task feedback events generated by users in different recommendation scenarios through an interactive data acquisition interface. Each exposure event includes at least an exposure event identifier, user identifier, recommendation object identifier, scenario identifier, exposure time, user characteristics, recommendation object characteristics, and scenario context characteristics. Task feedback events include at least a feedback event identifier, user identifier, recommendation object identifier, feedback event type, and feedback event occurrence time.
[0060] The sample construction module prioritizes associating task feedback events with corresponding exposure events based on exposure event identifiers. When a task feedback event does not carry an exposure event identifier, it searches for the nearest exposure event to that task feedback event within a preset association period based on the user identifier, recommended object identifier, and scene identifier, and establishes a relationship. The preset association period can be set according to the recommended task; for example, the association period for click prediction tasks can be set to 30 minutes, and the association period for conversion prediction tasks can be set to 7 days. After the association is completed, starting from the exposure event, subsequent feedback events such as clicks, valid dwell times, favorites, or conversions are arranged according to their occurrence time to form a scene interaction event chain.
[0061] Each exposure event can generate historical interaction samples corresponding to multiple recommendation tasks. The same exposure event may have already formed mature positive samples for click prediction tasks, while it may still be in a pending confirmation state for conversion prediction tasks, thus allowing different recommendation tasks to manage the sample status according to their respective feedback speeds.
[0062] The feedback maturity duration determination module reads historical interaction samples that have completed the observation period according to a preset update cycle. In this embodiment, the preset maximum observation period is set to 7 days, and historical interaction samples whose exposure time to sample statistics time reaches 7 days are determined as historical interaction samples that have completed the observation period. For samples that have generated corresponding task feedback events, the feedback delay is calculated based on the time difference between the feedback event occurrence time and the exposure time.
[0063] For the scenario Recommended tasks Based on the corresponding feedback delay, establish an empirical cumulative distribution function:
[0064]
[0065] in, This indicates that the feedback delay is no greater than The proportion of the number of task feedback events to the total number of corresponding task feedback events. Representing a scene The total number of task feedback events for the recommended task k. This indicates that the feedback delay is no greater than The number of task feedback events.
[0066] The feedback maturation time is determined according to the following formula:
[0067]
[0068] in, Representing a scene Recommended Tasks Feedback maturity time This indicates the preset feedback coverage ratio. This embodiment will... Setting it to 0.95 defines the feedback maturity time as the latency that can cover 95% of historical task feedback events. In practical applications, It can be adjusted between 0.90 and 0.99. Improvement It can reduce the probability of prematurely identifying delayed feedback samples as negative samples, but it will prolong the confirmation time; reducing η can improve the timeliness of samples entering training, but may increase false negative samples.
[0069] To avoid a feedback maturity time of zero or too short, a minimum feedback maturity time greater than zero can be set. The larger of the feedback maturity time obtained according to the cumulative distribution function and the minimum feedback maturity time can be used as the final feedback maturity time. For example, the minimum feedback maturity time for a click prediction task can be set to 1 minute, and the minimum feedback maturity time for a conversion prediction task can be set to 1 hour.
[0070] When the number of task feedback events for a particular scenario-task combination is less than a preset sample size threshold (e.g., less than 1000), a separate cumulative distribution of feedback delays for that scenario-task combination is not established. Instead, the feedback delays of the same recommendation task in other recommendation scenarios are aggregated, and the task-level feedback maturity time is determined based on the aggregated feedback delays. When the recommendation task also lacks historical data in other scenarios, a preset initial feedback maturity time can be used, for example, setting 30 minutes for click prediction tasks, 2 hours for effective dwell time prediction tasks, and 48 hours for conversion prediction tasks. Once the accumulated feedback events reach the preset sample size threshold, the actual feedback delays are recalculated.
[0071] At the current training moment The sample status determination module calculates the first step according to the following formula. A historical interaction sample for the scenario Recommended Tasks Feedback maturity:
[0072]
[0073] in, Indicates the maturity level of the feedback. Indicates the first A historical interaction sample in the scene The moment of exposure.
[0074] If a sample has already generated a corresponding task feedback event, it is classified as a mature positive sample and its task label is set to 1, regardless of whether the feedback maturity level reaches 1. If a sample has not yet generated a corresponding task feedback event and its feedback maturity level reaches 1, it is classified as a mature negative sample and its task label is set to 0. If the feedback maturity level is less than 1, it is classified as a sample awaiting confirmation. Samples awaiting confirmation are stored in the queue of samples awaiting confirmation and are not included in the calculation of the task loss for the corresponding recommendation task.
[0075] For example, a recommended product is exposed on the homepage feed at 10:00 AM on the first day. In this scenario, the feedback maturity time for conversion prediction is 48 hours. If no conversion feedback is generated by 10:00 AM on the second day, the feedback maturity of this sample is 0.5, and therefore it remains a sample awaiting confirmation. If a user generates conversion feedback before 10:00 AM on the third day, it is updated to a mature positive sample; if no conversion feedback is generated by 10:00 AM on the third day, it is updated to a mature negative sample.
[0076] The joint learning network comprises a feature encoding layer, a shared representation network, multiple scenario-task-specific correction networks, and multiple task output layers. The feature encoding layer encodes user features, recommendation object features, and scenario context features, respectively. User features may include user historical behavior sequences, activity levels, and interest categories; recommendation object features may include object categories, text features, statistical features, and publication time; scenario context features may include scenario identifiers, access time, terminal type, and network status.
[0077] For categorical features, the feature encoding layer can use embedding vectors for encoding; for continuous features, normalization can be performed; for user historical behavior sequences, recurrent neural networks, convolutional neural networks, or attention networks can be used to extract sequence features. The processed features are then concatenated to obtain the input feature vector.
[0078] A shared representation network can consist of a multi-layer fully connected network or a hybrid expert network, used to extract shared features that can be used by multiple recommendation scenarios and tasks from the input feature vector. The network parameters of the shared representation network form a shared parameter set. To facilitate hierarchical calculation and consistency updates, the shared parameter set can be divided into multiple shared parameter subgroups according to the network layers or parameter blocks of the shared representation network, where the j-th shared parameter subgroup corresponds to the j-th network layer or parameter block in the shared representation network.
[0079] Each scenario-task specific correction network corresponds to a scenario-task combination and is used to extract behavioral patterns unique to that scenario-task combination, generating specific correction features. Shared features and specific correction features can be combined using vector addition or linear mapping after vector concatenation. This embodiment uses vector addition:
[0080]
[0081] in, Indicates shared features, Representing a scene Recommended Tasks Its proprietary correction features This indicates the combined task characteristics.
[0082] The task output layer outputs prediction results based on the combined task features. For binary classification tasks such as clicks, valid dwell times, and conversions, a fully connected layer and a sigmoid function can be used to output prediction probabilities between 0 and 1. The network parameters of the task output layer are updated using regular backpropagation based on the corresponding task loss; the update paths for shared parameter sets and proprietary parameter sets are determined by the update path control module.
[0083] During training, for each scenario-task combination, mature positive samples and mature negative samples are read from the corresponding effective training sample set. For the first... For each valid training sample, the task loss can be achieved using binary cross-entropy loss. Recommended Tasks The task loss is:
[0084]
[0085] in, This indicates the number of samples in the corresponding valid training sample set. Indicates the sample task label. This represents the task prediction result output by the joint learning network. Samples awaiting confirmation are not included. It is also not included in the calculation of task loss.
[0086] Perform backpropagation on the task loss and calculate its relative to the th element in the shared parameter set. Gradients of a shared parameter subgroup:
[0087]
[0088] in, Indicates the first A shared parameter subgroup, Representing a scene Recommended Tasks Regarding the first The parameter update vector of a shared parameter subgroup.
[0089] For scenario-task combinations Scene-Task Combination The updated path control module calculates the values of the two for the first step according to the following formula. Update consistency of shared parameter subgroups:
[0090]
[0091] in, To prevent the default positive number with a denominator of zero, it can be set to The higher the update consistency, the closer the update directions of the two scenario-task combinations are; when the update consistency is less than zero, it indicates that there is a conflict in the update directions of the two.
[0092] To avoid frequent update path switching caused by a single training batch, this embodiment uses an exponential moving average method to smooth the update consistency of multiple consecutive training batches:
[0093]
[0094] in, Indicates the first Consistency of stable updates for each training batch The smoothing coefficient can be set between 0.8 and 0.95; in this embodiment, it is set to 0.9.
[0095] Subsequently, the sample weights are determined based on the number of effective training samples in other scenario-task combinations. The stable update consistency between the current scenario-task combination and other scenario-task combinations is then weighted to obtain the comprehensive update consistency. Scenario-task combinations with a larger number of samples have higher weights, but a maximum weight can be set to prevent scenarios with exceptionally large datasets from completely dominating the update direction of shared parameters.
[0096] In this embodiment, the sharing threshold is set to 0.2, and the conflict threshold is set to -0.1. When the overall update consistency is not less than 0.2, the current scene-task combination is added to the shared update set, and the parameter update vectors of each scene-task combination in the shared update set are weighted and aggregated according to the number of effective training samples. The corresponding shared parameter subgroups are then updated using the aggregated parameter update vectors.
[0097] When the overall update consistency is no greater than -0.1, stop using the scenario-task combination to update the corresponding shared parameter subgroup, and use its task loss to update the corresponding scenario-task specific correction network, so that the scenario-task combination can be trained along the specific parameter path, thereby reducing the impact of conflict updates on other scenarios and tasks.
[0098] When the overall update consistency is between -0.1 and 0.2, for scenario-task combinations where a parameter update path already exists, the parameter update path from the previous training cycle is maintained to avoid repeated path switching; for scenario-task combinations where a parameter update path from the previous training cycle does not exist, the dedicated parameter path is used for isolated training by default. The sharing threshold and conflict threshold can be adjusted based on the model stability on the validation set, but the conflict threshold should be kept lower than the sharing threshold.
[0099] When the number of valid training samples for a scenario-task combination is less than a preset threshold, that scenario-task combination will not participate in the consistency calculation or be used to update the shared parameter set. For scenario-task combinations that already have parameter update paths, the system saves the parameter update path from the previous training cycle and resumes training once the number of valid training samples reaches the threshold.
[0100] For newly launched scenario-task combinations that lack a parameter update path from the previous training cycle, the output of the corresponding scenario-task-specific correction network is initialized to zero. This ensures that recommendation predictions primarily rely on shared features generated by the shared representation network, with the corresponding task output layer producing the task prediction results. Before the number of effective training samples reaches a preset threshold, updates to the shared parameter set and corresponding specific parameter set using this new scenario-task combination are paused to prevent unstable updates caused by a small number of samples. Once the sample count reaches the threshold, the overall update consistency with existing scenario-task combinations is calculated, and either the shared parameter path or the specific parameter path is selected accordingly.
[0101] During the online recommendation phase, the recommendation output module receives user features, candidate recommendation object features, and scene context features for the target recommendation scenario, and obtains the input feature vector using the feature encoding layer. A shared representation network generates shared features, and a dedicated correction network for each recommendation task in the target recommendation scenario generates dedicated correction features. These features are then passed through the output layers of each task to obtain the click probability, effective dwell probability, and conversion probability, respectively.
[0102] For target recommendation scenarios, a recommendation score is calculated based on preset or trained task fusion weights:
[0103]
[0104] in, Indicates user In the scene Below are the candidate recommendation objects Recommended score, This represents the prediction result for recommendation task k. Representing a scene Recommended Tasks The system assigns task fusion weights, with the sum of all task fusion weights being 1. Candidate recommendation objects are sorted from highest to lowest recommendation score, and a predetermined number of top-ranked candidates are selected to generate a recommendation list, which is then sent to the client via the recommendation result output interface.
[0105] After the recommendation list is output, the feedback update module continuously receives actual task feedback events. For samples awaiting confirmation, if a task feedback event is generated within the corresponding feedback maturity period, it is updated as a mature positive sample; if no task feedback event is generated by the end of the feedback maturity period, it is updated as a mature negative sample.
[0106] For cases where a sample has been identified as a mature negative sample but subsequently generates a task feedback event, if the task feedback event occurs within the preset maximum correction period, the sample will be corrected from a mature negative sample to a mature positive sample. For example, if a conversion sample is identified as a mature negative sample 48 hours after exposure, but generates a conversion feedback 60 hours after exposure, and the preset maximum correction period is 7 days, then the sample will be corrected to a mature positive sample.
[0107] Samples whose feedback status changes are written to the supplementary training buffer. When the number of samples in the supplementary training buffer reaches a preset supplementary sample quantity threshold, such as 5000, or when a preset incremental training time is reached, such as at the top of each hour, replay samples are extracted from the effective training sample set of each scenario-task combination, and the replay samples and the samples in the supplementary training buffer are combined to form an incremental training batch. The ratio of replay samples to supplementary samples can be set to 1:1 or 3:1 to reduce the sudden changes in training data distribution caused by concentrated supplementation of late feedback.
[0108] The feedback maturity time can be recalculated daily. During recalculation, historical interaction samples that have completed the observation period within the last 30 days are selected as the rolling time window data. When the difference between the recalculated feedback maturity time and the current feedback maturity time reaches a preset update threshold, the feedback maturity time of the corresponding scenario-task combination is updated, and the feedback status of samples that have not generated feedback events within the rolling time window is reassessed. For samples whose status changes due to changes in feedback maturity time, they are written to the replenishment training buffer and incremental training is triggered.
[0109] Through the above methods, this embodiment can dynamically determine the sample maturity status according to the actual feedback speed of different recommendation scenarios and recommendation tasks, reducing the situation where unconfirmed samples are used as negative samples in training in advance; at the same time, it can dynamically select shared training or dedicated training paths according to the parameter update direction of different scenario-task combinations, thereby improving the stability of the multi-scenario multi-task joint learning process and enabling the joint learning network to continuously adapt to changes in feedback delay distribution and user behavior patterns.
[0110] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A multi-task joint learning recommendation system for multiple scenarios, characterized in that, It includes an interactive data acquisition interface, a memory, a processor, and a recommendation result output interface. The memory stores historical interaction samples from multiple recommendation scenarios and network parameters of the joint learning network. The processor is configured with: The sample construction module is used to acquire historical interaction data in multiple recommendation scenarios through the interaction data acquisition interface, associate the exposure event of the recommended object with multiple task feedback events that occur after the exposure event, and construct historical interaction samples corresponding to each recommendation scenario and each recommendation task. The feedback maturity duration determination module is used to determine the feedback maturity duration for each scenario-task combination based on the feedback delay from the exposure event to the corresponding task feedback event in the historical interaction samples of the completed observation period. The sample state determination module is used to determine the feedback state of historical interaction samples for each recommendation task based on the current training time, the exposure time of historical interaction samples, and the feedback maturity time of the corresponding scene-task combination. Specifically, historical interaction samples that have generated corresponding task feedback events are determined as mature positive samples, historical interaction samples that have not generated corresponding task feedback events but have reached the corresponding feedback maturity time after exposure are determined as mature negative samples, and historical interaction samples that have not generated corresponding task feedback events but have not reached the corresponding feedback maturity time after exposure are determined as samples to be confirmed. The mature positive samples and the mature negative samples are combined into a valid training sample set. The joint learning module is used to construct a joint learning network that includes a shared parameter set and a dedicated parameter set corresponding to each scenario-task combination. It calculates the task loss based on the effective training sample set of each scenario-task combination and obtains the parameter update vector of each scenario-task combination for the shared parameter set based on the task loss. The update path control module is used to calculate the update consistency between parameter update vectors of different scenario-task combinations, determine the parameter update path of the corresponding scenario-task combination based on the update consistency, use the parameter update vectors that meet the sharing conditions to update the shared parameter group, and use the task loss corresponding to the scenario-task combination that meets the conflict isolation conditions to update the corresponding proprietary parameter group. The recommendation output module is used to obtain user features, candidate recommendation object features and scene context features in the target recommendation scenario, obtain prediction results of multiple recommendation tasks through the updated joint learning network, determine the recommendation score of the candidate recommendation object based on the prediction results of the multiple recommendation tasks, generate a recommendation list according to the recommendation score, and output the recommendation list through the recommendation result output interface. The feedback update module is used to obtain the actual task feedback events generated after the recommendation list is output, update the samples to be confirmed and the historical interaction samples with confirmed feedback status according to the actual task feedback events, and redetermine the feedback maturity time of each scenario-task combination according to the preset update cycle.
2. The multi-task joint learning recommendation system for multiple scenarios according to claim 1, characterized in that, The historical interaction data includes at least user characteristics, recommendation object identifier, recommendation object characteristics, scene identifier, scene context characteristics, exposure time, task feedback event type, and task feedback event occurrence time; The sample construction module is used to associate task feedback events that occur within a preset associated time period with the corresponding exposure events based on the exposure event identifier, user identifier, and recommended object identifier, forming a chain of scene interaction events arranged according to the time of occurrence. The recommended task includes at least two of the following: click prediction task, effective stay prediction task, favorite prediction task, completion prediction task, and conversion prediction task. The historical interaction samples that have completed the observation period are historical interaction samples whose time distance from the exposure time to the sample statistics time reaches the preset maximum observation time.
3. The multi-task joint learning recommendation system for multiple scenarios according to claim 2, characterized in that, The feedback maturity time determination module is used to determine the scenario. Recommended tasks Based on historical interaction samples that have completed the observation period and generated corresponding task feedback events, a feedback delay set is determined, and a cumulative distribution of feedback delays is established based on the feedback delay set. The feedback maturation time is determined according to the following formula: in, Representing a scene Recommended Tasks Feedback maturity time This indicates a delay in feedback. This indicates that the feedback delay is no greater than The proportion of the number of task feedback events to the total number of corresponding task feedback events. This indicates the preset feedback coverage ratio, and ; When the scene Recommended Tasks When the number of corresponding task feedback events is less than a preset sample size threshold, the recommendation task is used. The cumulative distribution of feedback delays in other recommendation scenarios is used to determine the task-level feedback maturity time of recommendation task k, and the task-level feedback maturity time is used as the scenario. Recommended Tasks Feedback maturation time.
4. The multi-task joint learning recommendation system for multiple scenarios according to claim 3, characterized in that, The sample state determination module is used to determine the first according to the following formula. A historical interaction sample for the scenario Recommended Tasks Feedback maturity: in, Indicates the first A historical interaction sample for the scenario Recommended Tasks Feedback maturity Indicates the current training moment. Indicates the first A historical interaction sample in the scene The exposure time below; For historical interaction samples that did not generate corresponding task feedback events, when When, the historical interaction samples are identified as mature negative samples, when When the historical interaction sample is confirmed, it is identified as a sample to be confirmed; for historical interaction samples that have generated corresponding task feedback events, the historical interaction sample is identified as a mature positive sample, which is not limited by the feedback maturity.
5. The multi-task joint learning recommendation system for multiple scenarios according to claim 4, characterized in that, The joint learning network includes a feature encoding layer, a shared representation network, multiple scene-task specific correction networks, and multiple task output layers; the feature encoding layer is used to convert user features, recommendation object features, and scene context features into input feature vectors; the shared representation network is used to generate shared features shared by multiple recommendation scenes and multiple recommendation tasks based on the input feature vectors; Each scenario-task-specific correction network corresponds to a scenario-task combination and is used to generate specific correction features for the corresponding scenario-task combination based on the input feature vector. Each task output layer corresponds to a recommendation task and is used to output the prediction result of the recommendation task based on the shared features and the specific correction features of the corresponding recommendation task in the current recommendation scenario. The network parameters of the shared representation network form the shared parameter group, and the network parameters of each scenario-task specific correction network form the specific parameter group of the corresponding scenario-task combination.
6. The multi-task joint learning recommendation system for multiple scenarios according to claim 5, characterized in that, The joint learning module is used to perform the following processing for each scenario-task combination: read mature positive samples and mature negative samples from the corresponding effective training sample set, set the task label of the mature positive samples to 1, set the task label of the mature negative samples to 0, and prevent the samples to be confirmed from participating in the task loss calculation of the corresponding recommendation task; calculate the task loss based on the task label and the prediction result output by the corresponding task output layer, and normalize the task loss according to the number of samples in the effective training sample set; calculate the gradient of the normalized task loss relative to each shared parameter group, and use the gradient as the parameter update vector of the corresponding scenario-task combination for the corresponding shared parameter group. When the number of effective training samples for a scenario-task combination is less than a preset effective sample count threshold, the use of the scenario-task combination's parameter update vector to update the shared parameter set is paused. For scenario-task combinations that already have a parameter update path, the parameter update path from the previous training cycle is maintained. For new scenario-task combinations that do not have a parameter update path from the previous training cycle, before the number of effective training samples reaches the preset effective sample count threshold, the output of the corresponding scenario-task-specific correction network is initialized to zero, the shared representation network is called to generate the task prediction result, and the use of the new scenario-task combination to update the shared parameter set and the corresponding specific parameter set is paused.
7. The multi-task joint learning recommendation system for multiple scenarios according to claim 6, characterized in that, The update path control module is used for shared parameter groups. Calculate the scenario-task combination according to the following formula. Scene-Task Combination Update consistency: in, Representing a scenario-task combination Scene-Task Combination For shared parameter groups Update consistency, and These represent the scene-task combination respectively. Scene-Task Combination For shared parameter groups The parameter update vector, Indicates a preset positive number; The update path control module is also used to determine the stable update consistency based on the update consistency of multiple consecutive training batches, and to weight the stable update consistency between the current scenario-task combination and other scenario-task combinations based on the number of samples in each effective training sample set, so as to obtain the comprehensive update consistency of the current scenario-task combination. When the overall update consistency is not less than the shared threshold, the current scenario-task combination is added to the shared update set, and the corresponding shared parameter group is updated using the parameter update vectors of each scenario-task combination in the shared update set; when the overall update consistency is not greater than the conflict threshold, the update of the corresponding shared parameter group using the parameter update vectors of the current scenario-task combination is stopped, and the corresponding proprietary parameter group is updated using the task loss of the current scenario-task combination; when the overall update consistency is between the conflict threshold and the shared threshold, the parameter update path of the previous training cycle is maintained for scenario-task combinations that already have parameter update paths, and the update of the corresponding shared parameter group is stopped for new scenario-task combinations that do not have parameter update paths in the previous training cycle, and the corresponding proprietary parameter group is updated using its task loss; wherein, the conflict threshold is less than the shared threshold.
8. The multi-task joint learning recommendation system for multiple scenarios according to claim 7, characterized in that, The recommendation output module is used to combine the shared features output by the shared representation network with the specific modified features corresponding to each recommendation task in the target recommendation scenario, obtain the prediction results of each recommendation task through the corresponding task output layer, fuse the prediction results of each recommendation task according to the task fusion weight corresponding to the target recommendation scenario, obtain the recommendation score of the candidate recommendation object, and select a preset number of candidate recommendation objects with the highest recommendation scores to generate a recommendation list. The feedback update module is used to update the sample to be confirmed as a mature positive sample when a task feedback event is generated within the corresponding feedback maturity period of the sample to be confirmed; and to update the sample to be confirmed as a mature negative sample when no task feedback event is generated after the corresponding feedback maturity period has been reached. For historical interaction samples that have been identified as mature negative samples but generated task feedback events within the preset maximum correction period, they are corrected to mature positive samples. The feedback update module is also used to write historical interaction samples whose feedback status has changed into the back-up training buffer. When the number of samples in the back-up training buffer reaches the preset back-up batch number or reaches the preset incremental training time, replay samples are extracted from the effective training sample set of each scenario-task combination, and the replay samples and historical interaction samples in the back-up training buffer are used together for the incremental training of the joint learning network. The feedback update module is also used to redetermine the feedback maturity time based on the updated historical interaction samples within a preset rolling time window. When the difference between the redetermined feedback maturity time and the current feedback maturity time reaches a preset update threshold, the feedback maturity time of the corresponding scenario-task combination is updated, the feedback status of the historical interaction samples within the preset rolling time window is re-determined, and the incremental training of the joint learning network is triggered.
9. A multi-task joint learning recommendation method for multiple scenarios, characterized in that, Includes the following steps: S1. Obtain historical interaction data under multiple recommendation scenarios, associate the exposure event of the recommended object with multiple task feedback events that occur after the exposure event, and construct historical interaction samples corresponding to each recommendation scenario and each recommendation task; S2. For each scenario-task combination, determine the feedback maturity time of each scenario-task combination based on the feedback delay from the exposure event to the corresponding task feedback event in the historical interaction samples of the completed observation period. S3. Based on the current training time, the exposure time of historical interaction samples, and the feedback maturity time of the corresponding scene-task combination, the historical interaction samples that have generated corresponding task feedback events are determined as mature positive samples, the historical interaction samples that have not generated corresponding task feedback events but have reached the corresponding feedback maturity time after exposure are determined as mature negative samples, and the historical interaction samples that have not generated corresponding task feedback events but have not reached the corresponding feedback maturity time after exposure are determined as samples to be confirmed. The mature positive samples and the mature negative samples are combined into an effective training sample set. S4. Construct a joint learning network including a shared parameter set and a dedicated parameter set corresponding to each scenario-task combination. Calculate the task loss based on the effective training sample set of each scenario-task combination, and obtain the parameter update vector of each scenario-task combination for the shared parameter set based on the task loss. S5. Calculate the update consistency between parameter update vectors of different scenario-task combinations, determine the parameter update path of the corresponding scenario-task combination based on the update consistency, use the parameter update vector that meets the sharing condition to update the shared parameter group, and use the task loss corresponding to the scenario-task combination that meets the conflict isolation condition to update the corresponding proprietary parameter group. S6. Obtain user features, candidate recommendation object features, and scene context features in the target recommendation scenario. Obtain prediction results for multiple recommendation tasks through the updated joint learning network. Determine the recommendation score of the candidate recommendation object based on the prediction results of the multiple recommendation tasks. Generate a recommendation list according to the recommendation score and output the recommendation list. S7. Obtain the actual task feedback events generated after the recommendation list is output. When a task feedback event occurs within the feedback maturity time corresponding to the sample to be confirmed, update the sample to be confirmed as a mature positive sample. If no task feedback event occurs after the corresponding feedback maturity time has been reached, update the sample to be confirmed as a mature negative sample. For historical interaction samples that have been determined to be mature negative samples but have generated task feedback events within the preset maximum correction time, correct them to mature positive samples. Use the historical interaction samples whose feedback status has changed for incremental training of the joint learning network, and redetermine the feedback maturity time of each scenario-task combination according to the preset update cycle.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method described in claim 9.