Intelligent data recommendation method adopting data twinning and federal learning
Through the method of data twin combined with federated learning, the recommendation problem of intelligent recommendation system when new users and new content is cold-started, the effective recommendation and sorting model of new items are optimized, and the overall performance of the recommendation system is improved.
Patent Information
- Application Number
- CN202411406415.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2025-02-07
AI Technical Summary
The existing intelligent recommendation system is difficult to effectively recommend new users and new content when they are coldly launched, which has affected the healthy development of the user experience and content supply ecosystem.
Using the method of data twins combined with federated learning, a new item cold start sorting strategy is established, and the sorting model is optimized through multiple iterations, combined with the item cold start quantity guarantee mechanism to achieve intelligent data recommendation.
It improves the recommendation effect of new items, optimizes the sorting model, ensures the accuracy of recommendation results, increases the proportion of cold-start items becoming popular items, and reduces the negative impact on the consumer side.
Smart Images
Figure CN119397091A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of intelligent recommendation technology, and specifically relates to a data intelligent recommendation method using data twins combined with federated learning. Background Art
[0002] The essence of an intelligent recommendation system is an information publishing tool. This information distribution system has the characteristics of personalized distribution. It can automatically connect users with products, explore users' personalized needs, and help users find products that interest them. Whether it is recall, sorting, or re-arrangement, the recommendation algorithm relies more on the user's historical behavior data. However, the users and content of any benign Internet information platform are constantly growing and changing, and the recommendation system will face a steady stream of new users and new content. How to recommend suitable content to new users and how to distribute new content produced by creators all fall into the category of recommendation cold start. System cold start refers to how to build a recommendation system in the early stages of a product's development and launch.
[0003] The cold start problem is a very important and tricky problem in the recommendation system. A good cold start recommendation experience involves many aspects of work such as product, operation, data, algorithm, engineering, etc. Cold start is one of the important challenges of the recommendation system, and it is also one of the important issues that the recommendation system needs to consider and solve. Many Internet products put recommendations at the core of the product homepage. For new users, the content of cold start recommendations determines the user's first impression of the product and is a key factor in user retention. For content producers, if the newly released content is not displayed and fed back in a timely and sufficient manner, it will affect the enthusiasm of the producers, thereby affecting the healthy development of the content supply ecology of the information platform. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a data intelligent recommendation method using data twins combined with federated learning in response to the above-mentioned shortcomings of the prior art.
[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:
[0006] A data intelligent recommendation method using data twins combined with federated learning, comprising:
[0007] Establish a cold-start sorting strategy for new items, and use federated learning technology to iterate the fusion sorting model of the cold-start item queue and the normal item queue multiple times to achieve optimization and customized management and control of the sorting model;
[0008] A cold-start quantity preservation mechanism for items is established to calculate the process of the cold-start quantity preservation mechanism for the sorted recommendation data, and a digital twin model of the cold-start quantity preservation mechanism process is mapped to achieve data intelligent recommendation using data twins combined with federated learning.
[0009] To optimize the above technical solutions, the specific measures taken also include:
[0010] The above-mentioned new item cold start sorting strategy is as follows: the distribution chain of the recommendation system for cold start items includes recall, refinement and re-sorting;
[0011] In the rough and fine sorting stages, for the sorting model, cold start items can share the same model with normal items for scoring estimation, and an independent sorting model can be built for cold start items. When multi-objective fusion of model output is performed, firstly, independent fusion formulas and parameters are used to perform fusion scoring on cold start items. Secondly, the cold start content is sorted internally in combination with the quantity-preserving display speed control coefficient. Finally, the cold start items are sorted internally according to the product of the fusion score and the display speed control coefficient. This is the quantity-preserving display speed control mechanism for cold starts.
[0012] In the reordering phase, the cold start item queue and the natural item queue are sorted and merged: the sorted head item in the cold start queue is forcibly inserted into the specified position of the normal item queue.
[0013] The above-mentioned federated learning technology is specifically as follows: in the Nth training iteration, the user terminal performs local training based on the global model downloaded from the FL server using locally collected images / videos, and then the terminal reports the intermediate training results of the Nth iteration to the FL server through the 5G line channel. At the same time, the FL server sends the (N+1)th global model and training configuration to the user terminal. When the FL server aggregates the gradients collected in the Nth iteration, the terminal can perform the (N+1)th iteration training at the same time, and the federated aggregation outputs an updated global model for the next iteration training, and distributes the model together with the updated training configuration to the terminal.
[0014] The above-mentioned sorted recommendation data is used to calculate the process of the cold start quantity preservation mechanism of the item, including: the differentiated quantity preservation module calculates and allocates the quantity preservation budget to the item according to the current performance of the item; the quantity preservation display speed control module completes the optimal allocation of traffic for the given traffic budget; the loss calculation module records the actual exposure of the item and the consumption loss income, and assists the differentiated quantity preservation module to complete training and the next allocation.
[0015] The above-mentioned digital twin model uses computers to simulate the input and output control variables of the production line, as well as the impact of real-time combination changes in the control parameters of the simulation production line on the final product. The mapping process includes preparatory work and data collection, data feature engineering, model training and verification, model trial operation and optimization, and model online deployment.
[0016] The above-mentioned cold-start quantity preservation mechanism for items judges the quality of the items based on their basic characteristics when new items are generated, predicts the probability that the item will become popular in the future, and performs differentiated quantity preservation according to the probability value, giving "good" items a higher quantity preservation and "bad" items a lower quantity preservation. It makes full use of the quantity preservation traffic, increases the proportion of cold-start items that become popular items, and reduces the negative impact of quantity preservation on the consumer side.
[0017] The above-mentioned cold start volume retention mechanism estimates the probability of new items becoming popular in the future by training a high-heat prediction model. The model uses the exposure results of historical cold start items as samples, statistical features and content understanding features on the item side, as well as basic features and statistical features on the author side. It uses the full amount of data from the most recent period for offline prediction, obtains the distribution range of the predicted values, and performs differentiated volume retention traffic allocation online based on the distribution range of the predicted values, and calculates the high, medium and low thresholds of differentiated volume retention to ensure that the overall volume retention traffic on the day is basically stable.
[0018] The above-mentioned cold start quantity preservation mechanism supports the adjustment of the quantity preservation threshold strategy. The final quantity preservation threshold of the cold start item is determined by the model and strategy.
[0019] The above-mentioned cold start quantity guarantee mechanism for items adopts a multi-round quantity guarantee mechanism. During the distribution process, it uses the posterior action rate of the items to adjust the quantity guarantee value in real time to improve the overall cold start distribution quality; and optimizes the leverage efficiency of cold start traffic, including: estimating the quantity guarantee value of each item through a high-heat prediction model; the high-heat prediction model adds more content understanding features in the model iteration to reduce dependence on the author's historical data; and designs a pulse stimulation strategy for users who have not been "popular" to increase the quantity guarantee threshold.
[0020] The above-mentioned cold start quantity preservation mechanism for items can achieve the exposure quantity preservation target relatively evenly within the cold start time period for cold start items with a given quantity preservation threshold. Its quantity preservation display speed control adopts the Budget Pacing mechanism, which distributes the quantity preservation display target within the quantity preservation period according to the actual traffic curve of the platform, and uses the PID calculation method to dynamically adjust the display control speed coefficient according to the specific exposure situation, affecting the ranking of items in the recommendation system and the final exposure probability, so that the items are evenly distributed within the quantity preservation period and the quantity preservation target is achieved.
[0021] The present invention has the following beneficial effects:
[0022] The present invention aims to improve the problem of insufficient new item behavior. First, a cold-start sorting strategy for new items is proposed. Secondly, a real-time federated learning process is used to perform multiple iterations on the sorting model that sorts the cold-start item queue and the natural item queue, thereby completing the optimization and customized management and control of the sorting model.
[0023] The present invention calculates the process of the cold-start quantity-guarantee mechanism for items on the recommended data after fine-tuning to ensure the accuracy of the recommendation results; based on the digital model of machine learning, a digital twin model of the cold-start quantity-guarantee mechanism process for physical items is mapped as realistically as possible, to achieve data intelligent recommendation using data twins combined with federated learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of the method of the present invention;
[0025] Figure 2 Schematic diagram of the GME model for intelligent recommendation;
[0026] Figure 3 For real-time federated learning process;
[0027] Figure 4 Flow chart of the cold start guarantee mechanism for items;
[0028] Figure 5 The main steps of the digital twin model based on machine learning. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.
[0031] like Figure 1 As shown, a data intelligent recommendation method using data twins combined with federated learning of the present invention includes:
[0032] Step 1: Establish a cold-start sorting strategy for new items, and use federated learning technology to iterate the fusion sorting model of the cold-start item queue and the normal item queue multiple times to achieve optimization and customized management and control of the sorting model;
[0033] The recommendation cold start problem can be divided into new user cold start and new content cold start.
[0034] In fact, the cold start recall method of items is still based on the conventional recall method, which proposes improvements to the problem of insufficient new item behavior. This type of solution will cause the cold start exposure to be concentrated on the top items, making the top items appear too fast and difficult to control, which is prone to over-issuance and squeezes the overall cold start traffic of the system.
[0035] First, the present invention implements a cold-start sorting strategy for new items. Second, a real-time federated learning process is used to perform multiple iterations of the sorting model that sorts the cold-start item queue and the natural item queue, thereby completing the optimization and customized management and control of the sorting model. Figure 2 shown.
[0036] (1) New item cold start sorting strategy. The distribution of cold start items by the recommendation system is similar to that of normal items in the chain, and both need to go through the steps of recall, fine sorting, and re-sorting. This patent introduces the differences between cold start items and normal items in the sorting stage.
[0037] In the rough and fine ranking stages, for the sorting model, cold start items can share the same model with normal items for scoring estimation, or an independent sorting model can be built for cold start items. When multi-target fusion is performed on the model output, the cold start content will be quite different from the normal content. First of all, since the posterior indicators of cold start items such as Ctr distribution and global distribution are usually different, it is more reasonable to use independent fusion formulas and parameters for fusion scoring. Secondly, the internal sorting of cold start content also needs to be combined with the coefficient of quantity preservation and display speed control. Finally, the internal sorting of cold start items is carried out according to "fusion score x display speed control coefficient", and the quantity preservation and display speed control mechanism of cold start. In order to ensure relatively stable cold start traffic in the recommendation system, the cold start item queue and the natural item queue will be sorted and fused in the re-ranking stage: the head item in the cold start queue after sorting is usually forced to be inserted into the specified position of the normal item queue.
[0038] (2) A real-time federated learning process is used to perform multiple iterations on the sorting model that sorts the cold-start item queue and the natural item queue, thereby completing the optimization and customized management and control of the sorting model.
[0039] The real-time federated learning process is as follows Figure 3As shown, in the Nth training iteration, the user terminal performs local training using locally collected images / videos based on the global model downloaded from the FL server. The terminal then reports the intermediate training results of the Nth iteration (such as the gradient of the DNN) to the FL server through the 5G line channel. At the same time, the FL server sends the (N+1)th global model and training configuration to the user terminal. While the FL server aggregates the gradients collected in the Nth iteration, the terminal can perform training for the (N1)th iteration. Federated Aggregation outputs an updated global model for the next iteration training, and distributes the model to the terminal together with the updated training configuration.
[0040] In order to make full use of the training resources on the user terminal and minimize the training delay, Figure 3 The training sequence shown requires that the training result report of the (N-1)th iteration and the global model distribution and training parameter configuration of the (N+1)th iteration be completed during the training process of the UE's Nth iteration. Of course, in the actual training process, it is also possible to consider using a relaxed sequence, at the cost of sacrificing the convergence speed of the training, that is, performing non-real-time federated learning.
[0041] Step 2: Establish a cold-start quantity preservation mechanism for items. Perform calculations on the process of the cold-start quantity preservation mechanism for the sorted recommendation data, and map the digital twin model of the cold-start quantity preservation mechanism process to achieve data intelligent recommendation using data twins combined with federated learning.
[0042] To ensure the accuracy of the recommended data after fine sorting, the cold start quantity guarantee mechanism process of the item needs to be calculated. Then, based on the digital model of machine learning, the digital twin model of the cold start quantity guarantee mechanism process of the physical item is mapped as realistically as possible. The digital twin fully utilizes the data of physical models, sensor updates, operation history, etc., integrates multi-disciplinary, multi-physical quantity, multi-scale, and multi-probability simulation processes, completes the mapping in the virtual space, and thus reflects the full life cycle process of the corresponding physical equipment.
[0043] The purpose of simulating differentiated quantity preservation first is to judge the quality of new items based on their basic characteristics when they are produced, and to predict the probability that the items will become popular in the future. Then, the display speed control is performed to maintain the quantity. The key point is how to calculate the display speed control coefficient based on the error between the current expected cumulative exposure and the actual cumulative exposure of the cold start item. In the case of over-issuance of items, the display speed control coefficient should be less than 1, and in the case of insufficient exposure, the display speed control coefficient should be greater than 1.
[0044] (1) To ensure the accuracy of the recommended results, the recommended data after fine sorting needs to be calculated through the cold start quantity guarantee mechanism process. At present, the mainstream solution of the leading companies in the industry is to iterate based on the quantity guarantee mechanism as the main line. The flowchart of the cold start quantity guarantee mechanism is as follows: Figure 4 The differentiated volume-keeping module calculates and allocates a volume-keeping budget to items based on their current performance. The volume-keeping display speed control module completes the optimal traffic allocation for a given traffic budget. The loss calculation module records the actual exposure of items and consumption loss income, helping the differentiated volume-keeping module complete training and the next allocation.
[0045] (2) A digital twin model based on machine learning that maps the cold start quality assurance mechanism process of items as realistically as possible. This method uses a computer to simulate the input and output control variables of the production line, as well as the impact of changes in the real-time combination of control parameters of the production line on the final product. This method has five main steps, including preparatory work and data collection, data feature engineering, model training and verification, model trial operation and optimization, and model online deployment, such as Figure 5 shown.
[0046] (3) The purpose of differentiated quantity retention is to judge the quality of new items based on their basic characteristics when they are produced, and to predict the probability that the item will become popular in the future. Differentiated quantity retention is performed based on the probability value, giving "good" items a higher quantity retention and "bad" items a lower quantity retention, making full use of the quantity retention flow, thereby increasing the proportion of cold start items becoming popular items and reducing the negative impact of quantity retention on the consumer side.
[0047] Usually, the probability of a new item becoming popular in the future can be estimated by training a high-heat prediction model. The model can use the exposure results of historical cold-start items as samples. For example, the exposure two days after the item is released is used. Exposures greater than 10,000 are positive examples, and exposures less than 1,000 are negative examples. In terms of features, statistical features and content understanding features on the item side, as well as basic features and statistical features on the author side can be used. The model will use the full amount of data from the recent period of time offline for prediction and obtain the distribution range of the predicted values. Online, differentiated volume-keeping traffic allocation can be performed based on the distribution range of the predicted values, and differentiated high, medium and low-end thresholds for volume-keeping can be calculated to ensure that the overall volume-keeping traffic on the day is basically stable. At the same time, the differentiated volume-keeping service also needs to support the adjustment of volume-keeping threshold-related strategies, so the final volume-keeping threshold of the cold-start item is determined by the model and strategy.
[0048] There are some long-term optimization directions for differentiated volume maintenance:
[0049] The first is the multi-round quantity guarantee mechanism. In the above scheme, once the quantity guarantee threshold is determined, it will not be changed. If the item is estimated to be "hot", but the actual performance is poor, it will cause the consumption index to be damaged; for items that are estimated not to be "hot" but should be "hot", it will cause cold start failure. You can consider using a multi-round quantity guarantee method, using the posterior action rate of the item during the distribution process, and adjusting the quantity guarantee value in real time, thereby improving the overall cold start distribution quality.
[0050] The second is to optimize the leverage efficiency of cold start traffic. For very high-quality or very low-quality items, there is actually no need to guarantee the volume, or the guaranteed volume traffic can be extremely low. The model can be used to estimate the guaranteed volume value of each item. For example, if item A is given x guaranteed volume traffic, the final traffic is 2x, and if 2x guaranteed volume traffic is given, the final traffic reaches 50x; while item B is given x guaranteed volume traffic, the final traffic reaches 100x, and if 2x guaranteed volume traffic is given, it reaches 105x. The value of 2x guaranteed volume traffic given to item A will be greater than that given to B. The main features of the high-heat prediction model are all statistical features on the author side, which will cause an author to have a "hot" item, and the probability of the items released later being estimated to be "hot" will increase, which is equivalent to the guaranteed volume traffic will tilt towards the head. Therefore, more content understanding features need to be added in the model iteration to reduce the dependence on the author's historical performance. Or design a pulse stimulation strategy for users who have never been "hot" to increase the guaranteed volume threshold so that every creator has the opportunity to "become popular".
[0051] (4) To maintain the quantity and control the display speed, the core point is how to calculate the display speed control coefficient based on the error between the current expected cumulative exposure and the actual cumulative exposure of the cold start item. In the case of over-issuance of the item, the display speed control coefficient should be less than 1. In the case of insufficient exposure, the display speed control coefficient should be greater than 1.
[0052] It is hoped that for cold-start items with a given quantity preservation threshold, their exposure and quantity preservation targets can be achieved relatively evenly within the cold-start time period. The benefit of doing so is to improve the utilization rate of cold-start samples, ensure the publishing feedback experience on the author side, and avoid the loss of consumption indicators due to poor action rates caused by concentrated exposure. From the perspective of application technology, quantity preservation display speed control is similar to the Budget Pacing mechanism in the field of computational advertising. The quantity preservation display target is allocated within the quantity preservation period according to the actual traffic curve of the platform, and the display speed control coefficient is dynamically adjusted according to the specific exposure situation, which affects the ranking of items in the recommendation system and the final exposure probability, so that the items are evenly distributed within the quantity preservation period and the quantity preservation target is achieved. The core point here is how to calculate the display speed control coefficient based on the error between the current expected cumulative exposure and the actual cumulative exposure of the cold-start item. In the case of over-issuance of items, the display speed control coefficient should be less than 1, and in the case of insufficient exposure, the display speed control coefficient should be greater than 1. A typical control algorithm is the PID control algorithm. The PID (Proportion Integration Differentiation) algorithm is a control algorithm based on proportion, integration, and differentiation. The calculation method of PID is shown in the following formula. The first part is the proportional term, which takes the current error into consideration; the second part is the integral term, which takes the system steady-state error into consideration; the last part is the differential term, which is used to reduce oscillations during the control process.
[0053]
[0054] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
[0055] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A data intelligent recommendation method using data twins combined with federated learning, characterized in that: include: Establish a cold-start sorting strategy for new items, and use federated learning technology to iterate the fusion sorting model of the cold-start item queue and the normal item queue multiple times to achieve optimization and customized management and control of the sorting model; A cold-start quantity preservation mechanism for items is established to calculate the process of the cold-start quantity preservation mechanism for the sorted recommendation data, and a digital twin model of the cold-start quantity preservation mechanism process is mapped to achieve data intelligent recommendation using data twins combined with federated learning.
2. According to claim 1, a data intelligent recommendation method using data twins combined with federated learning is characterized in that: The new item cold start sorting strategy is: the distribution link of the recommendation system for cold start items includes recall, refinement and re-sorting; In the rough and fine sorting stages, for the sorting model, cold start items can share the same model with normal items for scoring estimation, and an independent sorting model can be built for cold start items. When multi-objective fusion of model output is performed, firstly, independent fusion formulas and parameters are used to perform fusion scoring on cold start items. Secondly, the cold start content is sorted internally in combination with the quantity-preserving display speed control coefficient. Finally, the cold start items are sorted internally according to the product of the fusion score and the display speed control coefficient. This is the quantity-preserving display speed control mechanism for cold starts. In the reordering phase, the cold start item queue and the natural item queue are sorted and merged: the sorted head item in the cold start queue is forcibly inserted into the specified position of the normal item queue.
3. According to claim 1, a data intelligent recommendation method using data twins combined with federated learning is characterized in that: The federated learning technology is specifically as follows: in the Nth training iteration, the user terminal performs local training based on the global model downloaded from the FL server using the locally collected images / videos, and then the terminal reports the intermediate training result of the Nth iteration to the FL server through the 5G line channel. At the same time, the FL server sends the (N+1)th global model and training configuration to the user terminal. When the FL server aggregates the gradients collected in the Nth iteration, the terminal can perform the (N+1)th iteration training, and the federated aggregation outputs an updated global model for the next iteration training, and distributes the model together with the updated training configuration to the terminal.
4. According to claim 1, a data intelligent recommendation method using data twins combined with federated learning is characterized in that: The sorted recommendation data is used to calculate the process of the cold start quantity preservation mechanism for items, including: a differentiated quantity preservation module calculates and allocates a quantity preservation budget to items according to their current performance; a quantity preservation display speed control module completes the optimal allocation of traffic for a given traffic budget; a loss calculation module records the actual exposure of items and consumption loss benefits, and assists the differentiated quantity preservation module in completing training and the next allocation.
5. According to claim 1, a data intelligent recommendation method using data twins combined with federated learning is characterized in that: The digital twin model simulates the input and output control variables of the production line through computers, as well as the impact of changes in the real-time combination of control parameters of the production line on the final product. The mapping process includes preparatory work and data collection, data feature engineering, model training and verification, model trial operation and optimization, and model online deployment.
6. According to claim 1, a data intelligent recommendation method using data twins combined with federated learning is characterized in that: When a new item is generated, the item cold start quantity guarantee mechanism judges the item quality based on the basic characteristics of the item, predicts the probability that the item will become popular in the future, and performs differentiated quantity guarantees based on the probability value, giving "good" items a higher quantity guarantee and "bad" items a lower quantity guarantee, thereby making full use of the quantity guarantee flow, increasing the proportion of cold start items that become popular items, and reducing the negative impact of quantity guarantee on the consumer side.
7. The data intelligent recommendation method using data twins combined with federated learning according to claim 6 is characterized in that: The item cold start volume preservation mechanism estimates the probability that a new item will become popular in the future by training a high-heat prediction model. The model uses the exposure results of historical cold start items as samples, statistical features and content understanding features on the item side, and basic features and statistical features on the author side. It uses the full data of the most recent period for offline prediction, obtains the distribution range of the predicted values, and performs differentiated volume preservation traffic allocation online based on the distribution range of the predicted values, calculates the high, medium and low thresholds of differentiated volume preservation, and ensures that the overall volume preservation traffic on the day is basically stable.
8. The data intelligent recommendation method using data twins combined with federated learning according to claim 7 is characterized in that: The item cold start quantity preservation mechanism supports the adjustment of the quantity preservation threshold strategy, and the final quantity preservation threshold of the cold start item is determined by the model and the strategy.
9. The data intelligent recommendation method using data twins combined with federated learning according to claim 7 is characterized in that: The item cold start quantity guarantee mechanism adopts a multi-round quantity guarantee mechanism, which uses the item's posterior action rate during the distribution process to adjust the quantity guarantee value in real time to improve the overall cold start distribution quality; And optimize the leverage efficiency of cold start traffic, including: estimating the value of each item through the high-heat prediction model; adding more content understanding features to the high-heat prediction model in the model iteration to reduce dependence on the author's historical data; designing a pulse stimulation strategy for users who have not been "popular" to increase the threshold for volume preservation.
10. The data intelligent recommendation method using data twins combined with federated learning according to claim 7, characterized in that: The item cold start quantity preservation mechanism achieves the exposure quantity preservation target relatively evenly within the cold start time period for cold start items with a given quantity preservation threshold; its quantity preservation display speed control adopts the Budget Pacing mechanism, distributes the quantity preservation display target within the quantity preservation period according to the actual platform traffic curve, and dynamically adjusts the display speed control coefficient according to the specific exposure situation by using the PID calculation method, which affects the ranking of items in the recommendation system and the final exposure probability, so that the items are evenly distributed within the quantity preservation period and the quantity preservation target is achieved.