Recommendation system cold start methods and devices, electronic equipment and storage media

Abstract

This application provides a cold start method, apparatus, electronic device, and storage medium for a recommendation system. The method includes: determining a list of historical clicked content corresponding to user tags based on acquired user tracking data and user tags; calculating feature data of the historical clicked content based on the user tracking data; smoothing the feature data to calculate the click-through rate (CTR) data corresponding to the historical clicked content, and determining a first candidate content list corresponding to the user tags based on the CTR data; determining a second candidate content list based on the first candidate content lists corresponding to multiple user tags under the same user; performing Thompson sampling on the CTR data of the candidate content in the second candidate content list to calculate a random number corresponding to the candidate content; and reordering the candidate content in the second candidate content list based on the random number to determine a user recommendation list. This application can provide users with more differentiated recommended content.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a cold start method and apparatus, electronic device and storage medium for a recommendation system. Background Technology

[0002] Currently, during the cold start process of recommendation systems, user tag-based recommendation methods may result in excessive repetition of recommendation lists displayed to different users, negatively impacting user experience. Therefore, how to improve the richness of recommended content for users during the cold start phase of recommendation systems has become an urgent technical problem to be solved. Summary of the Invention

[0003] The main objective of this application is to provide a cold start method, apparatus, electronic device, and storage medium for a recommendation system, aiming to improve the richness of user-recommended content during the cold start process of the recommendation system.

[0004] To achieve the above objectives, a first aspect of this application proposes a cold start method for a recommendation system. The method includes: determining a list of historical clicked content corresponding to a user tag based on acquired user tracking data and user tags; wherein the list of historical clicked content contains a plurality of historical clicked contents; calculating feature data of the historical clicked content based on the user tracking data; performing smoothing calculation on the feature data to calculate click-through rate data corresponding to the historical clicked content; determining a first candidate content list corresponding to the user tag based on the click-through rate data; wherein the first candidate content list includes a plurality of historical clicked contents arranged in order of the click-through rate data; determining a second candidate content list based on the first candidate content lists corresponding to multiple user tags under the same user; wherein the second candidate content list includes a plurality of candidate contents; performing Thompson sampling on the click-through rate data of the candidate contents in the second candidate content list to calculate a random number corresponding to the candidate contents; reordering the candidate contents in the second candidate content list according to the order of the random numbers to determine a user recommendation list; wherein the user recommendation list includes a plurality of recommended contents.

[0005] In some embodiments, the feature data includes click data and impression data.

[0006] In some embodiments, the step of smoothing the feature data to calculate the click-through rate (CTR) data corresponding to the historical clicked content includes: calculating the average click volume of each historical clicked content in the historical clicked content list based on the click volume data; calculating the average exposure volume of each historical clicked content in the historical clicked content list based on the exposure volume data; and performing a smoothing calculation based on the click volume data, the exposure volume data, the average click volume, and the average exposure volume to calculate the CTR data.

[0007] In some embodiments, determining a second candidate content list based on the first candidate content list corresponding to multiple user tags under the same user includes: sorting the historical clicked content in descending order according to the click-through rate data corresponding to the historical clicked content in the first candidate content list to obtain a list to be selected; determining the top N historical clicked content in the list to be selected as the candidate content to obtain the second candidate content list; wherein N is a positive integer less than or equal to a preset list threshold.

[0008] In some embodiments, the method further includes: if different first candidate content lists contain the same historical clicked content, determining the order of the historical clicked content in the candidate list based on the highest click-through rate data corresponding to the historical clicked content.

[0009] In some embodiments, performing Thompson sampling on the click-through rate data of the candidate content in the second candidate content list to calculate a random number corresponding to the candidate content includes: calculating the average click-through rate based on the click-through rate data corresponding to all historical clicked content in the second candidate content list; determining a first smoothing parameter and a second smoothing parameter based on the average click-through rate; determining a first parameter and a second parameter in the Thompson sampling based on the first smoothing parameter and the second smoothing parameter; and calculating the random number corresponding to the candidate content based on the first parameter and the second parameter.

[0010] In some embodiments, prior to the step of performing Thompson sampling on the click-through rate data of the candidate content in the second candidate content list, the method further includes: normalizing the click-through rate data of the candidate content in the second candidate content list.

[0011] To achieve the above objectives, a second aspect of this application proposes a cold start device for a recommendation system. The device includes: a first module, configured to determine a list of historical clicked content corresponding to a user tag based on acquired user tracking data and user tags; wherein the list of historical clicked content contains a plurality of historical clicked items; a second module, configured to calculate feature data of the historical clicked content based on the user tracking data; a third module, configured to perform smoothing calculations on the feature data to calculate click-through rate data corresponding to the historical clicked content; and a fourth module, configured to determine a first candidate content list corresponding to the user tag based on the click-through rate data; wherein the first candidate... The selected content list includes multiple historical clicked contents arranged in order of the click-through rate data; the fifth module is used to determine a second candidate content list based on the first candidate content list corresponding to multiple user tags under the same user; wherein, the second candidate content list includes several candidate contents; the sixth module is used to perform Thompson sampling on the click-through rate data of the candidate contents in the second candidate content list to calculate a random number corresponding to the candidate contents; the seventh module is used to reorder the candidate contents in the second candidate content list according to the order of the random numbers to determine a user recommendation list; wherein, the user recommendation list includes several recommended contents.

[0012] An electronic device includes a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for enabling communication between the processor and the memory, wherein the program, when executed by the processor, implements the method described in the first aspect above.

[0013] A storage medium, which is a computer-readable storage medium for computer-readable storage, stores one or more programs that can be executed by one or more processors to implement the method described in the first aspect above.

[0014] This application proposes a cold start method, apparatus, electronic device, and storage medium for a recommendation system. The method first determines a list of historical clicked content corresponding to each user tag based on acquired user tracking data and user tags; this list contains several historical clicked items. Then, feature data of the historical clicked content is calculated based on the user tracking data; the feature data is smoothed to calculate the click-through rate (CTR) data corresponding to the historical clicked content; and a first candidate content list corresponding to each user tag is determined based on the CTR data; this first candidate content list includes multiple historical clicked items arranged in order of CTR data. In the tag recall stage, a second candidate content list is first determined based on the first candidate content lists corresponding to multiple user tags under the same user; this second candidate content list includes several candidate items. Then, Thompson sampling is performed on the CTR data of the candidate items in the second candidate content list to calculate random numbers corresponding to the candidate items. Finally, the candidate items in the second candidate content list are reordered according to the order of the random numbers to determine a user recommendation list. This user recommendation list includes several recommended items. In this embodiment, firstly, feature data is smoothed to reduce the overestimation or underestimation of historical click content due to special cases such as click counts of 1 or 0. Secondly, this embodiment uses Thompson sampling to achieve an effective balance between the randomness and popularity of recommended content, thereby enabling the provision of more differentiated recommended content to multiple users with similar tags. Furthermore, the cold start method for the recommendation system proposed in this embodiment is relatively easy to implement and requires minimal manual intervention, making it widely applicable to various recommendation systems. Attached Figure Description

[0015] Figure 1 This is a flowchart of the cold start method of the recommendation system provided in the embodiments of this application;

[0016] Figure 2 yes Figure 1 The flowchart of step S103 in the process;

[0017] Figure 3 yes Figure 1 The flowchart of step S105 in the process;

[0018] Figure 4 yes Figure 1 The flowchart of step S106 in the process;

[0019] Figure 5 This is a schematic diagram of the cold start device of the recommendation system provided in the embodiments of this application;

[0020] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] 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.

[0022] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0024] First, let's analyze some of the terms used in this application:

[0025] Recommender systems are information filtering systems used to predict user preferences for items. By analyzing and mining user behavior, they discover users' personalized needs and interests, and recommend content that the user may be interested in. In some scenarios, recommender systems will filter content for specific users based on pre-defined strategies and recommend the filtered content to that user.

[0026] Cold start: In a recommendation system, new users and new content are unfamiliar to the system. The system needs to collect basic data about these new users or content to help complete the subsequent recommendation process. This process of accumulating basic data from scratch is called the cold start of the recommendation system. Generally, cold starts in recommendation systems include user cold start, content cold start, and system cold start. User cold start refers to making personalized recommendations to new users; content cold start refers to recommending new content to users who might be interested in it; system cold start refers to setting up recommendation strategies in a new recommendation system so that personalized recommendation needs can be met when the website is launched.

[0027] Data tracking: Data tracking refers to the technologies and implementation processes involved in capturing, processing, and transmitting data related to specific user behaviors or events. In the internet industry, data tracking is typically used to record user behavior events, analyze user paths and product metrics, thereby driving product optimization or guiding operations. Data tracking can be divided into front-end tracking and server-side tracking. Front-end tracking involves embedding capture code in the front-end page to collect user behavior events; server-side tracking involves embedding capture code in the server, primarily collecting data on how user behavior events are processed in the back-end server (e.g., data updates in the back-end database).

[0028] Currently, during the cold start process of recommendation systems, user tag-based recommendation methods can recommend content to users by recalling user tags. Specifically, this involves integrating and filtering different recommended content under user tags, determining the final recommendation list based on different tag combinations, and then displaying the recommendation list to the user. However, in this recommendation method, users with similar tag combinations may see highly similar recommendation lists. This is especially true for multiple users with relatively simple tags; these users may receive completely identical recommendation lists. For example, if user A's tag combination is [female, loves music], and user B's tag combination is also [female, loves music], then the recommendation method presented to user A and user B may be exactly the same.

[0029] It's conceivable that such a recommendation method firstly fails to meet users' personalized needs, resulting in a poor user experience and potentially impacting user retention and conversion rates. Secondly, to improve the accuracy of predicting user preferences and provide more tailored recommendations during system operation, the recommendation system needs to collect as much basic data as possible about new users browsing different types of content during the cold start process to accelerate the improvement of recommendation accuracy. However, in related recommendation methods, users with similar tags tend to browse content in highly similar recommendation lists, resulting in low data richness generated during user browsing. This makes it difficult for the recommendation system to quickly identify user preferences, potentially causing the system's recommendations to increasingly deviate from user interests, leading to a poor user experience. Furthermore, since recommendation systems can only recommend a limited amount of content to users each time, some recommendation methods consistently push similar content to new users with similar tags. This can lead to a situation where content that first appears in the recommendation list has an increasingly high click-through rate, and this content continues to be pushed to other users. Conversely, content not yet in the recommendation list maintains a low click-through rate, causing it to be considered unpopular by the user and thus not pushed to other users. Consequently, these recommendation methods may result in new content not receiving sufficient exposure, hindering its promotion among users.

[0030] Therefore, recommendation methods in related technologies may result in excessive repetition of recommendation lists displayed to different users, negatively impacting user experience. Thus, improving the richness of recommended content during the cold start phase of a recommendation system has become a pressing technical problem.

[0031] Based on this, embodiments of this application provide a cold start method, apparatus, electronic device, and storage medium for a recommendation system. The method first determines a list of historical clicked content corresponding to a user tag based on acquired user tracking data and user tags; wherein the list of historical clicked content contains several historical clicked items. Then, feature data of the historical clicked content is calculated based on the user tracking data; the feature data is smoothed to calculate the click-through rate data corresponding to the historical clicked content, and a first candidate content list corresponding to the user tag is determined based on the click-through rate data; wherein the first candidate content list includes multiple historical clicked items arranged in order of click-through rate data. In the tag recall stage, a second candidate content list is first determined based on the first candidate content lists corresponding to multiple user tags under the same user; wherein the second candidate content list includes several candidate items. Then, Thompson sampling is performed on the click-through rate data of the candidate items in the second candidate content list to calculate a random number corresponding to the candidate item. Finally, the candidate items in the second candidate content list are reordered according to the order of the random numbers to determine a user recommendation list. The user recommendation list includes several recommended items. In this embodiment, firstly, feature data is smoothed to reduce the overestimation or underestimation of historical click content due to special cases such as click counts of 1 or 0. Secondly, this embodiment uses Thompson sampling to achieve an effective balance between the randomness and popularity of recommended content, thereby enabling the provision of more differentiated recommended content to multiple users with similar tags. Furthermore, the cold start method for the recommendation system proposed in this embodiment is relatively easy to implement and requires minimal manual intervention, making it widely applicable to various recommendation systems.

[0032] Furthermore, the cold start method for the recommendation system provided in this application relates to the field of computer technology. The cold start method for the recommendation system provided in this application can be applied to a terminal or a server, and can also be implemented as software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the cold start method of the recommendation system, etc., but is not limited to the above forms.

[0033] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0034] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards of the relevant countries and regions. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data for the proper functioning of the embodiments of this application obtained.

[0035] The recommended methods, apparatuses, electronic devices, and storage media provided in the embodiments of this application are specifically described through the following embodiments. First, the recommended methods in the embodiments of this application are described.

[0036] Figure 1 This is a flowchart of the cold start method of the recommendation system provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0037] Step S101: Based on the obtained user tracking data and user tags, determine the list of historical clicked content corresponding to the user tags;

[0038] Specifically, to address the selection pressure brought about by massive amounts of data, various types of recommendation systems have emerged. Recommendation systems can recommend high-quality, personalized content to users, reducing the time and effort users spend blindly selecting content, improving the user experience, and thus increasing user retention rates. To meet diverse user needs, recommendation systems come in many forms. For example, there are systems that recommend physical goods or virtual services, such as shopping platforms or food delivery platforms; there are platforms that recommend content, such as knowledge payment systems and video websites; and there are platforms that recommend other users, such as social media platforms and recruitment platforms. Because recommendation systems can recommend a wide variety of content types, this application uses "recommended content" to summarize the specific recommendation items of the recommendation system. In different recommendation systems, continuously accepting new users and providing them with personalized recommendation services is an important direction for the development of recommendation systems; therefore, the cold start phase of a recommendation system is particularly important.

[0039] As mentioned above, the cold start of a recommendation system can be divided into user cold start, content cold start, and system cold start. In this embodiment, user cold start is the main application scenario. During the user cold start process, new users do not leave any records in the recommendation system. Therefore, in order for new users to enjoy certain personalized recommendation services from the very beginning, it is necessary to obtain user tracking data and user tags after confirming that the user is aware of and has given their consent.

[0040] In this embodiment, user tags are used to characterize the user's own features, such as gender, age, location, interests, etc. For example, for user A, their user tags might include [male, 30-35 years old, residing in location B, liking sport C], etc. Depending on the type of recommendation system, the system's operators can determine the target user group through the combination and filtering of tags.

[0041] As mentioned above, user tracking data is used to record user behavior events on the internet. When using the system, users typically express their attention to the current content by clicking on or hovering over it. Therefore, in this embodiment, the user tracking data mainly consists of the user's historical clicks, which are the user's past clicks on content.

[0042] Understandably, for different types of recommendation systems, obtaining user tracking data from similar systems to predict user preferences yields higher accuracy and a better user experience. For example, if user A was a user of video website B, then when user A uses a new video website C, user A's interests on video website B are clearly of great significance for video website C in recommending content to user A. Therefore, in actual implementation, obtaining user tracking data from similar systems can be prioritized as a data reference for the embodiments of this application.

[0043] Understandably, user interests and preferences can change over time; for example, user A might have preferred type B articles a year ago but now prefer type C. Therefore, the time period for acquiring event tracking data should be relatively close to the current time, such as acquiring user event tracking data from the past 7 days. After acquiring the user event tracking data, it is grouped according to user tags to determine the historical click content list corresponding to different user tags. This historical click content list contains several historical click items.

[0044] Step S102: Calculate the feature data of historical clicked content based on user tracking data;

[0045] Specifically, as mentioned above, user tracking data mainly consists of historical click data representing past user clicks on content. Therefore, feature data of historical click content can be calculated based on user tracking data. In this embodiment, this feature data includes click volume data (clickNum), exposure data (showNum), and click-through rate (ctr), etc. The click-through rate value in this step can be considered as the quotient obtained by dividing the click volume data by the exposure data.

[0046] Step S103: Perform smoothing calculation on the feature data to calculate the click-through rate data corresponding to the historical clicked content;

[0047] Specifically, in recommendation systems, click-through rate (CTR) is typically used as a key metric for content recommendation. However, in certain situations, CTR may be overestimated or underestimated. For example, if article A has a CTR of 1 and an impression of 1, its calculated CTR will also be 1, representing a 100% CTR for article A. This is clearly an overestimation of article A's appeal to users. Conversely, if article B has a CTR of 0, it might be calculated as having a CTR of 0. Article B may be new content in the recommendation system. If its CTR is 0, the system might consider article B to be unpopular, thus underestimating its potential appeal and reducing its recommendation, resulting in ineffective promotion of new content.

[0048] To address the aforementioned issues to some extent, this step proposes smoothing the feature data. Specifically, Bayesian smoothing is applied to data such as the exposure and click volume of historical clicked content. This aims to prevent the click-through rate from sliding to extremes like 1 or 0, and to some extent protects the opportunity for new content to be recommended, thereby further improving the richness of content recommended by the recommendation system to users.

[0049] Step S104: Based on the click-through rate data, determine the first candidate content list corresponding to the user tags;

[0050] Specifically, the smoothed click-through rate (CTR) data calculated in step S103 can reflect the degree of user attention to historically clicked content to a considerable extent. Therefore, in this step, a first candidate content list can be determined for each user tag based on the click-through rate data. The first candidate content list includes multiple historically clicked items, and these historically clicked items are arranged in descending order according to their corresponding CTR data.

[0051] Understandably, each user tag has a corresponding list of primary candidate content. Historical clicks on these primary candidate content lists may overlap across different tags. For example, the same article A might appear under both the [Male] and [Female] tags, meaning article A has an audience among both men and women. However, the click-through rate (CTR) for the same historically clicked content may differ across tags. For instance, for the same article A, men might show greater interest than women, resulting in a higher CTR for article A under the [Male] tag compared to the [Female] tag.

[0052] Through the above steps S101-S104, this embodiment of the application determines the first candidate content list under each user tag based on the user tracking data of the new user, and completes the data preparation work for the cold start of the recommendation system. The following describes how this embodiment of the application further enriches the content of the user recommendation list in the tag recall and recommendation generation stages.

[0053] Step S105: Determine the second candidate content list based on the first candidate content list corresponding to multiple user tags under the same user;

[0054] Specifically, in the tag recall phase, after obtaining user permissions, the recommendation system can collect multiple user tags for the same new user. For example, the user tags for new user A include tag X, tag Y, and tag Z, which correspond to the first candidate content list x, the first candidate content list y, and the first candidate content list y, respectively.

[0055] To provide personalized recommendations to user A, it is necessary to integrate the first suggestions under tags X, Y, and Z.

[0056] Selecting content lists x, y, and z yields a second candidate content list. It can be understood that this second candidate content list represents the content that the current user is interested in across different areas. For example, user B, aged 20-25, who enjoys music, might be more interested in articles related to career planning and music information. In this embodiment, the content in the second candidate content list is referred to as candidate content, and this second candidate content list includes several candidate contents.

[0057] Step S106: Perform Thompson sampling on the click-through rate data of the candidate content in the second candidate content list to calculate the random number corresponding to the candidate content;

[0058] Specifically, in some embodiments, the candidate content in the second candidate content list is arranged in descending order of click-through rate (CTR) data. Therefore, for different users with similar tag combinations, if the second candidate content list is directly pushed to these users as a user recommendation list, the order of the recommended content seen by these users may be the same. When users click to browse recommended content, they will also tend to browse content that appears earlier in the list, which will result in content with already low CTRs not being seen by users, further increasing the difficulty of promotion. Users will only be repeatedly recommended content with high CTRs, resulting in a poor user experience. Therefore, in this embodiment, a Thompson sampling method is proposed for the CTR data of the candidate content in the second candidate content list. The principle of Thompson sampling is based on the beta distribution. In a recommendation system, the parameter α of the beta distribution can be used to characterize the number of times a user clicks on a piece of content after the recommendation system recommends it, and the parameter β can be used to characterize the number of times a user does not click on a piece of content after the recommendation system recommends it. In this embodiment, parameter α is referred to as the first parameter, and parameter β is referred to as the second parameter. In this step, based on the click-through rate data and the first and second parameters, a random number corresponding to each candidate content is calculated using the probability density function of the beta distribution. The following formula is the probability density function of the beta distribution:

[0059]

[0060] Where x represents the click-through rate data, and Γ represents the gamma function.

[0061] The random numbers obtained from the above beta distribution all belong to the 0-1 interval, and the mean of the beta distribution is as follows:

[0062]

[0063] The larger the mean value, the closer the center of the probability density distribution of the beta distribution is to 1, and the closer the generated random number is to 1. Therefore, the higher the click-through rate of the content, the closer the generated random number is to 1. Thus, processing the click-through rate data using Thompson sampling can generate new random numbers. These random numbers are used to characterize the new order of candidate content in the second candidate content list. Therefore, the embodiments of this application can largely overcome the problem in related technologies that users with similar tag combinations receive similar recommended content orders. On the one hand, it provides users with richer recommended content, and on the other hand, it is more conducive to the promotion of new content among users.

[0064] The new data generated by Thompson sampling, since the probability of generating random numbers is still related to the click-through rate itself, can ensure the randomness of the candidate content in the second candidate content list and ensure that the new order of the candidate content is related to the original order, thereby avoiding the situation where the recommended content deviates too much from the user's preferences.

[0065] In some embodiments, before performing Thompson sampling on the second candidate content list, to simplify calculations and facilitate subsequent data processing, the click-through rate (CTR) data of the candidate content in the second candidate content list can be normalized. For example, the sum of the CTR data of the candidate content in the second candidate content list can be calculated first, denoted as sum. Then, the original CTR data can be replaced with the ratio of the number of clicks to the sum of the CTR data, cTR / sum, to obtain a new CTR, making the sum of all the new CTRs equal to 1. The subsequent Thompson sampling steps are then performed based on the new CTR.

[0066] After performing the above normalization steps, Thompson sampling is then applied to the click-through rate (CTR) data. Since a random number needs to be generated for each candidate content, a set of first and second parameters needs to be determined for each candidate content based on the CTR data. For example, let the first parameter... (To prevent α from being 0, we round α up here), the second parameter β = 100 - α, and then each candidate content generates a random number based on the corresponding first and second parameters.

[0067] Step S107: Reorder the candidate content in the second candidate content list according to the order of the random numbers to determine the user recommendation list;

[0068] Specifically, through step S106 above, a corresponding random number has been generated for all candidate items in the second candidate content list. Then, based on this random number, the order of the candidate items in the list can be reordered. It can be understood that the candidate items are rearranged in descending order of the random numbers, and the resulting list is called the user recommendation list, and the content in this list is called the recommended content.

[0069] It should be noted that the random numbers corresponding to candidate items in the second candidate list may be the same. Candidate items with the same random number can be sorted based on click-through rate data or randomly. Furthermore, when using a recommendation system, the number of recommended items in the user's recommendation list on the current page is generally limited. After browsing the current user recommendation list, users can refresh it on their devices by swiping up or down or clicking the refresh button. Therefore, the number of recommended items in the user recommendation list can be less than the number of candidate items in the second candidate list.

[0070] In some embodiments, after reordering the candidate content in the second candidate content list, for example, the first M candidate content can be selected as the recommended content to construct a user recommendation list, where M is a positive integer. When a user requests to refresh the current user recommendation list, the process can continue from the (M+1)th candidate content in the second candidate content list to the 2Mth candidate content as the second user recommendation list, and so on, until all candidate content in all current second candidate recommendation lists is pushed to the user. In other embodiments, when the user recommendation list is generated for the first time, the first M candidate content can be selected as the recommended content. When the user recommendation list needs to be generated again, the second candidate content list can be regenerated, or the random number of the remaining candidate content in the second candidate content list can be recalculated, and the remaining candidate content can be reordered again based on the random number to generate a new user recommendation list.

[0071] It should be noted that the method described above for refreshing the user recommendation list is applied to the cold start process of the recommendation system. Once the recommendation system has collected enough basic data about the new user, it can recommend personalized content based on the user's preset recommendation method.

[0072] Steps S101 to S107 of this embodiment first determine a list of historical clicked content corresponding to a user tag based on the acquired user tracking data and user tags; the list of historical clicked content contains several historical clicked items. Then, feature data of the historical clicked content is calculated based on the user tracking data; the feature data is smoothed to calculate the click-through rate (CTR) data corresponding to the historical clicked content, and a first candidate content list corresponding to the user tag is determined based on the CTR data; the first candidate content list includes multiple historical clicked items arranged in order of CTR data. In the tag recall stage, a second candidate content list is first determined based on the first candidate content lists corresponding to multiple user tags under the same user; the second candidate content list includes several candidate items. Then, Thompson sampling is performed on the CTR data of the candidate items in the second candidate content list to calculate a random number corresponding to the candidate item. Finally, the candidate items in the second candidate content list are reordered according to the order of the random numbers to determine a user recommendation list. The user recommendation list includes several recommended items. In this embodiment, firstly, feature data is smoothed to reduce the overestimation or underestimation of historical click content due to special cases such as click counts of 1 or 0. Secondly, this embodiment uses Thompson sampling to achieve an effective balance between the randomness and popularity of recommended content, thereby enabling the provision of more differentiated recommended content to multiple users with similar tags. Furthermore, the cold start method for the recommendation system proposed in this embodiment is relatively easy to implement and requires minimal manual intervention, making it widely applicable to various recommendation systems.

[0073] Please see Figure 2 In some embodiments, step S103 may include, but is not limited to, steps S201 to S203:

[0074] Step S201: Calculate the average number of clicks for each historical clicked item in the historical clicked content list based on the click data;

[0075] Specifically, based on the above, click volume data for each historical clicked content can be obtained from the user tracking data of the past 7 days. This click volume data includes the click data of different users under different tags for different content. For example, user B under tag A did not click on content C in the past 7 days, while user E under tag D clicked on content C 3 times in the past 7 days. Therefore, based on the click volume data, the average click volume corresponding to each historical clicked content in the historical clicked content list for each user tag can be calculated, which is the average number of times users under that user tag clicked that content.

[0076] Step S202: Calculate the average exposure of each historical clicked content in the historical clicked content list based on the exposure data;

[0077] Specifically, similar to the exposure volume mentioned above, in this step, the average exposure volume of each historical clicked content in the historical clicked content list can also be calculated using the exposure volume data in the user tracking data. This is the average number of times the recommendation system exposes the content under the current user tag.

[0078] Step S203: Perform smoothing calculations based on click data, impression data, average clicks, and average impressions to obtain click-through rate data;

[0079] Specifically, based on the above, Bayesian smoothing is performed on click-through rate (CTR) data from click-through data, impression data, average click-through rate, and average impression data. This CTR data is used for subsequent list content filtering and sorting. The specific formula for calculating the CTR data is as follows:

[0080]

[0081] Wherein, clickNum represents the number of clicks, showNum represents the number of impressions, clickAvg represents the average number of clicks, and showAvg represents the average number of impressions. Through the above steps S201-S203, this embodiment of the application realizes the smooth calculation of feature data and calculates the click-through rate data used for subsequent content recommendation filtering and sorting.

[0082] Please see Figure 3 In some embodiments, step S105 may include, but is not limited to, steps S301 to S302:

[0083] Step S301: Based on the click-through rate data of the historical clicked content in the first candidate content list, sort the historical clicked content in descending order to obtain the candidate list;

[0084] Specifically, as mentioned above, based on user tracking data, the first candidate content list under each user tag can be determined. When the recommendation system needs to determine the recommended content for the current new user, it obtains the first candidate content list corresponding to all tags of the current new user, and sorts all historical clicked content in the first candidate content list together based on click-through rate data. For example, the current user's tags correspond to the first candidate content lists List A and List B. If the articles (i.e., items) and their click-through rate data (i.e., CTR) in the first candidate content list are represented in the form of [item:ctr], then List A can be represented as List A = [item1:0.7, item2,0.6, item3,0.4, ...], and List B can be represented as List B = [item4:0.8, item5,0.5, item6,0.3, ...]. Then, based on the descending order of the click-through rate data in the two first candidate lists, the historical clicked items can be sorted to obtain the candidate list List T = [item4:0.8, item1:0.7, item2:0.6, item5:0.5, item3:0.4, item6:0.3]. This candidate list is a temporary list, and the second candidate list can be determined based on this candidate list in subsequent steps.

[0085] In other embodiments, there may be cases where the same historical clicked content exists in different tags of a user. For example, List A = [item1:0.7, item2,0.6, item3,0.4...] and List B = [item1:0.8, item5,0.5, item6,0.3...]. If item1 is included in both lists, then during sorting, the list is sorted according to the highest click-through rate of that historical clicked content. That is, [item1:0.8] is taken as the click-through rate data of item1, and the resulting list to be selected is List T = [item1:0.8, item2,0.6, item5,0.5, item3,0.4, item6,0.3].

[0086] Step S302: Determine the top N historical clicked items in the selection list as candidate items to obtain the second candidate item list;

[0087] Specifically, while the candidate lists listed above contain relatively few historical clicks, in reality, users generate and access a large amount of event tracking data daily. Even after removing duplicates, the final candidate list still contains a significant amount of content when integrating the first candidate content lists from all notes. Therefore, to provide users with content more aligned with their preferences and to facilitate browsing without overwhelming them with excessive recommendations, a list threshold is typically set. This threshold represents the maximum amount of content that can be accommodated in the second candidate content list. Then, the top N historical clicks from the candidate list are selected as candidates to construct the second candidate content list.

[0088] Where N is a positive integer less than or equal to a preset list threshold. Since the contents of the list to be selected are all...

[0089] The content is sorted in descending order based on the click-through rate (CTR) data. Therefore, the N candidate content items in the second candidate content list are also sorted in descending order based on the CTR data. In subsequent steps, sampling and other operations can be performed.

[0090] Through steps S301-S302, this application embodiment provides a specific process for determining a second candidate content list based on a first candidate content list.

[0091] Please see Figure 4 In some embodiments, step S106 may also include, but is not limited to, steps S401 to S404:

[0092] Step S401: Calculate the average click-through rate based on the click-through rate data corresponding to all historical clicked content in the second candidate content list;

[0093] Specifically, as mentioned above, before performing Thompson sampling on the click-through rate (CTR) data, the CTR data can be normalized for easier calculation. Furthermore, to further prevent overly extreme probabilities from being obtained in the beta distribution, a data smoothing operation can be performed again in some embodiments. Firstly, in this step, the mean CTR is calculated based on the CTR data corresponding to all historical clicked content in the second candidate content list; this mean CTR is denoted by "mean".

[0094] Step S402: Determine the first smoothing parameter and the second smoothing parameter based on the average click-through rate;

[0095] Specifically, after calculating the mean click-through rate (CTR) according to the above steps, another set of smoothing parameters is determined based on the mean CTR. Let the first smoothing parameter be represented by x, and the second smoothing parameter by y. We can let x = 100 * mean, and let y = 100 - x.

[0096] Step S403: Determine the first parameter and the second parameter in Thompson sampling based on the first smoothing parameter and the second smoothing parameter;

[0097] Specifically, the first parameter α is determined as α = α + x, and the second parameter β is determined as β = β + y.

[0098] Step S404: Calculate the random number corresponding to the candidate content based on the first parameter and the second parameter;

[0099] Specifically, based on the first and second parameters determined in step S403, random numbers are obtained based on the beta distribution, that is, beta(α, β) = beta(α+x, β+y).

[0100] Through the above steps S401-S404, this application embodiment provides a specific process for smoothing the click-through rate data again to reduce the occurrence of extreme probabilities during Thompson sampling.

[0101] In summary, based on one or more embodiments described above, this application provides a cold start method for a recommendation system. Firstly, during the determination of the first candidate content list, Bayesian smoothing is used to correct click-through rate (CTR) data, reducing the likelihood of CTRs sliding to extremes due to specific click and exposure conditions. Secondly, during the generation of the user recommendation list, Thompson sampling is used to reorder the recommended content, further enriching the recommended content in the user recommendation list. This allows the recommendation system to collect more click-through data from different domains from the current user more quickly, thereby improving the accuracy of personalized recommendations. Furthermore, the method proposed in this application is relatively simple to implement and requires minimal manual intervention, making it convenient, effective, and widely applicable in the field of cold start technology for recommendation systems.

[0102] Please see Figure 5This application also provides a cold start device for a recommendation system, which can implement the cold start method of the recommendation system described above. The device includes: a first module, a second module, a third module, a fourth module, a fifth module, a sixth module, and a seventh module. The system comprises the following modules: The first module determines a list of historical clicked content corresponding to a user tag based on the acquired user tracking data and user tags; this list contains several historical clicked content items. The second module calculates feature data for the historical clicked content based on the user tracking data. The third module performs smoothing calculations on the feature data to calculate the click-through rate (CTR) data corresponding to the historical clicked content. The fourth module determines a first candidate content list corresponding to a user tag based on the CTR data; this first candidate content list includes multiple historical clicked content items arranged in order of CTR data. The fifth module determines a second candidate content list based on the first candidate content lists corresponding to multiple user tags under the same user; this second candidate content list includes several candidate items. The sixth module performs Thompson sampling on the CTR data of the candidate items in the second candidate content list to calculate a random number corresponding to the candidate item. The seventh module reorders the candidate items in the second candidate content list according to the order of the random numbers to determine a user recommendation list; this user recommendation list includes several recommended items.

[0103] The specific implementation of the cold start device of the recommendation system is basically the same as the specific embodiment of the cold start method of the recommendation system described above, and will not be repeated here.

[0104] This application also provides an electronic device, which includes: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for communication between the processor and the memory. When the program is executed by the processor, it implements the cold start method of the recommended system described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0105] Please see Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0106] The processor 601 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0107] The memory 602 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 602 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 602 and is called by the processor 601 to execute the cold start method of the recommended system in the embodiments of this application.

[0108] The input / output interface 603 is used to implement information input and output;

[0109] The communication interface 604 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0110] Bus 605 transmits information between various components of the device (e.g., processor 601, memory 602, input / output interface 603, and communication interface 604);

[0111] The processor 601, memory 602, input / output interface 603, and communication interface 604 are connected to each other within the device via bus 605.

[0112] This application embodiment also provides a storage medium, which is a computer-readable storage medium for computer-readable storage. The storage medium stores one or more programs, which can be executed by one or more processors to implement the cold start method of the recommended system described above.

[0113] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0114] The recommendation method, device, electronic device, and storage medium provided in this application first determine a list of historical clicked content corresponding to a user tag based on acquired user tracking data and user tags; wherein the list of historical clicked content contains several historical clicked items. Then, feature data of the historical clicked content is calculated based on the user tracking data; the feature data is smoothed to calculate the click-through rate (CTR) data corresponding to the historical clicked content, and a first candidate content list corresponding to the user tag is determined based on the CTR data; wherein the first candidate content list includes multiple historical clicked items arranged in order of CTR data. In the tag recall stage, a second candidate content list is first determined based on the first candidate content lists corresponding to multiple user tags under the same user; wherein the second candidate content list includes several candidate items. Then, Thompson sampling is performed on the CTR data of the candidate items in the second candidate content list to calculate a random number corresponding to the candidate item. Finally, the candidate items in the second candidate content list are reordered according to the order of the random numbers to determine a user recommendation list. The user recommendation list includes several recommended items. In this embodiment, firstly, feature data is smoothed to reduce the overestimation or underestimation of historical click content due to special cases such as click counts of 1 or 0. Secondly, this embodiment uses Thompson sampling to achieve an effective balance between the randomness and popularity of recommended content, thereby enabling the provision of more differentiated recommended content to multiple users with similar tags. Furthermore, the cold start method for the recommendation system proposed in this embodiment is relatively easy to implement and requires minimal manual intervention, making it widely applicable to various recommendation systems.

[0115] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0116] It will be understood by those skilled in the art that Figure 1-4 The technical solutions shown do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0119] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0120] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0122] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A cold start method for a recommendation system, characterized in that, The method includes: Based on the acquired user tracking data and user tags, determine the list of historical clicked content corresponding to the user tags; The historical click content list contains several historical click content items; Based on the user tracking data, the feature data of the historical clicked content is calculated; The feature data is smoothed to obtain the click-through rate data corresponding to the historical clicked content; Based on the click-through rate data, determine the first candidate content list corresponding to the user tag; The first candidate content list includes multiple historical clicked contents arranged in order of the click-through rate data; A second candidate content list is determined based on the first candidate content list corresponding to multiple user tags under the same user; The second candidate content list includes several candidate contents; Thompson sampling is performed on the click-through rate data of the candidate content in the second candidate content list to calculate the random number corresponding to the candidate content; The candidate contents in the second candidate content list are reordered according to the order of the random numbers to determine the user recommendation list; The user recommendation list includes several recommended items; The step of determining the second candidate content list based on the first candidate content list corresponding to multiple user tags under the same user includes: Based on the click-through rate data corresponding to the historical clicked content in the first candidate content list, the historical clicked content is sorted in descending order to obtain a candidate list; The first N historical clicked items in the candidate list are determined as the candidate items to obtain the second candidate item list; Where N is a positive integer less than or equal to a preset list threshold.

2. The cold start method for a recommendation system according to claim 1, characterized in that, The feature data includes click data and impression data.

3. The cold start method for a recommendation system according to claim 2, characterized in that, The step of smoothing the feature data to calculate the click-through rate data corresponding to the historical clicked content includes: Based on the click volume data, the average click volume of each historical clicked item in the historical clicked content list is calculated; Based on the exposure data, the average exposure of each historical clicked item in the historical clicked content list is calculated; The click-through rate (CTR) is calculated by smoothing the click data, the exposure data, the average clicks, and the average exposures.

4. The cold start method for a recommendation system according to claim 1, characterized in that, The method further includes: If different first candidate content lists contain the same historical click content, the order of the historical click content in the candidate list is determined based on the highest click-through rate data corresponding to the historical click content.

5. The cold start method for a recommendation system according to claim 1, characterized in that, The step of performing Thompson sampling on the click-through rate data of the candidate content in the second candidate content list to calculate the random number corresponding to the candidate content includes: The average click-through rate is calculated based on the click-through rate data corresponding to all the historical clicked content in the second candidate content list; Based on the average click-through rate, determine the first smoothing parameter and the second smoothing parameter; The first and second parameters in the Thompson sampling are determined based on the first smoothing parameter and the second smoothing parameter. The random number corresponding to the candidate content is calculated based on the first parameter and the second parameter.

6. The cold start method for a recommendation system according to any one of claims 1-5, characterized in that, Prior to the step of performing Thompson sampling on the click-through rate data of the candidate content in the second candidate content list, the method further includes: The click-through rate data of the candidate content in the second candidate content list is normalized.

7. A cold start device for a recommendation system, characterized in that, The device includes: The first module is used to determine the list of historical clicked content corresponding to the user tags based on the acquired user tracking data and user tags; The historical click content list contains several historical click content items; The second module is used to calculate the feature data of the historical click content based on the user tracking data. The third module is used to perform smoothing calculations on the feature data to calculate the click-through rate data corresponding to the historical click content; The fourth module is used to determine the first candidate content list corresponding to the user tag based on the click-through rate data; The first candidate content list includes multiple historical clicked contents arranged in order of the click-through rate data; The fifth module is used to determine a second candidate content list based on the first candidate content list corresponding to multiple user tags under the same user; The second candidate content list includes several candidate contents; The sixth module is used to perform Thompson sampling on the click-through rate data of the candidate content in the second candidate content list and calculate the random number corresponding to the candidate content. The seventh module is used to reorder the candidate content in the second candidate content list according to the order of the random number to determine the user recommendation list; The user recommendation list includes several recommended items; The fifth module is specifically used for: Based on the click-through rate data corresponding to the historical clicked content in the first candidate content list, the historical clicked content is sorted in descending order to obtain a candidate list; The first N historical clicked items in the candidate list are determined as the candidate items to obtain the second candidate item list; Where N is a positive integer less than or equal to a preset list threshold.

8. An electronic device, characterized in that, The electronic device includes a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for enabling communication between the processor and the memory, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 6.

9. A storage medium, said storage medium being a computer-readable storage medium for computer-readable storage, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the steps of the method according to any one of claims 1 to 6.

Images