A deep learning driven personalized content recommendation method

Abstract

The application relates to the technical field of content recommendation, and discloses a personalized content recommendation method based on deep learning driving, which comprises the following steps: collecting historical data behaviors of a user, obtaining the number of historical data behaviors collected for the user, setting a behavior quantity threshold, marking the user as an inactive user when the number of historical data behaviors is less than the behavior quantity threshold, and executing a sparse data processing strategy.The application can better capture potential semantic information and internal correlation of user behaviors by mapping discretized historical behavior data to a continuous dense low-dimensional vector space through embedding technology in deep learning, can generate personalized recommendation with the aid of behavior data of similar active users by matching basic attributes of the user through a Siamese network, and can relieve the troubles caused by data sparsity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content recommendation, in particular to a personalized content recommendation method based on deep learning driving. BACKGROUND

[0002] The personalized content recommendation method is a technology and strategy that provides customized content or services to users by analyzing their interests, behaviors and preferences. This method aims to improve user experience and engagement, and common application scenarios include e-commerce product recommendation, news information recommendation, video and music recommendation, etc.

[0003] It mainly extracts content features according to the content that the user has liked or consumed before, and then recommends new content with similar features. However, data sparsity problems are prevalent in most recommendation systems, especially for less active users. Due to insufficient historical behavior data, the model is difficult to capture the user's real interests, which makes the recommendation method that simply relies on user historical records ineffective. SUMMARY

[0004] In view of the problems existing in the prior art, the purpose of the present application is to provide a personalized content recommendation method based on deep learning driving, which can reasonably and accurately analyze the sparse data of inactive users to generate personalized content that meets the user's preference expectations.

[0005] In order to achieve the above purpose, the present application provides the following technical scheme: a personalized content recommendation method based on deep learning driving, the method comprising the following steps:

[0006] Collect the historical data behavior of the user, and obtain the number of historical data behaviors collected for the user, set a behavior quantity threshold, when the number of historical data behaviors is less than the behavior quantity threshold, mark it as an inactive user, and execute a sparse data processing strategy;

[0007] The sparse data processing strategy includes establishing an active user attribute library, the active user attribute library includes the historical data behavior and basic attributes of the active user, and obtaining the basic attributes of the inactive user, using a Siamese network to compare and match the basic attributes of the active user and the basic attributes of the inactive user, selecting the active user with the highest similarity to the basic attributes of the inactive user and marking it as a similar user;

[0008] The content features in the historical data behaviors generated by the inactive user are extracted, compared with the content features in the historical data behaviors generated by all similar users, and the similar user with the most similar content features to the inactive user is marked as a template user and the highest feature similarity is obtained, a similarity threshold is set, the highest feature similarity is compared with the similarity threshold, and a corresponding response is made according to the comparison result;

[0009] If the highest feature similarity is greater than or equal to the similarity threshold, the inactive user is temporarily regarded as a template user, and personalized content is recommended to the inactive user according to the template user until the historical data behaviors generated by the inactive user reach a behavior quantity threshold; if the highest feature similarity is less than the similarity threshold, personalized content is generated according to the content features extracted from the historical data behaviors of the inactive user and the content features of the template user.

[0010] In some embodiments, when the highest feature similarity is less than the similarity threshold, a data quantity proportion value is obtained by dividing the number of historical data behaviors of the inactive user by the behavior quantity threshold, and when personalized content is recommended to the inactive user, content conforming to the content features of the inactive user is recommended according to the proportion of the data quantity proportion value, and the remaining personalized content is recommended according to the content features of the template user.

[0011] In some embodiments, when the highest feature similarity is greater than or equal to the similarity threshold, the daily average active times of the inactive user are obtained according to the number of historical data behaviors of the inactive user, and the adjusted active times are obtained within a preset time after the sparse data processing strategy is executed, the adjusted active times are compared with the daily average active times, and a corresponding response is made according to the comparison result.

[0012] In some embodiments, the daily average active times are obtained by obtaining the number of days from the date of establishing an account to the present, and dividing the number of historical data behaviors of the inactive user by the number of days from the date of establishing an account to the present to obtain the daily average active times; the adjusted active times are obtained by obtaining the number of historical data behaviors generated by the inactive user within a preset time after the sparse data processing strategy is executed, and dividing the number of historical data behaviors by the preset time to obtain the adjusted active times.

[0013] In some embodiments, when the adjusted active times are greater than the daily average active times, the personalized content is generated for the user according to the content features of the template user, and no additional operation is performed; when the adjusted active times are less than or equal to the daily average active times, personalized content optimization is performed, the proportion of the content features of the inactive user in the personalized recommendation is increased, and the dependence on the features of the template user is reduced.

[0014] In some embodiments, the specific way of performing the personalized content optimization is that half of the generated content is generated according to the content features of the inactive user, and the other half is generated according to the content features of the template user, or the highest feature similarity of the user is considered to be less than a similarity threshold.

[0015] In some embodiments, when the adjusted active frequency is greater than the daily average active frequency, an adjusted promotion ratio is obtained, the specific way of obtaining the adjusted promotion ratio is that the difference between the adjusted active frequency and the daily average active frequency is obtained to obtain a daily average additional frequency, and the adjusted promotion ratio is obtained by dividing the daily average additional frequency by the daily average active frequency, and a promotion threshold is set, the adjusted promotion ratio is compared with the promotion threshold, and a corresponding response is made according to the comparison result.

[0016] In some embodiments, if the adjusted promotion ratio is greater than or equal to the promotion threshold, the personalized content is continuously generated for the user according to the content features of the template user, and no additional operation is performed; if the adjusted promotion ratio is less than the promotion threshold, further user behavior judgment should be performed.

[0017] In some embodiments, the further user behavior judgment specifically includes: generating 10% of the generated content according to the content features of the user, and marking this part of the content as user feature content, and increasing the priority of the user feature content so that the user can be preferentially recommended to the user feature content when using the product, if the historical data behavior frequency estimation value generated when browsing the user feature content is greater than the historical data behavior frequency estimation value generated when browsing the remaining generated content, the adjusted active frequency of the user is considered to be less than or equal to the daily average active frequency, and the personalized content optimization is performed to reduce the dependence on the template user features; if the historical data behavior frequency estimation value generated when browsing the user feature content is less than or equal to the historical data behavior frequency estimation value generated when browsing the remaining generated content, the personalized content is continuously generated for the user according to the content features of the template user, and no additional operation is performed.

[0018] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above-mentioned personalized content recommendation method based on deep learning driving.

[0019] Compared with the prior art, the technical scheme provided by the application has the beneficial effects as follows:

[0020] Firstly, the application can better capture the potential semantic information and internal correlation of user behavior by mapping the discretized historical behavior data to a continuous dense low-dimensional vector space through embedding technology in deep learning, improve the accuracy of personalized recommendation, and solve the problem of insufficient historical data behavior of inactive users by establishing an active user attribute library and matching the basic attributes of users through Siamese network, which can assist in generating personalized recommendations by using the behavior data of similar active users, and alleviate the problem caused by data sparsity.

[0021] Secondly, when the content feature similarity between inactive users and active users is high, the application can directly generate recommendations by using template users, otherwise, it generates recommended content by mixing the features of inactive users and template users, taking into account the personal preferences of users and the experience of similar users, to ensure that the recommended content is more in line with the actual interests of users.

[0022] Thirdly, the application can monitor the recommendation effect in real time by comparing the adjusted active times and the daily average active times, and further evaluating the adjusted promotion ratio, and when the data change is not significant or the promotion is insufficient, it can also increase the proportion of the features of inactive users in the recommendation, which ensures the dynamic optimization of the recommendation strategy and the system robustness.

[0023] Fourthly, the application can further reduce the dependence on template user features when the estimated data shows that users are more positive to their own feature content by using the method of dividing the generated content proportion and calculating the corresponding historical data behavior times, which realizes truly personalized recommendation, and not only can deal with natural fluctuations in user behavior or external interference such as overall platform activity, but also can use real-time behavior data feedback to judge the recommendation effect. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The figure is a flowchart of the personalized content recommendation method based on deep learning driving. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0026] It can be understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of one element can be one, and in another embodiment, the number of the element can be multiple, and the term "one" cannot be understood as a limitation on the number.

[0027] The present application provides a personalized content recommendation method based on deep learning driving, as shown in Figure 1 The method comprises the following steps:

[0028] Step one, collect the user's historical data behavior, including the number of clicks, browsing time, payment amount, number of collections, number of comments and other basic behaviors, use data preprocessing technology to clean, normalize and standardize the data, so that different behavior data has comparability, use embedding technology in deep learning to map discrete behavior data to continuous dense low-dimensional vector space, in this process, embedding layer can be used to represent the user's various behaviors, so as to capture the potential meaning and the association between behaviors. And get the number of historical data behaviors collected for the user, set the behavior quantity threshold, compare the number of historical data behaviors with the behavior quantity threshold, when the number of historical data behaviors is greater than or equal to the behavior quantity threshold, it indicates that the user has generated a large number of historical data behaviors, and it is marked as an active user, and the content features are extracted according to the user's historical data behaviors, and new content with similar content features is recommended to the user; When the number of historical data behaviors is less than the behavior quantity threshold, it indicates that the user generates a small amount of historical data behaviors, and it is marked as an inactive user, because the user generates insufficient historical data behaviors, a sparse data processing strategy is executed;

[0029] Step two, the sparse data processing strategy includes establishing an active user attribute library, the active user attribute library includes the historical data behavior and basic attributes of the active user, the basic attributes include the gender, age, geographic location and device information of the user, and the basic attributes of the inactive user are obtained, the basic attributes of the inactive user are compared with the active user attributes in the active user database, the comparison method is specifically to use Siamese network or contrast learning architecture, compare and match the basic attributes of the active user and the basic attributes of the inactive user, calculate the cosine similarity or Euclidean distance to measure the similarity between users, select the active user with the highest similarity to the basic attributes of the inactive user, and mark the active user as a similar user;

[0030] Step three, extracting the content features in the historical data behavior generated by the inactive user, comparing the content features with the historical data behavior generated by all similar users, and marking the similar user with the most similar content features to the inactive user as a template user, while obtaining the content feature similarity between the inactive user and the template user and marking it as the highest feature similarity. The highest feature similarity represents the greatest similarity of the content features that can be extracted from the historical data behavior of the similar user to the inactive user, and also represents that the similar user is most similar to the inactive user in the content features extracted from the historical data behavior. A similarity threshold is set, and the highest feature similarity is compared with the similarity threshold. According to the comparison result, the corresponding reaction is made;

[0031] Step four, if the highest feature similarity is greater than or equal to the similarity threshold, it indicates that there is an active user in the active user attribute library that is similar to the inactive user in historical behavior data. Therefore, the inactive user is temporarily regarded as a template user, and personalized content is recommended to the inactive user according to the template user until the historical data behavior generated by the inactive user reaches the behavior quantity threshold. If the highest feature similarity is less than the similarity threshold, it indicates that there is no active user in the active user attribute library that is similar to the inactive user in historical behavior data. Therefore, personalized content is generated according to the content features extracted from the historical data behavior of the inactive user (hereinafter referred to as content features) and the content features of the template user;

[0032] Specifically, when the highest feature similarity is less than the similarity threshold, the data quantity ratio is obtained by dividing the number of historical data behaviors of the inactive user by the behavior quantity threshold. When personalized content is recommended to the inactive user, the content that meets the content features of the user is recommended according to the proportion of the data quantity ratio, and the remaining personalized content is recommended according to the content features of the template user. For example, it is assumed that the number of collected user historical data behaviors is 20 times, the behavior quantity threshold is set to 50 times, and the data quantity ratio is 0.4. Since the number of user historical data behaviors is less than the behavior quantity threshold, the user is marked as an inactive user, and a template user is matched in the active user attribute library. It is assumed that the highest feature similarity between the inactive user and the template user is 50%, and the similarity threshold is set to 75%. Since the highest feature similarity is less than the similarity threshold, when personalized content is generated for the inactive user, 40% of the generated content is generated according to the content features of the inactive user, and the remaining 60% is generated according to the content features of the template user.

[0033] When the highest feature similarity is greater than or equal to the similarity threshold value after the sparsity data processing strategy, personalized content generated by the template user is recommended to the inactive user, at this time, the daily active frequency of the inactive user should also be obtained according to the historical data behavior frequency of the inactive user, and the adjusted active frequency is obtained within the preset time after the sparsity data processing strategy is executed. The adjusted active frequency and the daily active frequency are compared in size. When the adjusted active frequency is greater than the daily active frequency, it indicates that the historical data behavior frequency of the user has been improved after the sparsity data processing strategy is executed, which reflects that the personalized content recommended according to the template user meets the preferences of the inactive user, and the user is willing to improve the browsing volume and usage rate of the product. Then, the personalized content is generated for the user according to the content features of the template user, and no additional operation is performed. When the adjusted active frequency is less than or equal to the daily active frequency, it indicates that the historical data behavior frequency of the user has not been improved or has been reduced after the sparsity data processing strategy is executed, which reflects that the personalized content generated according to the template user does not meet the preferences of the inactive user, and the user's browsing and usage intention of the product is reduced. Then, the personalized content optimization is performed, the proportion of the content features of the inactive user in the personalized recommendation is increased, and the dependence on the template user features is reduced. The specific way of performing the personalized content optimization is that half of the generated content is generated according to the content features of the inactive user, and the other half is generated according to the content features of the template user, or the highest feature similarity of the user is regarded as being less than the similarity threshold value, the content meeting the content features of the user is recommended to the user according to the proportion of the data volume ratio, and the remaining personalized content is recommended according to the content features of the template user.

[0034] The specific way to obtain the daily average active number is to obtain the number of days from the account establishment day to the present day, and divide the historical data behavior number of the inactive user by the number of days from the account establishment day to the present day to obtain the daily average active number. The specific way to obtain the adjusted active number is to obtain the historical data behavior number of the inactive user generated within a preset time after executing the sparsity data processing strategy, and divide the historical data behavior number by the preset time to obtain the adjusted active number. In combination with the above examples, it is assumed that the highest feature similarity between the inactive user and the template user is 80%, and since the similarity threshold is 75%, the highest feature similarity is greater than the similarity threshold, the personalized content generated by the template user is recommended to the inactive user, it is assumed that the number of days from the account establishment day to the present day of the inactive user is 20 days, and the generated historical data behavior number is 20 times, then the daily average active number is 1 time, it is assumed that the preset time is 5 days, and the historical data behavior number generated by the inactive user within the preset time after executing the sparsity data processing strategy is 10 times, then the adjusted active number is 2 times, since the adjusted active number is greater than the daily average active number, it reflects that the personalized content recommendation generated according to the template user conforms to the preference of the inactive user, and then the personalized content is generated for the user according to the content features of the template user.

[0035] Meanwhile, when the adjusted active number is greater than the daily average active number, the adjusted promotion ratio should also be obtained, and the promotion threshold is set, and the adjusted promotion ratio is compared with the promotion threshold, and the corresponding response is made according to the comparison result. Specifically, if the adjusted promotion ratio is greater than or equal to the promotion threshold, it indicates that the historical data behavior number generated by the user after executing the sparsity data processing strategy has been significantly improved, reflecting that the personalized content recommendation generated according to the template user conforms to the preference of the inactive user, and the personalized content is generated for the user according to the content features of the template user, and no additional operation is performed; if the adjusted promotion ratio is less than the promotion threshold, it indicates that the historical data behavior number generated by the user after executing the sparsity data processing strategy has been improved, but the promotion amount is not obvious, and it cannot be indicated that the change of the generated content is the cause of the change, because the small promotion amount is not enough to exclude the influence of other external factors or random fluctuations of user behavior, and it cannot be definitely proved that the effect of the improved recommendation strategy (i.e. the recommended content generated by the template user) is the real decisive factor. That is, the insufficient promotion may be caused by natural fluctuations of user habits, changes in overall platform activity or other interference factors, rather than the direct effect of the generated content strategy, at which time further user behavior judgment should be performed;

[0036] The specific manner of obtaining the adjusted promotion ratio is to obtain the daily average additional number of times by subtracting the adjusted active number of times from the daily average active number of times, and to obtain the adjusted promotion ratio by dividing the daily average additional number of times by the daily average active number of times. The further user behavior judgment includes: generating a small part of content amount (such as 10% of the content amount) in the generated content according to the content characteristics of the user, marking the part of the content amount as user characteristic content, increasing the priority of the user characteristic content, so that the user can be preferentially recommended to the user characteristic content when using the product, and comparing the historical data behavior number of times generated by the user when browsing the user characteristic content with the historical data behavior number of times generated by the user when browsing the remaining generated content, if the historical data behavior number of times generated when browsing the user characteristic content is greater than the historical data behavior number of times generated when browsing the remaining generated content, it indicates that the user reacts more actively to the content customized based on the user's characteristics, and the click, stay or other data behavior indicators are better, which indicates that the user's content preference for the user's characteristics is more obvious, and the adjusted active number of times of the user is considered to be less than or equal to the daily average active number of times, and the personalized content optimization is performed to reduce the dependence on the template user characteristics; if the historical data behavior number of times generated when browsing the user characteristic content is less than or equal to the historical data behavior number of times generated when browsing the remaining generated content, it indicates that the user does not have a more positive reaction to the content customized based on the user's characteristics, which indicates that the user's content preference for the user's characteristics is insufficient, and the personalized content is generated for the user according to the content characteristics of the template user without additional operation. For example, it is assumed that the number of days from the establishment of the account of the inactive user to the present is 20 days, and the generated historical data behavior number of times is 20 times, so the daily average active number of times is 1 time, the preset time is set to 5 days, the historical data behavior number of times generated by the inactive user after performing the sparse data processing strategy within the preset time is 6 times, the adjusted active number of times is 1.2 times, the adjusted active number of times is greater than the daily average active number of times, and the daily average additional number of times is 0.2 times, and the adjusted promotion ratio is 20%. Since the adjusted promotion ratio is less than the promotion amount threshold, the further user behavior judgment is performed, 10% of the content amount in the generated content is generated according to the content characteristics of the user, and the part of the content amount is marked as user characteristic content, and the remaining 90% of the content amount is still generated according to the content characteristics of the template user. It is assumed that the user generates 5 historical data behavior numbers when browsing the user characteristic content, and generates 10 historical data behavior numbers when browsing the remaining generated content (content generated according to the template user characteristics), so the historical data behavior number of times generated when browsing the user characteristic content is 5÷0.1=50, and the historical data behavior number of times generated when browsing the remaining generated content is 10÷0.9=11.1, that is, when all the generated content is generated based on the user feature content, the estimated number of historical data behaviors generated by the user is 50 times, and when all the generated content is generated based on the template user feature, the estimated number of historical data behaviors generated is 11.1 times, which indicates that the user's content preference for his own features is more obvious. The adjusted active number of the user is less than or equal to the average daily active number, and personalized content optimization is performed.

[0037] In summary, the present application aims to design a deep learning driven personalized content recommendation method. In view of the problem that the real interests of low active users are difficult to be accurately analyzed due to the sparsity of data, the present application maps the discretized historical behavior data to a continuous dense low-dimensional vector space by using embedding technology in deep learning, which can better capture the potential semantic information and internal correlation of user behavior, and improve the accuracy of personalized recommendation. In view of the problem of insufficient historical data behavior of inactive users, by establishing an active user attribute library and using Siamese network or contrast learning to match the basic attributes of users, the behavior data of similar active users can be used to assist in generating personalized recommendations, and the problem caused by data sparsity can be alleviated. When the content feature similarity between inactive users and active users is high, the recommendation can be directly generated by using the template user, otherwise the recommendation content is generated by mixing the inactive user's own features and the template user's features, taking into account the user's personal preferences and similar user's experience, to ensure that the recommended content is more in line with the user's actual interests. By comparing the adjusted active number with the average daily active number and further evaluating the adjusted improvement ratio, the recommendation effect can be monitored in real time, and when the data changes are not significant or the improvement is insufficient, the proportion of inactive user's own features in the recommendation can also be increased, which ensures the dynamic optimization of the recommendation strategy and the system robustness. Moreover, by using the method of dividing the proportion of generated content and calculating the corresponding historical data behavior times by using the estimated value, the user's real reaction to the content of his own features can be more intuitively judged. When the estimated data shows that the user's reaction to the content of his own features is more positive, the system will further reduce the dependence on the template user features, realize the real user-centered recommendation, and measure the adjusted improvement ratio, which can not only cope with the natural fluctuations of user behavior or external disturbances such as the overall activity of the platform, but also use real-time behavior data feedback to judge the recommendation effect, avoiding the mistake of mistaking small improvements caused by external factors as effective recommendation strategies. By dynamically adjusting the content recommendation proportion according to the user's historical data behavior and real-time feedback, the recommended content is finally more in line with the user's real interests, so as to improve the user's click, browse and interaction behavior, and enhance the platform user stickiness and product activity.

[0038] The processes described above with reference to the flowcharts can be implemented as computer software programs in accordance with embodiments of the present disclosure. Embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication section, and / or installed from a detachable medium. When the computer program is executed by a central processing unit, the above-described functions defined in the methods of the present application are performed. It should be noted that the computer readable medium of the present application can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, but not limited to, be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical fiber, a portable compact disk read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus or device. In the present application, the computer readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, in which a computer readable program code is carried. Such a propagated data signal can take many forms, including but not limited to, an electromagnetic signal, an optical signal or any suitable combination of the above. The computer readable signal medium can also be any computer readable medium that can send, propagate or transfer a program for use by or in connection with an instruction execution system, apparatus or device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to, wireless, wire, optical cable, RF or any suitable combination of the above.

[0039] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0040] Those skilled in the art should understand that the above description is only a specific embodiment of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A personalized content recommendation method based on deep learning, characterized in that, The method includes the following steps: Collect users' historical data behavior and obtain the number of historical data behaviors collected for each user. Set a behavior volume threshold. When the number of historical data behaviors is less than the behavior volume threshold, mark the user as an inactive user and execute a sparse data processing strategy. The sparse data processing strategy includes establishing an active user attribute database, which includes the historical data behavior and basic attributes of active users, and obtaining the basic attributes of inactive users. A Siamese network is used to compare and match the basic attributes of active users with the basic attribute features of inactive users, and the active users with the highest similarity to the basic attributes of inactive users are selected and marked as similar users. Extract content features from the historical data behavior of inactive users, compare them with the content features from the historical data behavior of all similar users, mark the similar users with the most similar content features to inactive users as template users and obtain the highest feature similarity, set a similarity threshold, compare the highest feature similarity with the similarity threshold, and take appropriate action based on the comparison results. If the highest feature similarity is greater than or equal to the similarity threshold, the inactive user will be temporarily regarded as a template user, and personalized content will be recommended to the inactive user based on the template user's personalized content until the inactive user's historical data behavior reaches the behavior volume threshold; if the highest feature similarity is less than the similarity threshold, personalized content recommendations will be generated based on the content features extracted from the inactive user's historical data behavior and the template user's content features. After temporarily treating inactive users as template users, if the number of active users of inactive users does not show an increase, perform personalized content optimization. Half of the generated content is generated based on the content characteristics of the inactive user, and the other half is generated based on the content characteristics of the template user, or the user's highest feature similarity is regarded as less than the similarity threshold. When the adjusted number of active users exceeds the average daily number of active users, the adjusted increase ratio is obtained. Specifically, the difference between the adjusted number of active users and the average daily number of active users is used to obtain the daily excess number of active users. This daily excess number of active users is then divided by the average daily number of active users to obtain the adjusted increase ratio. An increase threshold is set, and the adjusted increase ratio is compared to this threshold. If the adjusted increase ratio is greater than or equal to the threshold, personalized content is generated for the user based on the template user's content characteristics, and no further action is taken. If the adjusted increase ratio is less than the threshold, further user behavior judgment should be performed. 10% of the generated content is generated based on the user's content characteristics, and this portion of content is marked as user characteristic content. The priority of the user characteristic content is increased so that users can be recommended to user characteristic content first when using the product. If the estimated number of historical data behaviors generated when browsing user characteristic content is greater than the estimated number of historical data behaviors generated when browsing other generated content, then the user's adjusted activity count should be regarded as less than or equal to the daily average activity count, and personalized content optimization should be performed to reduce the dependence on template user characteristics. If the estimated number of historical data behaviors generated when browsing user-featured content is less than or equal to the estimated number of historical data behaviors generated when browsing other generated content, then personalized content will continue to be generated for that user based on the content characteristics of the template user, without any additional operations.

2. The personalized content recommendation method based on deep learning as described in claim 1, characterized in that, When the highest feature similarity is less than the similarity threshold, the data volume ratio is obtained by dividing the number of historical data behaviors of inactive users by the behavior volume threshold. When recommending personalized content to inactive users, content that matches their own content characteristics is recommended to the user according to the proportion of the data volume ratio. The remaining personalized content is recommended according to the content characteristics of the template user.

3. The personalized content recommendation method based on deep learning according to claim 2, characterized in that, When the highest feature similarity is greater than or equal to the similarity threshold, the average daily active count of inactive users is obtained based on the historical data behavior count of inactive users. After executing the sparse data processing strategy, the adjusted active count is obtained within a preset time. The adjusted active count is compared with the average daily active count, and a corresponding response is made based on the comparison result.

4. The personalized content recommendation method based on deep learning according to claim 3, characterized in that, The specific method to obtain the average daily active users is to obtain the number of days since the inactive user's account was created, and then divide the number of historical data behaviors of the inactive user by the number of days since the account was created to obtain the average daily active users. The specific method to obtain the adjusted active count is to obtain the number of historical data behaviors generated by inactive users from the time the sparse data processing strategy is implemented until a preset time, and then divide the number of historical data behaviors by the preset time to obtain the adjusted active count.

5. The personalized content recommendation method based on deep learning as described in claim 4, characterized in that, When the adjusted number of active users exceeds the average daily number of active users, continue to generate personalized content for the user based on the content characteristics of the template user, without taking any additional actions. When the adjusted number of active users is less than or equal to the average daily number of active users, personalized content optimization will be performed to increase the proportion of inactive users' own content characteristics in personalized recommendations and reduce the reliance on template user characteristics.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement a deep learning-driven personalized content recommendation method as described in any one of claims 1-5.

Images