A short video recommendation method and device

Abstract

The present application relates to a short video recommendation method and device, belonging to the technical field of video recommendation, the method comprising: obtaining a preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category, obtaining a first recommendation list according to the preference portrait score, obtaining a viewing ratio of each feature category according to the second playback time of the short video in the first recommendation list, obtaining video feedback of each feature category according to the viewing ratio of each feature category, updating the first recommendation list according to the video feedback to obtain a second recommendation list, and recommending videos to the user according to the second recommendation list; the recommended videos obtained based on the preference portrait score, the viewing ratio and the video feedback are more in line with the user's preferences. In addition, since there is no need to write and maintain two sets of codes, the development and maintenance costs can be effectively reduced.

Description

Technical Field

[0001] The present application relates to the technical field of video recommendation, and in particular to a short video recommendation method and device. Background Art

[0002] Short videos have extremely high commercial value, which is reflected in advertising and brand promotion, e-commerce sales and product promotion, community building and user interaction, content payment and knowledge realization, cross-border cooperation and business expansion. With the continuous advancement of technology and the continuous changes in the market, the commercial value of short videos will continue to expand and improve. With the continuous innovation of technology, the improvement of network bandwidth and the gradual popularization of 5G technology, the loading speed and viewing experience of short videos have been greatly improved, which provides strong hardware support for the rise of short videos.

[0003] In the prior art, short video recommendations use the Lambda architecture of the recommendation system. The Lambda architecture is an architectural model for efficiently processing large-scale data. It combines batch processing and real-time processing to meet the needs of big data processing and provide high-performance, low-latency and fault-tolerant data processing solutions. However, in actual applications, the Lambda architecture needs to maintain two systems, the batch processing layer and the real-time processing layer, which increases the complexity of development and maintenance. Developers need to write and maintain two sets of code to ensure data consistency and accuracy between the two systems. Since the batch processing layer and the real-time processing layer need to run at the same time, the Lambda architecture will consume more computing resources and storage resources, resulting in high development and maintenance costs. Summary of the invention

[0004] To this end, the present invention provides a short video recommendation method and device to solve the problem of high development and maintenance costs in the prior art.

[0005] In a first aspect, a short video recommendation method is provided, the method comprising:

[0006] Based on the first playback time of the user in each feature category, the preference portrait score corresponding to each feature category is obtained; the feature categories include short video tags, short video authors, short video series, short video major categories, short video secondary categories and short video item portraits;

[0007] Obtaining a first recommendation list according to the preference portrait score; the first recommendation list includes a plurality of first recommended short videos;

[0008] Obtaining a viewing ratio of each feature category according to the second playback duration of the short video in the first recommendation list;

[0009] Obtaining video feedback for each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback;

[0010] The first recommendation list is updated according to the video feedback to obtain a second recommendation list, and videos are recommended to the user according to the second recommendation list.

[0011] Furthermore, obtaining the preference portrait score corresponding to each feature category based on the first playback duration of the user in each feature category includes:

[0012] The current preference score is obtained according to the first preference score formula, and the current preference score formula is:

[0013] P 当前 =T1 / T2;

[0014] Among them, P 当前 is the first preference score; T1 is the first playback duration, that is, the total playback duration of the feature classification within the preset number of days before the current time node; T2 is the total playback duration of all feature classifications within the preset number of days before the current time node;

[0015] The preference portrait score is obtained according to the preference portrait score formula, and the preference portrait score formula is:

[0016] P 偏好 =P 当前 ×s+P 历史 ×z;

[0017] Among them, P 偏好 is the preference profile score; P 当前 is the first preference score; s is the first preset coefficient; P 历史 is the average of each of the first preference scores in the past; z is the second preset coefficient.

[0018] Furthermore, obtaining a first recommendation list according to the preference portrait score includes:

[0019] Obtaining different numbers of first recommended short videos from the feature categories according to the preference portrait scores;

[0020] The first recommended short videos are sorted according to the feature classification and creation time to obtain a first recommendation list; the creation time is the upload time of the first recommended short video.

[0021] Furthermore, the step of sorting the first recommended short videos according to the feature classification and the creation time point to obtain a first recommendation list includes:

[0022] Arrange the plurality of first recommended short videos in each of the feature categories in descending order according to the timestamps of the creation times to obtain a feature category video list;

[0023] Arrange the plurality of feature-classified video lists in descending order according to the preference portrait scores and then merge them to obtain the first recommendation list.

[0024] Furthermore, obtaining the viewing ratio of each feature category according to the second play time of the first recommended short video in the first recommendation list includes:

[0025] When the user watches the first recommended short video in the first recommendation list, the viewing ratio of the first feature classification is obtained according to the viewing ratio formula, and the viewing ratio formula is:

[0026] G = T3 / T4;

[0027] Among them, G is the viewing ratio of the first feature category; T3 is the second playback time, that is, the real-time total playback time of the first feature category with the short video recommended from the user's viewing recommendation list as the starting point; T4 is the real-time total playback time of all feature categories with the short video recommended from the user's viewing recommendation list as the starting point.

[0028] Furthermore, obtaining video feedback of each feature classification according to the viewing ratio of each feature classification includes:

[0029] If the viewing ratio is greater than or equal to a preset ratio, the video feedback of the feature classification corresponding to the viewing ratio is positive feedback;

[0030] If the viewing ratio is less than a preset ratio, the video feedback of the feature classification corresponding to the viewing ratio is negative feedback.

[0031] Furthermore, the method further comprises:

[0032] If the first recommended short video in the first recommendation list is liked or collected, the video feedback of the corresponding feature classification is positive feedback;

[0033] If the first recommended short video in the first recommendation list is disliked or blacklisted, the video feedback of the corresponding feature classification is negative feedback.

[0034] Further, updating the first recommendation list according to the video feedback to obtain a second recommendation list includes:

[0035] Traversing all video feedbacks corresponding to the feature classification at preset time intervals, and if the video feedback is positive feedback, obtaining a preset number of second recommended short videos from the corresponding feature classification of the total video library;

[0036] The second recommended list is obtained by adding the second recommended short video to the first recommended list.

[0037] Furthermore, the method further comprises:

[0038] If the number of recommended short videos in the second recommendation list is less than a preset number, the number of recommended short videos is supplemented to the preset number based on the preference portrait score.

[0039] In a second aspect, a short video recommendation device is provided, the device comprising:

[0040] A portrait score module, which obtains a preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category; the feature categories include short video tags, short video authors, short video series, short video major categories, short video secondary categories and short video item portraits;

[0041] A first recommendation module, configured to obtain a first recommendation list according to the preference portrait score; the first recommendation list includes a plurality of first recommended short videos;

[0042] A viewing ratio module, used to obtain a viewing ratio of each feature category according to the second playback duration of the short video in the recommendation list;

[0043] A feedback module, used for obtaining video feedback of each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback;

[0044] The target recommendation module is used to update the first recommendation list according to the video feedback to obtain a second recommendation list, and recommend videos to the user according to the second recommendation list.

[0045] The present invention adopts the above technical solution and has at least the following beneficial effects:

[0046] A short video recommendation method and device are provided. The preference portrait score corresponding to each feature category is obtained based on the first playback time of the user in each feature category, a first recommendation list is obtained according to the preference portrait score, a viewing ratio of each feature category is obtained according to the second playback time of the short videos in the first recommendation list, video feedback of each feature category is obtained according to the viewing ratio of each feature category, the first recommendation list is updated according to the video feedback to obtain a second recommendation list, and videos are recommended to the user according to the second recommendation list; the recommended videos obtained based on the preference portrait score, the viewing ratio and the video feedback are more in line with the user's preferences. In addition, since there is no need to write and maintain two sets of codes, the development and maintenance costs can be effectively reduced.

[0047] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 is a flow chart of a short video recommendation method shown in an exemplary embodiment of the present invention;

[0050] Figure 2 It is a schematic block diagram of a short video recommendation device shown in an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0052] In the prior art, the short video reservoir adopts the Lambda architecture of the recommendation system. The Lambda architecture is an architecture mode for efficiently processing large-scale data. It combines batch processing and real-time processing to meet the needs of big data processing and provide high-performance, low-latency and fault-tolerant data processing solutions. However, in practical applications, it is also necessary to select and adjust according to specific scenarios and needs. Problems with the prior art: High development and maintenance costs: The Lambda architecture needs to maintain two systems, the batch processing layer and the real-time processing layer, at the same time, which increases the complexity of development and maintenance. Developers need to write and maintain two sets of code to ensure data consistency and accuracy between the two systems. High resource consumption: Since the batch processing layer and the real-time processing layer need to be run at the same time, the Lambda architecture consumes more computing resources and storage resources. This may lead to rising costs, especially when the amount of data is very large. Data redundancy: In the Lambda architecture, the batch processing layer and the real-time processing layer may generate redundant data views. This may lead to data inconsistency and additional storage overhead. In order to maintain data consistency, additional mechanisms are required to synchronize and merge the data in the two views. Complex deployment and migration: The deployment and migration process of the Lambda architecture is relatively complex. Because the configuration and dependencies of both the batch and real-time processing layers need to be considered, it may take more time and resources to complete the deployment and migration work.

[0053] The embodiments of the present application provide a short video recommendation method and device, in which the recommended videos obtained based on the preference portrait score, viewing ratio and video feedback are more in line with the user's preferences; since there is no need to write and maintain two sets of codes, the development and maintenance costs can be effectively reduced.

[0054] The method in this application is described below through specific examples.

[0055] See also Figure 1 , Figure 1 is a flowchart of a short video recommendation method shown in an exemplary embodiment of the present invention, see Figure 1 , the method comprising:

[0056] Step S11, obtaining a preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category; the feature categories include short video tags, short video authors, short video series, short video categories, short video secondary categories and short video item portraits;

[0057] Step S12: obtaining a first recommendation list according to the preference portrait score; the first recommendation list includes a plurality of first recommended short videos;

[0058] Step S13, obtaining a viewing ratio of each feature category according to the second playback duration of the short video in the first recommendation list;

[0059] Step S14, obtaining video feedback of each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback;

[0060] Step S15: Update the first recommendation list according to the video feedback to obtain a second recommendation list, and recommend videos to the user according to the second recommendation list.

[0061] It should be noted that the short video recommendation method provided in this embodiment is applicable to scenarios including, but not limited to: short video recommendation reservoir and other scenarios involving video recommendation in specific practice.

[0062] It can be understood that the short video recommendation method provided in this embodiment obtains the preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category, obtains the first recommendation list according to the preference portrait score, obtains the viewing ratio of each feature category according to the second playback time of the short video in the first recommendation list, obtains video feedback of each feature category according to the viewing ratio of each feature category, updates the first recommendation list according to the video feedback to obtain the second recommendation list, and recommends videos to the user according to the second recommendation list; since there is no need to write and maintain two sets of codes, the development and maintenance costs can be effectively reduced. At the same time, the recommended videos obtained based on the preference portrait score, viewing ratio and video feedback are more in line with the user's preferences.

[0063] In specific practice, in step S11, "obtaining the preference portrait score corresponding to each feature category based on the first playback duration of the user in each feature category" includes: obtaining the current preference score according to the first preference score formula, and the current preference score formula is: P 当前 =T1 / T2; where P 当前 is the first preference score; T1 is the first playback duration, that is, the total playback duration of the feature classification within the preset number of days before the current time node; T2 is the total playback duration of all feature classifications within the preset number of days before the current time node;

[0064] The preference portrait score is obtained according to the preference portrait score formula. The preference portrait score formula is: P 偏好 =P 当前 ×s+P 历史 ×z; where P 偏好 is the preference profile score; P 当前 is the first preference score; s is the first preset coefficient; P 历史 is the mean of each previous first preference score; z is the second preset coefficient.

[0065] It should be noted that the preset number of days can be set according to specific business needs, and the default setting is 14 days.

[0066] Specifically, the preference portrait score is the user's preference for the short video attribute feature classification. The calculated result is a decimal between 0 and 1. The larger the value, the more the user likes short videos of this type. Each calculation can obtain the current 14-day playback behavior record to represent the current user's recent preferences.

[0067] In specific practice, "obtaining a first recommendation list based on preference portrait scores" in step S12 includes: obtaining different numbers of first recommended short videos from feature categories based on preference portrait scores; sorting the first recommended short videos based on feature categories and creation time to obtain a first recommendation list; the creation time is the upload time of the first recommended short video.

[0068] It should be noted that the number of first recommended short videos obtained in each feature category is: the total number of list settings multiplied by the preference portrait score of the feature category.

[0069] Specifically, the first recommended short videos are sorted according to feature classification and creation time point to obtain a first recommendation list, including: arranging multiple first recommended short videos in each feature classification in descending order according to the timestamp of the creation time to obtain a feature classification video list; arranging multiple feature classification video lists in descending order according to the preference portrait score and then merging them to obtain the first recommendation list.

[0070] In specific practice, step S13 "obtaining the viewing ratio of each feature category based on the second playback time of the short video in the first recommendation list" includes: in the process of the user watching the first recommended short video in the first recommendation list, obtaining the viewing ratio of the first feature category according to the viewing ratio formula, and the viewing ratio formula is: G=T3 / T4; wherein G is the viewing ratio of the first feature category; T3 is the second playback time, that is, the real-time total playback time of the first feature category from the time when the user watches the recommended short video in the recommendation list; T4 is the real-time total playback time of all feature categories from the time when the user watches the recommended short video in the recommendation list.

[0071] In specific practice, step S14 "obtaining video feedback for each feature category based on the viewing ratio of each feature category" includes: if the viewing ratio is greater than or equal to the preset ratio, the video feedback for the feature category corresponding to the viewing ratio is positive feedback; if the viewing ratio is less than the preset ratio, the video feedback for the feature category corresponding to the viewing ratio is negative feedback.

[0072] It should be noted that the preset ratio is set according to the specific accuracy requirements and is set to 0.25 by default.

[0073] Specifically, if the first recommended short video in the first recommendation list is liked or collected, the video feedback of the corresponding feature classification is positive feedback; if the first recommended short video in the first recommendation list is disliked or blacklisted, the video feedback of the corresponding feature classification is negative feedback.

[0074] In specific practice, step S15 "update the first recommendation list based on video feedback to obtain the second recommendation list" includes: traversing the video feedback corresponding to all feature categories every preset time period, if the video feedback is positive feedback, obtaining a preset number of second recommended short videos from the corresponding feature category of the total video library; adding the second recommended short video to the first recommendation list to obtain the second recommendation list.

[0075] It should be noted that the preset duration is set according to the update requirements and is generally set to 5 minutes.

[0076] It can be understood that the recommendation method provided in this embodiment will update the recommended short videos in the recommendation list at preset intervals, so that the user can always browse the appropriate video content.

[0077] In a specific embodiment, if the number of recommended short videos in the second recommendation list is less than a preset number, the number of recommended short videos is supplemented to the preset number based on the preference portrait score.

[0078] It should be noted that the preset number is set according to the storage space and the default setting is 2000.

[0079] In a specific embodiment, a first recommendation list may be obtained based on a content recommendation algorithm, wherein content recommendation includes: recalling the same album, calculating short videos of the same album, and sorting by creation time, with the latest short video at the front.

[0080] Similar recall of the same category label calculates the similarity between short videos under the same category based on secondary classification, key type, scene, region, voice, key IP, character, emotion and other short video labels, and sorts them by creation time, with the latest short video at the front.

[0081] Similarity recall of different classification labels, based on secondary classification, key types, screen scenes, regions, voices, key IPs, characters, emotions and other label short video labels, calculates the similarity between short videos under different major categories, and sorts them by creation time, with the latest short videos sorted at the front.

[0082] Unlabeled similarity recall, no labels are the same, there is no label relationship between short videos, sorted by creation time, the latest short video is sorted at the front.

[0083] Recall the new and hot videos in major categories, aggregate and calculate the latest short videos under the same major category, and sort them by creation time, with the latest short videos at the front.

[0084] The reservoir is initialized and recalled, large categories are broken up, short videos of the same series are discontinuous, sorted by creation time, with the latest short video at the front, and initial scores are assigned.

[0085] Personalized recommendations for user portraits are based on user short video portrait preferences and short video features for similarity calculation. Based on the user portrait's preference score for features, the short video's popularity UV and online time, the most likely short videos to be of interest are recommended based on the calculated scores and sorted.

[0086] User footprint recall, short videos that are relevant to historical short videos are calculated through user historical behaviors, such as footprint recall of the same series, footprint recall of similar tags, footprint recall of TV series, and footprint recall of popular shows in the same category.

[0087] In a specific embodiment, the first recommendation list can be obtained based on a collaborative filtering algorithm. Collaborative filtering collects user feedback logs and analyzes user behavior data such as clicks, plays, likes, shares, and dislikes to establish a model to form a user-short video rating matrix, where each row represents a user, each column represents a short video, and the elements in the matrix represent the user's rating of the short video. The cosine similarity formula is used to calculate the similar short videos watched by users. The cosine similarity calculation formula is:

[0088]

[0089] Where u and v are the rating vectors of user u and user v respectively, · represents the vector dot product, and ||*|| represents the modulus of the vector.

[0090] In a specific embodiment, the first recommendation list can be obtained based on the recommendation algorithm of the machine learning model, and the random forest supervised learning and BERT unsupervised pre-training model are combined to train and calculate the high-quality short video recall, which specifically includes: analyzing user feedback logs to establish training data such as: short video id, number of plays less than or equal to 2 seconds, number of plays, number of people who play less than or equal to 2 seconds, number of people who play, average play percentage (excluding those less than 2 seconds and = 100%), number of days online, number of likes, number of likes, number of full-length / feature-length views, number of full-length / feature-length views, number of completed views, number of completed views, number of created orders, number of successful orders, number of long video created orders, number of successful long video orders, etc. Random forests achieve reasonable standards for indicators such as accuracy, precision, recall, and F1 score through continuous decision tree construction and parameter setting. The core idea of ​​the BERT model is to perform unsupervised pre-training on massive data, and to train the model by fine-tuning parameters such as learning rate, batch size, number of pre-training steps, number of layers, hidden layer size, number of attention heads, etc., so that the model's accuracy, precision, recall rate, F1 score and other indicators reach reasonable standards.

[0091] See also Figure 2 , Figure 2 is a schematic block diagram of a short video recommendation device according to an exemplary embodiment of the present invention, see Figure 2 , the short video recommendation device 100 includes:

[0092] The portrait score module 101 obtains the preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category; the feature categories include short video tags, short video authors, short video series, short video major categories, short video secondary categories and short video item portraits;

[0093] The first recommendation module 102 is used to obtain a first recommendation list according to the preference portrait score; the first recommendation list includes multiple first recommended short videos;

[0094] A viewing ratio module 103, configured to obtain a viewing ratio of each feature category according to the second play time of the short video in the recommendation list;

[0095] Feedback module 104, used to obtain video feedback of each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback;

[0096] The target recommendation module 105 is used to update the first recommendation list according to the video feedback to obtain a second recommendation list, and recommend videos to the user according to the second recommendation list.

[0097] It should be noted that the short video recommendation device provided in this embodiment is applicable to scenarios including, but not limited to: a short video recommendation reservoir and other scenarios involving video recommendation in specific practice.

[0098] Specifically, the preset nutritional elements are set according to specific application scenarios, and are generally nutritional elements that are relatively important to the human body.

[0099] Specifically, the short video item portrait information comes from the existing SVS system.

[0100] It can be understood that the short video recommendation device provided in this embodiment obtains the preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category, obtains the first recommendation list according to the preference portrait score, obtains the viewing ratio of each feature category according to the second playback time of the short video in the first recommendation list, obtains video feedback of each feature category according to the viewing ratio of each feature category, updates the first recommendation list according to the video feedback to obtain the second recommendation list, and recommends videos to the user according to the second recommendation list; since there is no need to write and maintain two sets of codes, the development and maintenance costs can be effectively reduced. At the same time, the recommended videos obtained based on the preference portrait score, the viewing ratio and the video feedback are more in line with the user's preferences.

[0101] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

[0102] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0103] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A short video recommendation method, characterized in that: The method comprises: Based on the first playback time of the user in each feature category, the preference portrait score corresponding to each feature category is obtained; the feature categories include short video tags, short video authors, short video series, short video major categories, short video secondary categories and short video item portraits; Obtaining a first recommendation list according to the preference portrait score; the first recommendation list includes a plurality of first recommended short videos; Obtaining a viewing ratio of each feature category according to the second playback duration of the short video in the first recommendation list; Obtaining video feedback for each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback; The first recommendation list is updated according to the video feedback to obtain a second recommendation list, and videos are recommended to the user according to the second recommendation list.

2. The recommendation method according to claim 1, characterized in that: The method of obtaining the preference portrait score corresponding to each feature category based on the first playback duration of the user in each feature category includes: The current preference score is obtained according to the first preference score formula, and the current preference score formula is: P 当前 =T1 / T2; Among them, P 当前 is the first preference score; T1 is the first playback duration, that is, the total playback duration of the feature classification within the preset number of days before the current time node; T2 is the total playback duration of all feature classifications within the preset number of days before the current time node; The preference portrait score is obtained according to the preference portrait score formula, and the preference portrait score formula is: P 偏好 =P 当前 ×s+P 历史 ×z; Among them, P 偏好 is the preference profile score; P 当前 is the first preference score; s is the first preset coefficient; P 历史 is the average of each of the first preference scores in the past; z is the second preset coefficient.

3. The recommendation method according to claim 1, characterized in that: The obtaining of a first recommendation list according to the preference portrait score includes: Obtaining different numbers of first recommended short videos from the feature categories according to the preference portrait scores; The first recommended short videos are sorted according to the feature classification and creation time to obtain a first recommendation list; the creation time is the upload time of the first recommended short video.

4. The recommendation method according to claim 3, characterized in that: The step of sorting the first recommended short videos according to the feature classification and the creation time point to obtain a first recommendation list includes: Arrange the plurality of first recommended short videos in each of the feature categories in descending order according to the timestamps of the creation times to obtain a feature category video list; Arrange the plurality of feature-classified video lists in descending order according to the preference portrait scores and then merge them to obtain the first recommendation list.

5. The recommendation method according to claim 1, characterized in that: The obtaining of the viewing ratio of each feature category according to the second playback duration of the first recommended short video in the first recommendation list includes: When the user watches the first recommended short video in the first recommendation list, the viewing ratio of the first feature classification is obtained according to the viewing ratio formula, and the viewing ratio formula is: G = T3 / T4; Among them, G is the viewing ratio of the first feature category; T3 is the second playback time, that is, the real-time total playback time of the first feature category with the short video recommended from the user's viewing recommendation list as the starting point; T4 is the real-time total playback time of all feature categories with the short video recommended from the user's viewing recommendation list as the starting point.

6. The recommendation method according to claim 5, characterized in that: The video feedback of each feature classification is obtained according to the viewing ratio of each feature classification, including: If the viewing ratio is greater than or equal to a preset ratio, the video feedback of the feature classification corresponding to the viewing ratio is positive feedback; If the viewing ratio is less than a preset ratio, the video feedback of the feature classification corresponding to the viewing ratio is negative feedback.

7. The recommendation method according to claim 6, characterized in that: The method further comprises: If the first recommended short video in the first recommendation list is liked or collected, the video feedback of the corresponding feature classification is positive feedback; If the first recommended short video in the first recommendation list is disliked or blacklisted, the video feedback of the corresponding feature classification is negative feedback.

8. The recommendation method according to any one of claims 6 or 7, characterized in that: The updating the first recommendation list according to the video feedback to obtain a second recommendation list includes: Traversing all video feedbacks corresponding to the feature classification at preset time intervals, and if the video feedback is positive feedback, obtaining a preset number of second recommended short videos from the corresponding feature classification of the total video library; The second recommended list is obtained by adding the second recommended short video to the first recommended list.

9. The recommendation method according to claim 8, characterized in that: The method further comprises: If the number of recommended short videos in the second recommendation list is less than a preset number, the number of recommended short videos is supplemented to the preset number based on the preference portrait score.

10. A short video recommendation device, characterized in that: The device comprises: A portrait score module, which obtains a preference portrait score corresponding to each feature category based on the first playback time of the user in each feature category; the feature categories include short video tags, short video authors, short video series, short video major categories, short video secondary categories and short video item portraits; A first recommendation module, configured to obtain a first recommendation list according to the preference portrait score; the first recommendation list includes a plurality of first recommended short videos; A viewing ratio module, used to obtain a viewing ratio of each feature category according to the second playback duration of the short video in the recommendation list; A feedback module, used for obtaining video feedback of each feature classification according to the viewing ratio of each feature classification; the video feedback includes positive feedback and negative feedback; The target recommendation module is used to update the first recommendation list according to the video feedback to obtain a second recommendation list, and recommend videos to the user according to the second recommendation list.

Images