Short video recommendation method and system for sparse scene
By using benchmark short video queue evaluation and dynamic adjustment methods, the problems of cold start for new users and data sparsity were solved, enabling diversity and personalization of short video recommendations, and improving recommendation performance and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing short video recommendation methods rely too heavily on sparse data or pursue short-term metrics during the cold start phase for new users, resulting in insufficient interest exploration, premature homogenization of recommendations, difficulty in accurately identifying user interests and preferences, and a tendency to create information cocoons.
A benchmark short video recommendation queue is used to evaluate the interests of new users. Through category tag filtering and secondary recommendation, combined with the interest distribution concentration index, dynamic adjustment is carried out to ensure the diversity and personalization of recommended content and avoid interest solidification.
It enables the rapid capture of user interests in sparse scenarios, improves recommendation accuracy and user engagement, prevents information cocoons, enhances system adaptability and robustness, and promotes the healthy distribution of the content ecosystem.
Smart Images
Figure CN122019832A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of video recommendation technology, specifically relating to a short video recommendation method and system for sparse scenarios. Background Technology
[0002] With the rapid development of the mobile internet, short video platforms have sprung up like mushrooms after rain and quickly become popular. In their leisure time, people are increasingly inclined to use short videos to obtain information, relax, and socialize.
[0003] Existing short video recommendation methods, whether widely used collaborative filtering algorithms or emerging models based on graph neural networks or deep learning, all face significant limitations when handling such scenarios. Specifically, they mainly rely on sufficient historical user behavior data to construct user profiles or establish interest associations. However, for new users, their historical interaction information is completely lacking, making it difficult for traditional collaborative filtering algorithms to find similar users for effective recommendations. Content-based recommendations also struggle due to the inability to obtain users' initial preferences, a typical cold start problem. Even after users have engaged in some initial behavior, the fragmented and rapid switching characteristics of short video consumption mean that effective feedback data generated by individual users is often very scarce and highly limited. Randomness leads to an extremely sparse user-item interaction matrix, making it difficult for traditional models to accurately extract stable and reliable long-tail interests, resulting in significant fluctuations in recommendation quality in the early stages. Secondly, existing methods are prone to causing imbalances in the interest exploration process for new users in practice, and prematurely trapping them in information cocoons. For example, many recommendation algorithms, after capturing a user's brief stay or accidental interaction with a certain type of content, will quickly amplify the proportion of such content pushed in order to improve short-term interaction metrics. However, this strategy ignores the broad-spectrum exploration needs of new users' interests and may prematurely solidify their interest judgments due to the randomness of their early behavior, limiting their opportunities to access content in other potential interest areas, which is detrimental to the healthy development of the platform ecosystem and long-term user retention.
[0004] To address the aforementioned issues, this invention proposes a short video recommendation method and system for sparse scenarios. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a short video recommendation method and system for sparse scenarios, solving the problems of insufficient interest exploration and premature homogenization of recommendations caused by over-reliance on sparse data or pursuit of short-term metrics in the cold start phase for new users and data sparse phases.
[0006] The objective of this invention can be achieved through the following technical solutions: A short video recommendation method for sparse scenarios, the method comprising: Step 1: Acquire new users based on the short video platform, perform recommendation operations on the pre-built benchmark short video recommendation queue to the new users, record the viewing data of each benchmark short video from the time the new user starts viewing the benchmark short video recommendation queue to the time the viewing is completed, evaluate the new user's interest index for each benchmark short video, and use it as the interest index of the major category tag associated with each benchmark short video. Step 2: Calculate the short video recommendation ratio associated with each major category tag based on the interest index among the major category tags; Based on the short video recommendation ratio, the short videos under all subcategories of major categories are filtered, a secondary recommendation sequence of short videos associated with each major category is constructed, and the recommendation operation is performed to new users, and secondary viewing data is recorded. Step 3: Based on the secondary browsing data, reassess the new users' interest index on major category tags, construct a concentration index of the new users' interest distribution, perform solidification verification, determine whether to carry out balancing adjustment based on the solidification results, and perform recommendation operations.
[0007] As a further aspect of the present invention, in step one, the specific method for acquiring new users based on the short video platform and performing recommendation operations on the pre-constructed benchmark short video recommendation queue to the new users is as follows: Acquire new users who register on the short video platform, and denote them as X; Take a pre-built benchmark short video recommendation queue, where any benchmark short video corresponds to a major category of tags; Take the total number of pre-defined category tags j in the short video platform; Get any major category label YBi, where i is the counting index, with a value ranging from 1 to j; Get the number of likes, comments, shares, and favorites for all short videos under the major category tag YBi; Obtain the preset weights for calculating likes, comments, shares, and favorites; The selection index associated with each short video under the category tag YBi is calculated by weighted summation. The short video with the highest selection index is selected as the benchmark short video associated with the category tag YBi, and is denoted as Vi. Similarly, determine the benchmark short videos associated with each of the j major category tags, and randomly arrange the j major category tags, denoted as the major category tag sequence YB1, YB2, ..., YBj; A baseline short video recommendation queue V1, V2, ..., Vj is constructed based on the category label sequence, where YBi corresponds to Vi; Recommendation operations are performed on new user X based on the baseline short video recommendation queues V1, V2, ..., Vj.
[0008] As a further aspect of the present invention, the specific method for assessing the interest index of new users for each benchmark short video in step one is as follows: S31. Record the time span from when a new user X starts browsing the baseline short video recommendation queue to when they finish browsing, as the first monitoring period; S32. Obtain the viewing data of any benchmark short video Vi viewed by new user X within the monitoring period, including actual playback time PTi, total short video duration ATi, like flag Li, comment flag Ci, forward flag Si, and favorite flag Fi. If new user X performs a like operation on benchmark short video Vi, then Li=1; otherwise, Li=0. The same applies to Ci, Si, and Fi. S33. Calculate the completion rate Ri and effective dwell ratio ERi of new user X for the baseline short video Vi, where Ri = PTi / TTi. If Ri≥1, then ERi=1, otherwise ERi=Ri; S34. Obtain the preset playback behavior weight α1 and interaction behavior base weight α2 of the operator, where α1+α2=1; S35. Calculate the interest component P_Scorei=ERi based on playback behavior; Calculate the interest component U_Scorei based on interactive behavior: U_Scorei=(Li×w1+Ci×w2+Si×w3+Fi×w4) / (w1+w2+w3+w4), where w1, w2, w3, and w4 are the preset weights for likes, comments, shares, and favorites, respectively. S36. Calculate the interest index Qi of new user X on the benchmark short video Vi, which is Qi = α1*P_Scorei + α2*U_Scorei. Use the interest index Qi as the interest index of new user X on the category tag YBi associated with the benchmark short video Vi. S37. Similarly, determine the interest index of new user X for each benchmark short video and the interest index of the major category tags associated with each benchmark short video, and form the interest index sequence Q1, Q2, ..., Qj.
[0009] As a further aspect of the present invention, the specific method for calculating the short video recommendation ratio associated with each major category tag based on the interest index between the major category tags in step two is as follows: Calculate the sum of the j interest indices in the interest index sequence Q1, Q2, ..., Qj, denoted as the total interest index Tot_Q = Q1 + Q2 + ... + Qj; Calculate the initial recommendation ratio Ei = Qi / Total_Q for the major category label YBi; Obtain the preset minimum recommendation ratio threshold ε, where 0 < ε < 1 / j; If Ei≥ε, let the undetermined proportion of the major category label YBi be DEi=Ei; Conversely, let the undetermined proportion DEi = ε; Similarly, by determining the undetermined proportions of each major category label, we obtain DE1, DE2, ..., DEj; Calculate the total undetermined proportion Tot_DE = DE1 + DE2 + ... + DEj; Calculate the short video recommendation ratio Ki=DEi / Tot_DE for the major category tag YBi; Similarly, by determining the short video recommendation ratio K1 for each major category tag, we obtain K1, K2, ..., Kj.
[0010] As a further aspect of the present invention, the specific method for constructing the secondary recommendation short video sequences associated with each of the major category tags in step two is as follows: Take the preset total number of secondary recommended short videos, denoted as M; Based on the short video recommendation ratio of each major category tag, the number of short videos to be assigned in the secondary recommendation for each major category tag is calculated. Among them, the number of short videos to be assigned for the major category tag YBi is Ni=M*Ki. Ni is rounded down and the rounded value is assigned to the number of short videos to be assigned Ni. Obtain the overall popularity value of all subcategories under the major category tag YBi. The overall popularity value is calculated based on the weighted average sum of the number of views and interactions of all short videos under the corresponding subcategory tag at the current time. Play count is the number of times all short videos under the subcategory tag have been played, and interaction count is the sum of the number of likes, comments, shares, and favorites for all short videos under the subcategory tag. Sort all subcategories under the main category YBi in descending order of overall popularity value. Starting from the first subcategory, select the short videos with the highest individual popularity value under each subcategory until Ni short videos are selected. The individual popularity value is calculated based on the average weighted sum of the views and interactions of any single short video. Arrange the Ni videos in the order of selection into a secondary recommendation short video sequence associated with the major category tag YBi, denoted as Hi; Similarly, determine the sequence of short videos for secondary recommendations for each major category tag, denoted as H1, H2, ..., Hj.
[0011] As a further aspect of the present invention, step two further includes performing a recommendation operation on the determined secondary recommendation short video sequence to the new user X, wherein the recommendation method is a fully random recommendation, until all short videos in the j secondary recommendation short video sequences H1, H2, ..., Hj have been recommended, and the secondary viewing data of the new user X for each short video is recorded, including the actual playback time, the total duration of the short video, likes, comments, shares, and favorites.
[0012] As a further aspect of the present invention, the specific method for reassessing the new user's interest index in the category tags based on the secondary viewing data in step three is as follows: Obtain the secondary viewing data of new user X for all short videos in the secondary recommended short video sequence H1, H2, ..., Hj, and determine the interest index of all short videos based on the secondary viewing data of each short video in the manner described in steps S32 to S36. The interest index of all short videos under any major category tag YBi is averaged, and the result is recorded as the interest index Qi' to be updated. Similarly, determine the interest indices to be updated for each major category tag, forming the interest index sequence Q1', Q2', ..., Qj'.
[0013] As a further aspect of the present invention, in step three, the specific method for constructing the concentration index of the interest distribution of new users and performing the solidification verification is as follows: Calculate the average value of the interest index sequence Q1', Q2', ..., Qj' to be updated: AVG_Q' = (Q1' + Q2' + ... + Qj') / j; Calculate the absolute deviation of each interest index to be updated from the average value AVG_Q', where the absolute deviation of any interest index Qi' to be updated is Di=|Qi'-AVG_Q'|. The mean absolute deviation (MAD) is calculated using MAD=(D1+D2+...+Dj) / j, and the concentration index of the interest distribution of new user X is calculated using ZB=MAD / AVG_Q'. Obtain the operator's preset interest fixation threshold θ, where θ > 0; If ZB≥θ, it is determined that the interest distribution of the new user X has become fixed and balance adjustment needs to be performed; Conversely, if the distribution of interests of new user X has not become fixed and no balancing adjustment is needed, short videos are recommended to new user X based on the short video recommendation ratios K1, K2, ..., Kj as the final ratio parameters.
[0014] As a further aspect of the present invention, the specific method for performing balance regulation and executing the recommended operation in step three is as follows: Based on the interest index sequence Q1', Q2', ..., Qj' to be updated, the short video recommendation ratio of each major category tag is recalculated, and the calculated short video recommendation ratio is used as the final ratio parameter. Based on the final ratio parameter, short videos are recommended to new user X.
[0015] A short video recommendation system for sparse scenarios, the system comprising: The new user initialization module acquires new users based on the short video platform, performs recommendation operations on the pre-built benchmark short video recommendation queue to the new users, records the viewing data of each benchmark short video from the time the new user starts viewing the benchmark short video recommendation queue to the time when the viewing is completed, evaluates the new user's interest index for each benchmark short video, and uses it as the interest index of the major category tag associated with each benchmark short video. The tag association filtering module calculates the proportion of short videos recommended by each major category of tags based on the interest index between the major categories of tags; Based on the short video recommendation ratio, the short videos under all subcategories of major categories are filtered, a secondary recommendation sequence of short videos associated with each major category is constructed, and the recommendation operation is performed to new users, and secondary viewing data is recorded. The interest update and control module, based on secondary browsing data, reassesses the interest index of new users on major category tags, constructs a concentration index of new users' interest distribution, performs solidification verification, determines whether to perform balance control based on the solidification results, and performs recommendation operations.
[0016] The beneficial effects of this invention are: This invention designs a progressive recommendation mechanism for sparse new user scenarios, employing a closed-loop process of benchmark testing, dynamic adjustment, and balance verification to solve the cold start problem. First, it uses a benchmark recommendation queue to quickly capture users' initial interests and efficiently builds interest profiles using broad category tags. Then, it dynamically allocates recommendation ratios based on interest indices, enabling interest exploration and deepening from coarse-grained to fine-grained, significantly improving recommendation accuracy and user engagement. Finally, it uses the concentration index of interest distribution for solidification verification and balance control, avoiding premature solidification of recommended content and preventing information cocoons, thus enhancing the system's adaptability and robustness. This ensures personalized and diverse recommendation results even under sparse conditions. This invention pre-constructs a baseline short video recommendation queue based on broad category tags, uses weighted indicators such as likes and comments to filter high-quality content, ensuring broad and high-quality recommendation coverage, and randomly recommends content to new users to provide a diverse initial experience. Subsequently, it monitors user playback behavior and interaction data in real time, calculates playback completion rate and interest index, and accurately assesses user interest preferences for each tag. By combining content popularity with real-time user feedback, it achieves an efficient combination of comprehensive and personalized recommendations. This allows for a rapid breakthrough in overcoming the cold start problem for new users, quickly capturing interests through data-driven approaches, thereby increasing user engagement, enhancing stickiness, and optimizing the overall recommendation effect of the platform. This invention introduces a minimum recommendation ratio threshold to ensure that even broad category tags with low interest indices receive a certain number of recommendation opportunities. This effectively prevents excessive concentration of recommended content, improves the fairness and diversity of the system, and avoids users getting trapped in information cocoons. Based on this, the recommendation ratio is calculated by combining interest index and popularity value, and videos are filtered layer by layer from broad categories to subcategories. This achieves a precise balance between personalized preferences and content popularity, making the recommendation results both relevant to user interests and broadly appealing. A fully random recommendation method is used to display video sequences in an orderly manner, and secondary viewing data is recorded to continuously optimize the recommendation algorithm, thereby driving adaptive system improvement and enhancing the user experience. This invention improves the accuracy and timeliness of interest index determination by introducing a dynamic evaluation mechanism based on secondary browsing data, and more sensitively captures short-term changes in user interests. By constructing an interest distribution concentration index, it quantifies the dispersion of user interests, intelligently identifies whether interests are solidified, and thus autonomously decides whether to initiate balancing regulation. It effectively breaks the information cocoon problem that traditional recommendation systems are prone to, and while ensuring personalized recommendations, it proactively introduces diverse content, achieving an adaptive balance between deepening user interests and expanding exploration, and promoting the healthy distribution of the content ecosystem. Attached Figure Description
[0017] The invention will now be further described with reference to the accompanying drawings.
[0018] Figure 1 This is a schematic diagram of the system described in this invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] like Figure 1 As shown, this application provides a short video recommendation system for sparse scenarios; As an embodiment 1 of this application, it specifically includes: The new user initialization module acquires new users based on the short video platform, performs recommendation operations on the pre-built benchmark short video recommendation queue to the new users, records the viewing data of each benchmark short video from the time the new user starts viewing the benchmark short video recommendation queue to the time when the viewing is completed, evaluates the new user's interest index for each benchmark short video, and uses it as the interest index of the major category tag associated with each benchmark short video. The tag association filtering module calculates the proportion of short videos recommended by each major category of tags based on the interest index between the major categories of tags; Based on the short video recommendation ratio, the short videos under all subcategories of major categories are filtered, a secondary recommendation sequence of short videos associated with each major category is constructed, and the recommendation operation is performed to new users, and secondary viewing data is recorded. The interest update and control module, based on the secondary viewing data, reassesses the interest index of new users on major category tags, constructs the interest distribution concentration index of new users, performs solidification verification, determines whether to perform balance control based on the solidification results, and performs recommendation operations.
[0021] Example 2 This embodiment further discloses a short video recommendation method for sparse scenarios based on Embodiment 1, specifically including the following: This method is a short video recommendation method for sparse scenarios, which focuses on solving the cold start and interest exploration problems of short video recommendation systems. The core goal is to quickly and dynamically explore and adapt to user interests when facing new users without historical data, while preventing interests from becoming fixed too early and entering the information cocoon.
[0022] This method mainly includes the following steps: First, new users are acquired based on the short video platform. A pre-built benchmark short video recommendation queue is then recommended to these new users. Viewing data for each benchmark short video is recorded from the start of viewing the queue until the viewing is completed. The new user's interest index for each benchmark short video is assessed and used as the interest index for the major category tags associated with each benchmark short video. Specifically: Blind testing was conducted on new users using a pre-built benchmark short video recommendation queue covering multiple major categories to quickly map their general interest profiles. Both major and minor categories are determined based on tags customized by the short video platform. For example, if there is a major category tag of "sports", then its minor categories include: basketball, football, e-sports, etc.
[0023] Next, this diverse benchmark short video recommendation queue is pushed to new users, and their viewing data is recorded throughout the process. Based on the viewing data, the interest index of the new user for each benchmark short video is calculated, which serves as the initial interest index of the new user for each major category tag. It should be noted that the scope of the benchmark short video recommendation queue needs to be as broad as possible, covering as many major categories tags as possible in the short video platform.
[0024] Then, the recommendation ratio of short videos associated with each major category tag is calculated based on the interest index between the major category tags; Based on the short video recommendation ratio, short videos under all subcategories of major categories are filtered, and a secondary recommendation sequence of short videos associated with each major category is constructed. This sequence is then recommended to new users, and secondary viewing data is recorded. Specifically: This step aims to make more accurate and proportionate recommendations based on the initial interest profile, and to collect more in-depth feedback. Based on the interest indices of each major category obtained in the previous step, calculate the proportion of each major category tag in the recommendation operation of the next step, and construct a secondary recommendation short video sequence.
[0025] We will push more personalized short video sequences to new users and record secondary viewing data in detail again to reflect the new users’ true preferences for content they may be interested in.
[0026] Finally, based on the secondary viewing data, the interest index of new users on major category tags is reassessed, a concentration index of new user interest distribution is constructed, a solidification verification is performed, and based on the solidification results, it is determined whether to perform balancing adjustments and execute recommendation operations. Specifically: This step aims to assess whether new users' interests have become too concentrated too early, and to intervene if necessary, in order to maintain the exploratory nature and ecological health of the recommendation system. Based on the more accurate secondary browsing data in the second step, the interest index of new users on major category tags is recalculated, and the score at this time is more accurate than that in the first step. Calculate an interest distribution concentration index. For example, the interest distribution concentration index quantifies whether a user's interests are widely dispersed across multiple categories or extremely concentrated in one or two categories. If the concentration index shows that the user's interests are very concentrated and stable, for example, more than 95% of the interactions are concentrated in the "games" category, then the solidification verification is considered to have passed. Once it is confirmed that an interest has become too fixed, the system will not completely comply, but will actively carry out balancing adjustments. That is, in subsequent recommendations, other major categories will be consciously and slightly mixed in, even high-quality content with low current interest index, in order to test the user's potential other interests and break the information cocoon. Based on whether balancing adjustments are carried out and the operation process of balancing adjustments, the final recommendation list is generated and pushed to new users.
[0027] Example 3 This embodiment, based on embodiment 2, further discloses a method for evaluating new users' interest index in various benchmark short videos, specifically including the following: This embodiment provides a detailed explanation of the implementation of the first step described in Embodiment 2, specifically outlining how to construct the baseline recommendation queue and the detailed method for calculating the interest index, as follows: First, obtain the short video platform to be processed, and obtain any newly registered user from the short video platform. Mark this new user as X, and use the new user X as an example for subsequent processing.
[0028] First, obtain a pre-built benchmark short video recommendation queue, where any benchmark short video represents a major category of tags. Obtain the total number of pre-labeled major category tags in the short video platform in real time and denot it as j. For example, retrieve all predefined category tags from the platform, with a total of j=10, representing 10 categories: sports, food, beauty, technology, film and television, music, games, comedy, travel, and cute pets.
[0029] Get any one of the j major category labels and denote it as YBi, where i is the counting index, and the value range is from 1 to j; Next, obtain the number of likes, comments, shares, and favorites for all short videos under the major category tag YBi; It also obtains the weights preset by the operator for calculating likes, comments, shares, and favorites. By employing a weighted summation method, the optimization index associated with each short video under the major category tag YBi is calculated. The short video with the highest optimization index is then used as the benchmark short video Vi associated with the major category tag YBi. In other words, the short video with the highest overall popularity is selected for blind testing on new users to improve the success rate.
[0030] Next, repeat the above steps to determine the benchmark short videos associated with each of the j major category tags, resulting in j benchmark short videos. Then, sort the j benchmark short videos based on the randomized sorting of the j major category tags YB1, YB2, ..., YBj, to obtain the benchmark short video recommendation queue V1, V2, ..., Vj, where any major category tag YBi corresponds to a benchmark short video Vi.
[0031] Based on the established baseline short video recommendation queues V1, V2, ..., Vj, recommendation operations are performed on new user X to ensure that the new user can quickly access all mainstream content directions within the short video platform from the outset.
[0032] Next, the interest index of new users for each benchmark short video is evaluated. First, it is necessary to determine the time span from when new user X starts browsing the benchmark short video recommendation queue until the browsing of the benchmark short video recommendation queue is completed, and this time span is used as the first monitoring period.
[0033] Then, the viewing data of new user X during the monitoring period is obtained, including the actual playback duration, total video duration, likes, comments, shares, and favorites of each benchmark short video. For ease of explanation, any benchmark short video Vi in the benchmark short video recommendation queue V1, V2, ..., Vj is used as an example. The viewing data of benchmark short video Vi includes: actual playback duration PTi, total video duration ATi, likes Li, comments Ci, shares Si, and favorites Fi.
[0034] It needs to be explained that if a new user X performs a "like" operation on the benchmark short video Vi, then the like flag Li=1; otherwise, the like flag Li=0. The same applies to the comment flag Ci, the share flag Si, and the favorite flag Fi.
[0035] Next, the completion rate Ri of the new user X for the baseline short video Vi is calculated by using Ri=PTi / TTi; Next, the playback completion rate Ri is judged. If Ri≥1, the effective dwell ratio ERi of new user X on the benchmark short video Vi is set to 1; otherwise, the effective dwell ratio ERi is set to Ri. It should be noted that if Ri≥1, the viewing time may exceed the video itself due to repeated playback or pausing, so ERi=1; otherwise, ERi=Ri, and the upper limit of the playback completion rate is locked at 100% to prevent data anomalies.
[0036] Next, obtain the preset playback behavior weight α1 and the preset interaction behavior base weight α2 of the operator for the playback behavior, where α1+α2=1 and α1 and α2 are both greater than 0.
[0037] By using P_Scorei=ERi, the interest component P_Scorei of the new user X's playback behavior during the viewing of the benchmark short video Vi is determined; The interest component U_Scorei of a new user X's interactive behavior during the viewing of a benchmark short video Vi is calculated using U_Scorei=(Li×w1+Ci×w2+Si×w3+Fi×w4) / (w1+w2+w3+w4), where w1, w2, w3, and w4 are the preset weights for likes, comments, shares, and favorites, respectively.
[0038] Next, the interest index Qi of new user X on benchmark short video Vi is calculated by using Qi=α1*P_Scorei+α2*U_Scorei, and the calculated interest index Qi is used as the interest index of new user X on the category tag YBi associated with benchmark short video Vi.
[0039] By repeating the above steps, we can determine the interest index of new user X for each benchmark short video in the benchmark short video recommendation queue V1, V2, ..., Vj, as well as the interest index for the major category tags associated with each benchmark short video. Finally, we can form the interest index sequence Q1, Q2, ..., Qj.
[0040] Example 4 This embodiment, based on embodiment 3, further discloses a method for constructing secondary recommendation short video sequences associated with each major category tag and recommending them to new users, while recording secondary viewing data. Specifically, it includes the following: First, the recommendation ratio of short videos associated with each major category tag is calculated based on the interest index among the major category tags; Calculate the sum of the j interest indices in the interest index sequence Q1, Q2, ..., Qj, and record the calculated sum as the total interest index Tot_Q = Q1 + Q2 + ... + Qj to determine the overall interest level shown by the new user X in the initial blind test.
[0041] Next, the initial recommendation ratio Ei of any major category tag YBi is calculated using the formula: Ei=Qi / Total_Q, and the interest index is converted into a percentage.
[0042] Then obtain the minimum recommended ratio threshold ε preset by the operator, where 0 < ε < 1 / j.
[0043] The determined initial recommendation ratio Ei is compared with the minimum recommendation ratio threshold ε. If Ei ≥ ε, the undetermined ratio DEi of the category tag YBi is set to Ei. If Ei < ε, the undetermined ratio DEi is set to ε. This prevents some category tags from being completely abandoned in subsequent recommendations due to the initial interest of new users being unclear or zero, thus preventing them from falling into an information cocoon.
[0044] Repeat the above steps to calculate the undetermined proportions of each major category label, and arrange them in the order of the major category label sequence YB1, YB2, ..., YBj, represented as: DE1, DE2, ..., DEj.
[0045] Next, the total undetermined proportion Tot_DE is calculated using Tot_DE=DE1+DE2+...+DEj.
[0046] Then, the short video recommendation ratio Ki of the major category tag YBi associated with the new user X is calculated by using Ki=DEi / Tot_DE; At this point, the short video recommendation ratio Ki for the major category tag YBi has been determined. Based on this method, the short video recommendation ratios for the remaining major category tags are determined and arranged in the order of the major category tag sequence YB1, YB2, ..., YBj, represented as: K1, K2, ..., Kj.
[0047] Obtain the total number M of secondary recommended short videos preset by the operator for the secondary recommended short video sequence. In the end, the total number of secondary recommended short videos in the actual calculated secondary recommended short video sequence may not be equal to M, depending on the actual situation. Then, combining the short video recommendation ratio of each major category tag, calculate the number of short videos to be allocated in the secondary recommendation for each major category tag. For example, the number of short videos to be allocated for the major category tag YBi is Ni=M*Ki. If Ni is not an integer, then Ni is rounded down (e.g., Ni=10.3, then Ni=10). Finally, the rounded value is assigned to the number of short videos to be allocated, Ni.
[0048] Next, we obtain the overall popularity value of all subcategories under the main category YBi. It should be noted that the overall popularity value of a subcategory is calculated by weighted sum of the number of views and interactions of all short videos under the corresponding subcategory at the current time. The number of views is the number of times all short videos under the subcategory are played (without considering the playback time), and the number of interactions is the sum of the number of likes, comments, shares, and favorites of all short videos under the subcategory.
[0049] Based on the above method, calculate the overall popularity value of all subcategories under the major category tag YBi, and sort all subcategories under the major category tag YBi in descending order of overall popularity value. Starting from the first subcategory tag in the sorting results, select the short video with the highest individual popularity value under the corresponding subcategory tag in turn, until Ni short videos are selected. If Ni is greater than the total number of all subcategories under the major category tag YBi, then re-traverse all subcategories and select the short video with the second highest individual popularity value under the corresponding subcategory tag.
[0050] It should be noted that the purpose of this step is to prioritize the most popular videos in the most popular subcategories under the major categories, which can maximize the retention and interaction probability of new users during the interest exploration stage, because popular content usually has a wider appeal. Finally, the selected Ni videos are arranged in the order of selection and recorded as the secondary recommended short video sequence associated with the major category tag YBi, and labeled as Hi (representing a sequence).
[0051] By repeating the above steps, the secondary recommendation short video sequences for each major category tag can be determined and arranged in the order of the major category tag sequence YB1, YB2, ..., YBj, denoted as H1, H2, ..., Hj.
[0052] Based on the determined j-sequences of secondary recommended short videos H1, H2, ..., Hj, all secondary recommended short videos in the j-sequences are extracted and arranged in a completely random manner. Recommendations are then made to new users based on the randomized results until all short videos in the j-sequences H1, H2, ..., Hj have been recommended. During this process, secondary viewing data for each short video viewed by new user X is recorded. This secondary viewing data includes the actual playback duration, total short video duration, likes, comments, shares, and favorites.
[0053] Example 5 This embodiment, based on embodiment 4, further discloses a method for constructing a concentration index of interest distribution for new users, performing a solidification verification, and balancing and regulating the solidified interest distribution. Specifically, it includes the following: Obtain the secondary viewing data of all short videos in the secondary recommended short video sequence H1, H2, ..., Hj for the new user X determined in Example 4; And based on the secondary viewing data of each recommended short video, the interest index of new user X for each short video is calculated; Next, obtain the interest index of all short videos under any major category tag YBi, and average the interest index of all short videos, and record the average result as the interest index Qi' to be updated.
[0054] Repeat this step to determine the interest indices to be updated associated with each of the j major category labels, and arrange them in the order of the major category label sequence to form the interest index sequence to be updated Q1',Q2',...,Qj'.
[0055] Next, based on the interest index sequence Q1', Q2', ..., Qj' to be updated, the concentration index of the new user's interest distribution is evaluated, and a solidification validation is performed on the new user X based on the concentration index of the interest distribution, as follows: First, the average value AVG_Q' of the j interest indices to be updated in the interest index sequence Q1',Q2',...,Qj' is calculated by using AVG_Q'=(Q1'+Q2'+...+Qj') / j; Then, the absolute deviation Di between any interest index Qi' to be updated and the average value AVG_Q' is calculated using Di=|Qi'-AVG_Q'|. The mean absolute deviation (MAD) is then calculated using MAD=(D1+D2+...+Dj) / j, and the concentration index of interest distribution associated with new user X, ZB=MAD / AVG_Q', is calculated using ZB. MAD is a robust indicator that measures the dispersion of all interest indices. The larger the MAD, the more concentrated and uneven the interest distribution is, indicating that some interest indices are much higher than the average and some are much lower than the average. The Interest Distribution Concentration Index (ZB) is a normalized index that eliminates the influence of the overall interaction level of new user X. It purely reflects the relative concentration of interest distribution. A high ZB value indicates that interests are highly concentrated, and new user X has a strong interest in only a few major categories. A low ZB value indicates that interests are relatively evenly distributed, and new user X has some interest in multiple major categories.
[0056] Next, obtain the operator's preset interest fixation threshold θ, where θ > 0; If ZB≥θ, it is determined that the interest distribution of new user X has become solidified, with interests excessively concentrated on a few points, and there is a risk of falling into an information cocoon, requiring balancing and regulation. Conversely, if the interest distribution remains dispersed and no intervention is needed, the existing interest differences should be reinforced. If the interest distribution of new user X has not become fixed and no balancing adjustment is needed, short videos should be recommended to new user X based on the short video recommendation ratios K1, K2, ..., Kj.
[0057] Finally, if balancing adjustments are required, the short video recommendation ratios for each major category tag are recalculated based on the determined interest index sequence Q1',Q2',...,Qj' to be updated, and the calculated short video recommendation ratios are used as the final ratio parameters. Short videos are then recommended to new user X based on these final ratio parameters.
[0058] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.
[0059] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.
[0060] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.
Claims
1. A short video recommendation method for sparse scenarios, characterized in that, The method includes: Step 1: Acquire new users based on the short video platform, perform recommendation operations on the pre-built benchmark short video recommendation queue to the new users, record the viewing data of each benchmark short video from the time the new user starts viewing the benchmark short video recommendation queue to the time the viewing is completed, evaluate the new user's interest index for each benchmark short video, and use it as the interest index of the major category tag associated with each benchmark short video. Step 2: Calculate the short video recommendation ratio associated with each major category tag based on the interest index among the major category tags; Based on the short video recommendation ratio, the short videos under all subcategories of major categories are filtered, a secondary recommendation sequence of short videos associated with each major category is constructed, and the recommendation operation is performed to new users, and secondary viewing data is recorded. Step 3: Based on the secondary browsing data, reassess the new users' interest index on major category tags, construct a concentration index of the new users' interest distribution, perform solidification verification, determine whether to carry out balancing adjustment based on the solidification results, and perform recommendation operations.
2. The method according to claim 1, characterized in that, In step one, the specific method for acquiring new users based on the short video platform and performing recommendation operations on the pre-built baseline short video recommendation queue for the new users is as follows: Acquire new users who register on the short video platform, and denote them as X; Take a pre-built benchmark short video recommendation queue, where any benchmark short video corresponds to a major category of tags; Take the total number of pre-defined category tags j in the short video platform; Get any major category label YBi, where i is the counting index, with a value ranging from 1 to j; Get the number of likes, comments, shares, and favorites for all short videos under the major category tag YBi; Obtain the preset weights for calculating likes, comments, shares, and favorites; The selection index associated with each short video under the category tag YBi is calculated by weighted summation. The short video with the highest selection index is selected as the benchmark short video associated with the category tag YBi, and is denoted as Vi. Similarly, determine the benchmark short videos associated with each of the j major category tags, and randomly arrange the j major category tags, denoted as the major category tag sequence YB1, YB2, ..., YBj; A baseline short video recommendation queue V1, V2, ..., Vj is constructed based on the category label sequence, where YBi corresponds to Vi; Recommendation operations are performed on new user X based on the baseline short video recommendation queues V1, V2, ..., Vj.
3. The method according to claim 2, characterized in that, In step one, the specific method for assessing new users' interest in each benchmark short video is as follows: S31. Record the time span from when a new user X starts browsing the baseline short video recommendation queue to when they finish browsing, as the first monitoring period; S32. Obtain the viewing data of any benchmark short video Vi viewed by new user X within the monitoring period, including actual playback time PTi, total short video duration ATi, like flag Li, comment flag Ci, forward flag Si, and favorite flag Fi. If new user X performs a like operation on benchmark short video Vi, then Li=1; otherwise, Li=0. The same applies to Ci, Si, and Fi. S33. Calculate the completion rate Ri and effective dwell ratio ERi of new user X for the baseline short video Vi, where Ri = PTi / TTi. If Ri≥1, then ERi=1, otherwise ERi=Ri; S34. Obtain the preset playback behavior weight α1 and interaction behavior base weight α2 of the operator, where α1+α2=1; S35. Calculate the interest component P_Scorei=ERi based on playback behavior; Calculate the interest component U_Scorei based on interactive behavior: U_Scorei=(Li×w1+Ci×w2+Si×w3+Fi×w4) / (w1+w2+w3+w4), where w1, w2, w3, and w4 are the preset weights for likes, comments, shares, and favorites, respectively. S36. Calculate the interest index Qi of new user X on the benchmark short video Vi, which is Qi = α1*P_Scorei + α2*U_Scorei. Use the interest index Qi as the interest index of new user X on the category tag YBi associated with the benchmark short video Vi. S37. Similarly, determine the interest index of new user X for each benchmark short video and the interest index of the major category tags associated with each benchmark short video, and form the interest index sequence Q1, Q2, ..., Qj.
4. The method according to claim 3, characterized in that, In step two, the specific method for calculating the short video recommendation ratio associated with each major category tag based on the interest index between the major category tags is as follows: Calculate the sum of the j interest indices in the interest index sequence Q1, Q2, ..., Qj, denoted as the total interest index Tot_Q = Q1 + Q2 + ... + Qj; Calculate the initial recommendation ratio Ei = Qi / Total_Q for the major category label YBi; Obtain the preset minimum recommendation ratio threshold ε, where 0 < ε < 1 / j; If Ei≥ε, let the undetermined proportion of the major category label YBi be DEi=Ei; Conversely, let the undetermined proportion DEi = ε; Similarly, by determining the undetermined proportions of each major category label, we obtain DE1, DE2, ..., DEj; Calculate the total undetermined proportion Tot_DE = DE1 + DE2 + ... + DEj; Calculate the short video recommendation ratio Ki=DEi / Tot_DE for the major category tag YBi; Similarly, by determining the short video recommendation ratio K1 for each major category tag, we obtain K1, K2, ..., Kj.
5. The method according to claim 4, characterized in that, In step two, the specific method for constructing the secondary recommendation short video sequences associated with each of the major category tags is as follows: Take the preset total number of secondary recommended short videos, denoted as M; Based on the short video recommendation ratio of each major category tag, the number of short videos to be assigned in the secondary recommendation for each major category tag is calculated. Among them, the number of short videos to be assigned for the major category tag YBi is Ni=M*Ki. Ni is rounded down and the rounded value is assigned to the number of short videos to be assigned Ni. Obtain the overall popularity value of all subcategories under the major category tag YBi. The overall popularity value is calculated based on the weighted average sum of the number of views and interactions of all short videos under the corresponding subcategory tag at the current time. Play count is the number of times all short videos under the subcategory tag have been played, and interaction count is the sum of the number of likes, comments, shares, and favorites for all short videos under the subcategory tag. Sort all subcategories under the main category YBi in descending order of overall popularity value. Starting from the first subcategory, select the short videos with the highest individual popularity value under each subcategory until Ni short videos are selected. The individual popularity value is calculated based on the average weighted sum of the views and interactions of any single short video. Arrange the Ni videos in the order of selection into a secondary recommendation short video sequence associated with the major category tag YBi, denoted as Hi; Similarly, determine the sequence of short videos for secondary recommendations for each major category tag, denoted as H1, H2, ..., Hj.
6. The method according to claim 5, characterized in that, Step two further includes performing a recommendation operation on the determined secondary recommendation short video sequence to the new user X. The recommendation method is fully random recommendation, until all short videos in the j secondary recommendation short video sequences H1, H2, ..., Hj have been recommended. The secondary viewing data of the new user X for each short video is recorded, including the actual playback time, the total duration of the short video, likes, comments, shares, and favorites.
7. The method according to claim 6, characterized in that, In step three, the specific method for reassessing the new user's interest index in the major category tags based on the secondary viewing data is as follows: Obtain the secondary viewing data of new user X for all short videos in the secondary recommended short video sequence H1, H2, ..., Hj, and determine the interest index of all short videos based on the secondary viewing data of each short video in the manner described in steps S32 to S36. The interest index of all short videos under any major category tag YBi is averaged, and the result is recorded as the interest index Qi' to be updated. Similarly, determine the interest indices to be updated for each major category tag, forming the interest index sequence Q1', Q2', ..., Qj'.
8. The method according to claim 7, characterized in that, In step three, the specific method for constructing the concentration index of new users' interest distribution and performing solidification verification is as follows: Calculate the average value of the interest index sequence Q1', Q2', ..., Qj' to be updated: AVG_Q' = (Q1' + Q2' + ... + Qj') / j; Calculate the absolute deviation of each interest index to be updated from the average value AVG_Q', where the absolute deviation of any interest index Qi' to be updated is Di=|Qi'-AVG_Q'|. The mean absolute deviation (MAD) is calculated using MAD=(D1+D2+...+Dj) / j, and the concentration index of the interest distribution of new user X is calculated using ZB=MAD / AVG_Q'. Obtain the operator's preset interest fixation threshold θ, where θ > 0; If ZB≥θ, it is determined that the interest distribution of the new user X has become fixed and balance adjustment needs to be performed; Conversely, if the distribution of interests of new user X has not become fixed and no balancing adjustment is needed, short videos are recommended to new user X based on the short video recommendation ratios K1, K2, ..., Kj as the final ratio parameters.
9. The method according to claim 8, characterized in that, In step three, the specific method for performing balance adjustment and executing the recommended operation is as follows: Based on the interest index sequence Q1',Q2',...,Qj' to be updated, the short video recommendation ratio of each major category tag is recalculated, and the calculated short video recommendation ratio is used as the final ratio parameter. Based on the final ratio parameter, short videos are recommended to new user X.
10. A short video recommendation system for sparse scenarios, characterized in that, The system includes: The new user initialization module acquires new users based on the short video platform, performs recommendation operations on the pre-built benchmark short video recommendation queue to the new users, records the viewing data of each benchmark short video from the time the new user starts viewing the benchmark short video recommendation queue to the time when the viewing is completed, evaluates the new user's interest index for each benchmark short video, and uses it as the interest index of the major category tag associated with each benchmark short video. The tag association filtering module calculates the proportion of short videos recommended by each major category of tags based on the interest index between the major categories of tags; Based on the short video recommendation ratio, the short videos under all subcategories of major categories are filtered, a secondary recommendation sequence of short videos associated with each major category is constructed, and the recommendation operation is performed to new users, and secondary viewing data is recorded. The interest update and control module, based on secondary browsing data, reassesses the interest index of new users on major category tags, constructs a concentration index of new users' interest distribution, performs solidification verification, determines whether to perform balance control based on the solidification results, and performs recommendation operations.