CN-122019832-A - Short video recommendation method and system for sparse scene

CN122019832ACN 122019832 ACN122019832 ACN 122019832ACN-122019832-A

Abstract

The invention discloses a short video recommending method and a short video recommending system for sparse scenes, which relate to the technical field of video recommending and evaluate interest indexes of new users on all standard short videos and related major labels by acquiring new users, recommending pre-constructed standard short video queues to the new users and recording reading data; and then, re-evaluating the interest index of the new user on the large-class labels according to the secondary reading data, constructing interest distribution concentration index for solidification verification, determining whether balance regulation is carried out, and finally carrying out short video recommendation to the new user.

Inventors

LIU DONGHAI
SUN JIANSHAN

Assignees

合肥工业大学

Dates

Publication Date: 20260512
Application Date: 20260113

Claims (10)

1. The short video recommendation method for sparse scenes is characterized by comprising the following steps of: Step one, acquiring a new user based on a short video platform, executing recommendation operation on a pre-constructed reference short video recommendation queue to the new user, recording reading data of each reference short video in a period from when the new user starts to read the reference short video recommendation queue to when the new user finishes reading, evaluating the interest index of the new user for each reference short video, and taking the interest index as the interest index of a large label associated with each reference short video; step two, calculating short video recommendation proportion associated with each large class of labels based on interest indexes among the large class of labels; Based on short video recommendation proportion, screening short videos of all subclasses of labels under all the subclasses of labels, constructing secondary recommendation short video sequences respectively associated with the subclasses of labels, performing recommendation operation on new users, and recording secondary reading data; And thirdly, evaluating interest indexes of the new user on the large-class labels again based on the secondary reading data, constructing interest distribution concentration indexes of the new user, executing solidification verification, determining whether balance regulation is carried out or not based on solidification results, and executing recommendation operation.
2. The method according to claim 1, wherein in the first step, the specific way of obtaining the new user based on the short video platform and performing the recommendation operation on the pre-constructed reference short video recommendation queue to the new user is as follows: Acquiring a new user registered with the short video platform and recording X; taking a pre-constructed reference short video recommendation queue, wherein any reference short video corresponds to a large label; taking the total number j of the large class labels calibrated in advance in the short video platform; any large label YBi is obtained, wherein i is a counting index, and the value range is 1 to j; obtaining praise amount, comment amount, forwarding amount and collection amount of all short videos under the large-class label YBi; Acquiring a preset praise calculation weight, a comment calculation weight, a forwarding calculation weight and a collection calculation weight; calculating the preference index associated with each short video under the large-class label YBi by adopting weighted summation, selecting the short video with the highest preference index as the reference short video associated with the large-class label YBi, and marking as Vi; similarly, determining reference short videos associated with j major class labels respectively, and performing random arrangement on the j major class labels, and recording the j major class labels as major class label sequences YB1, YB2, YBj; constructing a reference short video recommendation queue V1, V2, vj based on the large class tag sequences, wherein YBi corresponds to Vi; based on the reference short video recommendation queues V1, V2, vj performs a recommendation operation to the new user X.
3. The method according to claim 2, wherein in the first step, the specific way of evaluating the interest index of the new user for each reference short video is: s31, recording a time span from a start of reading a reference short video recommendation queue to completion of reading of a new user X, and taking the time span as a first monitoring period; S32, acquiring reading data of a new user X for reading any reference short video Vi in a monitoring period, wherein the reading data comprises an actual playing time PTi, a short video total time ATi, a praise mark Li, a comment mark Ci, a forwarding mark Si and a collection mark Fi, if the new user X carries out praise operation on the reference short video Vi, li=1, otherwise, li=0, ci, si and Fi are the same; S33, calculating the playing completion rate Ri and the effective stay ratio ERi of the new user X to the reference short video Vi, wherein Ri=PTi/TTi, If Ri is greater than or equal to 1, eri=1, otherwise eri=ri; s34, acquiring a play behavior weight alpha 1 and an interaction behavior basic weight alpha 2 preset by an operator, wherein alpha 1 +alpha 2 = 1; S35, calculating an interest component p_ Scorei =eri based on playing behavior; calculating an interest component U_ Scorei = (Lixw1+Cixw2+Sixw3+fi xw4)/(w1+w2+w3+w4) based on interaction behavior, wherein w1, w2, w3 and w4 are respectively preset praise weight, comment weight, forwarding weight and collection weight; s36, calculating an interest index Qi=α1×P_Score+α2×U_Score of the new user X on the reference short video Vi, and taking the interest index Qi as an interest index of the new user X on the major label YBi associated with the reference short video Vi; And S37, determining interest indexes of the new user X for all the reference short videos and interest indexes of the major labels associated with all the reference short videos by the same method, and forming interest index sequences Q1, Q2.
4. The method according to claim 3, wherein in the second step, the specific manner of calculating the short video recommendation ratio associated with each major category label based on the interest index between each major category label is: Calculating a sequence of interest indices Q1, Q2..j the sum of the j interest indices in Qj is noted as total interest index tot_q=q1+q2.+ Qj; Calculating a preliminary recommended proportion ei=qi/total_q of the large class label YBi; acquiring a preset minimum recommended proportion threshold epsilon, wherein 0< epsilon <1/j; if Ei is greater than or equal to epsilon, enabling the undetermined proportion DEi=Ei of the large class label YBi; Otherwise, let dei=ε to be defined; likewise, determining the undetermined proportion of each major class of tags to obtain DE1, DE2, DEj; Calculating the total to-be-determined ratio tot_de=de1+de2+ & DEj; Calculating short video recommendation ratio ki=dei/tot_de of the large class label YBi; and similarly, determining a short video recommendation proportion K1 of each major type of tag to obtain K1, K2.
5. The method according to claim 4, wherein in the second step, the specific manner of constructing the secondary recommended short video sequence associated with each major label is as follows: taking the total number of preset secondary recommended short videos, and marking the total number as M; Calculating the number of short videos to be distributed in secondary recommendation of each large-class label based on the short video recommendation proportion of each large-class label, wherein the number Ni=M×Ki of the short videos to be distributed of the large-class label YBi, performing downward rounding on Ni, and assigning the rounded value to the number Ni of the short videos to be distributed; Acquiring comprehensive heat values of all subclasses of tags under the large class of tags YBi, wherein the comprehensive heat values are calculated based on the average weighted summation of the play amount and the interaction amount of all short videos under the corresponding subclasses of tags when the current time is taken as the end time; The play quantity is the play times of all short videos under the subclass labels, and the interaction quantity is the sum of the praise, comment, forwarding and collection times of all short videos under the subclass labels; Sequencing all subclasses of labels under the large class of labels YBi in descending order according to the comprehensive heat value, starting from the first subclass label, sequentially selecting the short video with the highest individual heat value until Ni short videos are selected, wherein the individual heat value is calculated based on the play amount and the average weighted sum of the interaction amount of any single short video; arranging the Ni videos into a secondary recommendation short video sequence associated with the major label YBi according to the selection sequence, and marking the secondary recommendation short video sequence as Hi; similarly, a secondary recommended short video sequence for each major class of tags was determined, and was designated H1, H2, hj in that order.
6. The method according to claim 5, further comprising performing a recommendation operation on the determined secondary recommended short video sequences to the new user X, wherein the recommendation mode is a full random recommendation until all short videos in the j secondary recommended short video sequences H1, H2, hj are recommended, and recording secondary viewing data of the new user X for viewing each short video, including an actual playing duration, a short video total duration, a praise flag, a comment flag, a forwarding flag, and a collection flag.
7. The method according to claim 6, wherein in the third step, based on the secondary viewing data, the new user's interest index in the large class label is rated again in the following specific manner: acquiring secondary reading data of all short videos in secondary recommended short video sequences H1, H2, hj of a new user X, and determining interest indexes of all short videos by combining the secondary reading data of all short videos based on the modes in the steps S32 to S36; averaging the interest indexes of all short videos under any one large class label YBi, and marking the average result as an interest index Qi' to be updated; similarly, interest indexes to be updated of the tags of each major class are determined, and interest index sequences Q1', Q2'.
8. The method according to claim 7, wherein in the third step, the interest distribution concentration index of the new user is constructed, and the curing verification is performed in the following specific manner: calculating an average avg_q ' = (q1 ' +q2' +, +qj ')/j of the index sequence of interest Q1', Q2', qj ' to be updated; calculating absolute deviation of each interest index to be updated and the average value AVG_Q ', wherein the absolute deviation Di= |Qi' -AVG_Q '| of any interest index Qi' to be updated; calculating an average absolute deviation MAD by using mad= (d1+d2+ & gt.+ Dj)/j, and calculating an interest distribution concentration index ZB of the new user X by zb=mad/avg_q'; acquiring an interest curing threshold value theta, theta >0 preset by an operator; if ZB is more than or equal to theta, judging that the interest distribution of the new user X is solidified, and executing balance regulation; And otherwise, judging that the interest distribution of the new user X is not solidified, and recommending short videos to the new user X based on the final proportion parameters by taking the short video recommendation proportions K1, K2.
9. The method of claim 8, wherein in the third step, the balance adjustment is performed in a specific manner of performing the recommended operation: Based on the sequence of interest indices Q1', Q2', recalculate the short video recommendation ratio for each broad category of tags, and recommending the short video to the new user X based on the final proportion parameter by taking the calculated short video recommendation proportion as the final proportion parameter.
10. A short video recommendation system for sparse scenes, the system comprising: The new user initialization module is used for acquiring a new user based on the short video platform, executing recommendation operation on the pre-built standard short video recommendation queue to the new user, recording reading data of each standard short video in a period from when the new user starts to read the standard short video recommendation queue to when the new user finishes reading, evaluating the interest index of the new user for each standard short video, and taking the interest index as the interest index of a large label associated with each standard short video; The tag association screening module calculates short video recommendation proportion associated with each large class of tags based on interest indexes among the large classes of tags; Based on short video recommendation proportion, screening short videos of all subclasses of labels under all the subclasses of labels, constructing secondary recommendation short video sequences respectively associated with the subclasses of labels, performing recommendation operation on new users, and recording secondary reading data; And the interest updating regulation and control module is used for evaluating interest indexes of the new user on the large-class labels again based on the secondary reading data, constructing interest distribution concentration indexes of the new user, executing solidification verification, determining whether balance regulation and control are carried out or not based on a solidification result, and executing recommendation operation.

Description

Short video recommendation method and system for sparse scene Technical Field The invention belongs to the technical field of video recommendation, and particularly relates to a short video recommendation method and system for sparse scenes. Background With the rapid development of the mobile internet, short video platforms emerge and spread rapidly like spring bamboo shoots after rain, and people tend to acquire information, entertain and relax and socialize through short videos in leisure time. The conventional short video recommendation method is widely applied collaborative filtering algorithm or an emerging model based on graph neural network or deep learning, is subject to remarkable limitation in processing such scenes, and particularly mainly depends on sufficient historical user behavior data to construct user portraits or carry out interest association, however, for new users, the historical interaction information is completely lost, so that the conventional collaborative filtering algorithm is difficult to find similar users to effectively recommend, the recommendation based on the content is trapped into a dilemma due to the fact that the initial preference of the users cannot be obtained, namely, a typical cold start problem, even after the users have some preliminary behaviors, effective feedback data generated by single users are very rare and highly random due to fragmentation and rapid switching characteristics of short video consumption, so that a user and project interaction matrix are extremely sparse, a traditional model is difficult to accurately mine stable and reliable long-tail interest preference from the user interaction matrix, the recommendation quality is large in initial fluctuation, and the current method is easy to cause the interest exploration process of the new users in practice, and is sunk into an information cocoon house too early, for example, a plurality of recommendation algorithms capture the user to stay in the category, the user is difficult to quickly capture the initial interest of the user, the interaction is difficult to quickly expand the interest of the user interaction, the interaction is easy to expand the user interaction with a certain user, the interaction is easy to expand the interaction index, the situation is easy to be influenced by the user has a short-term interest, and the user has a poor possibility of the situation is easy to develop, and the user has a poor possibility of easily-stage is in the situation, and has a serious potential problem. In order to solve the problems, the invention provides a short video recommendation method and a short video recommendation system for sparse scenes. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a short video recommendation method and a short video recommendation system for sparse scenes, which solve the problems of insufficient interest exploration and premature recommendation homogenization caused by excessively relying on sparse data or pursuing short-term indexes in the cold start and data sparse stages of new users in the prior art. The aim of the invention can be achieved by the following technical scheme: a short video recommendation method facing sparse scenes comprises the following steps: Step one, acquiring a new user based on a short video platform, executing recommendation operation on a pre-constructed reference short video recommendation queue to the new user, recording reading data of each reference short video in a period from when the new user starts to read the reference short video recommendation queue to when the new user finishes reading, evaluating the interest index of the new user for each reference short video, and taking the interest index as the interest index of a large label associated with each reference short video; step two, calculating short video recommendation proportion associated with each large class of labels based on interest indexes among the large class of labels; Based on short video recommendation proportion, screening short videos of all subclasses of labels under all the subclasses of labels, constructing secondary recommendation short video sequences respectively associated with the subclasses of labels, performing recommendation operation on new users, and recording secondary reading data; And thirdly, evaluating interest indexes of the new user on the large-class labels again based on the secondary reading data, constructing interest distribution concentration indexes of the new user, executing solidification verification, determining whether balance regulation is carried out or not based on solidification results, and executing recommendation operation. In the first step, a new user is acquired based on a short video platform, and the specific way of executing recommendation operation to the new user by using the pre-constructed reference short video recommendation queue is as follows: Acquiring a new user regis