User behavior prediction method and device

Abstract

The invention provides a user behavior prediction method and device, relates to the technical field of behavior prediction, and aims to accurately predict knowledge search requirements generated in a short video browsing process of a user, firstly, a short video currently browsed by the user is analyzed in two dimensions of meaning and picture so as to obtain a first semantic analysis result; analyzing the information of the comment area to obtain a second semantic analysis result so as to obtain a first to-be-recommended entry set, then determining basic knowledge information and expanded search information based on a specific operation of a user on a specific comment so as to obtain a second to-be-recommended entry set, and finally, obtaining a second to-be-recommended entry set. And updating the first vocabulary entry blacklist based on the operation information of the user on the target recommended vocabulary entry. According to the method, the to-be-recommended entries are determined from multiple dimensions, the to-be-recommended entries are further expanded according to the specific operation information of the user, and the to-be-recommended entries are filtered based on the browsing portrait of the user and the entry blacklist, so that the recommended entries can better meet the actual requirements of the user.

Description

Technical Field

[0001] The present invention belongs to the technical field of behavior prediction, and in particular, relates to a method and device for user behavior prediction. Background Art

[0002] When users use short video apps, they usually receive recommended videos that match their preferences, which are obtained based on big data analysis. The convenience of short video browsing can also be used to provide college students with an independent learning platform, that is, to display the knowledge points that students are interested in in a fragmented way, thereby improving students' learning efficiency.

[0003] However, currently short video apps can usually only recommend short videos based on the user's browsing history, which is too single in terms of recommendation dimensions and cannot meet user needs. In addition, in response to the user's real-time knowledge acquisition needs when browsing short videos, the existing technology can only recommend to users in text form, which is too single in terms of recommendation format and cannot meet the user's actual needs. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for predicting user behavior to solve the technical problem in the prior art that the dimension of knowledge acquisition demand prediction is too single.

[0005] A user behavior prediction method, characterized in that the method comprises:

[0006] S1: Determine a first user portrait according to a short video browsing record of the first user within a preset time period;

[0007] The short video browsing record includes the first browsing times of each short video by the first user;

[0008] S2: Determine D second short videos according to the first number of views, and determine a first supplementary user portrait based on the D second short videos;

[0009] S3: obtaining a third short video currently browsed by the first user, determining a plurality of first terms to be recommended according to the third short video and the first user portrait, and forming a first term set to be recommended from the plurality of first terms to be recommended;

[0010] The S3 comprises the following sub-steps:

[0011] S31: performing semantic expansion according to the third short video to obtain a plurality of first candidate terms;

[0012] S32: filtering the first candidate terms based on the first user portrait to obtain a plurality of second candidate terms;

[0013] S33: selecting a preset number of the second candidate terms as first terms to be recommended, and forming a first term set to be recommended from the plurality of the first terms to be recommended;

[0014] S4: filtering the first set of terms to be recommended by using the first supplementary user portrait to determine a plurality of second terms to be recommended and a plurality of first short videos to be recommended, and forming a second set of terms to be recommended by the plurality of second terms to be recommended;

[0015] S5: In response to the first operation issued by the first user, a first number of second terms to be recommended are selected from the second set of terms to be recommended and displayed on the first interface, and a second number of first short videos to be recommended are selected from the plurality of first short videos to be recommended and displayed on the first interface;

[0016] S6: In response to the second operation issued by the first user, the content displayed on the first interface is refreshed, and the first entry blacklist is updated.

[0017] Preferably, S1 comprises the following sub-steps:

[0018] S11: Acquire A first short videos browsed by the first user within a preset time period, wherein each of the first short videos corresponds one-to-one to a first tag;

[0019] S12: Performing a first frequency analysis on the A first short videos using the first tag to obtain B first short videos and B first viewing times;

[0020] S13: performing cluster analysis on the B first short videos according to the B first viewing times to obtain C first cluster centers;

[0021] Among them, A, B and C are all natural numbers greater than 0, and A>B>C.

[0022] Preferably, the S13 includes the following sub-steps:

[0023] S131: performing semantic analysis on the B first short videos one by one to obtain B first semantic analysis information;

[0024] S132: weighting the B pieces of first semantic analysis information based on the B first viewing times to obtain B pieces of first weight information corresponding to the B first short videos;

[0025] S133: performing cluster analysis on the B first short videos according to the B first weight information and the B first semantic analysis information to obtain C first cluster centers;

[0026] S134: Determine a first user portrait based on the C first cluster centers.

[0027] Preferably, S2 comprises the following sub-steps:

[0028] S21: Determine the first short video whose first number of views is greater than a preset value as a second short video, thereby obtaining D second short videos; wherein D is a natural number less than C;

[0029] S22: For each of the second short videos, obtain D first key browsing information of the first user;

[0030] S23: classifying the D first key browsing information into two categories: first text browsing information and first multimedia browsing information;

[0031] S24: inputting the first text browsing information and the first multimedia browsing information into a first supplementary portrait determination model and a second supplementary portrait determination model respectively, so as to obtain a first text supplementary portrait and a first multimedia supplementary portrait;

[0032] S25: For each piece of the first multimedia browsing information, establish a first mapping relationship between the first multimedia browsing information and the first theme information of the corresponding second short video;

[0033] S26: Construct the first supplementary user portrait according to the first text-based supplementary portrait, the first multimedia supplementary portrait and the first mapping relationships.

[0034] Preferably, S4 includes the following sub-steps:

[0035] S41: Acquire the first text-based supplementary portrait, the first multimedia supplementary portrait, and the plurality of first mapping relationships from the first supplementary user portrait;

[0036] S42: performing similarity matching between the first text-based supplementary portrait and the plurality of first terms to be recommended one by one, and taking the first terms to be recommended whose similarity meets preset requirements as second terms to be recommended;

[0037] S43: Filtering the plurality of second terms to be recommended by using the first term blacklist to obtain a set of second terms to be recommended;

[0038] S44: Extract second multimedia browsing information and second theme information from the third short video, and traverse multiple first mapping relationships according to the second multimedia browsing information and the second theme information. If a traversal result exists, determine the second short video corresponding to the traversal result as the first short video to be recommended.

[0039] Preferably, S6 comprises the following sub-steps:

[0040] S61: Refreshing the content displayed on the first interface in response to a click operation of the first user on the first control;

[0041] S62: updating the first term blacklist according to the replaced second term to be recommended and the first short video to be recommended.

[0042] The present application also proposes a user behavior prediction system for implementing the above-mentioned user behavior prediction method.

[0043] The present application proposes a method and device for predicting user behavior, which generates a first user portrait based on the user's short video browsing history, and generates a first supplementary user portrait based on the user's browsing behavior for short video soundtracks, texts, and comments. When the user generates a knowledge search demand while browsing short videos, the first user portrait and the first supplementary user portrait are used to generate recommended terms and recommended short videos for the user, and the term blacklist can be updated based on the user's feedback. Through the technical solution of the present application, search behavior recommendations can be made for users from multiple dimensions such as user historical preferences, short video semantics, texts, comments, and soundtracks, and the recommendation results are more in line with the actual needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0045] Figure 1 It is an execution flow chart of a user behavior prediction method of the present invention.

[0046] Figure 2 It is a schematic diagram of the first interface in the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the illustrative embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0049] A method and device for predicting user behavior of the present invention are described in detail below.

[0050] This embodiment proposes a user behavior prediction method for predicting the user's operation behavior in a short video APP. The method flow is as follows: Figure 1 As shown, the specific steps include:

[0051] S1: Determine a first user portrait based on the short video browsing record of the first user within a preset time period.

[0052] The first user refers to a user for whom user behavior prediction is to be performed.

[0053] In one embodiment, the user behavior prediction method is applied to a short video learning platform in colleges and universities, and student users can conduct autonomous learning by browsing short videos on the platform.

[0054] The S1 comprises the following sub-steps:

[0055] S11: Acquire A first short videos viewed by the first user within a preset time period, wherein each of the first short videos corresponds one-to-one to a first tag.

[0056] The preset time period refers to the time period when the user is relatively active on the short video platform. For example, if the first user is only active on the short video platform for one hour a day, the preset time period can be taken as the one hour when the first user is active on the short video platform. If the first user is active on the short video platform for more than six hours a day, the preset time period can be taken as the six hours when the first user is active on the short video platform.

[0057] S12: Perform a first frequency analysis on the A first short videos using the first tag to obtain B first short videos and B first viewing times.

[0058] For example, when a student user browses short videos on a short video platform, if he encounters a video that he does not understand, he may browse it repeatedly and may even play it back. Therefore, statistical analysis of the playback frequency of the first short video of the first user can be conducted to further analyze the browsing preferences of the first user.

[0059] S13: Performing cluster analysis on the B first short videos according to the B first viewing times to obtain C first cluster centers.

[0060] Among them, A, B and C are all natural numbers greater than 0, and A>B>C.

[0061] The cluster analysis is performed based on the semantic information of the B first short videos.

[0062] The S13 comprises the following sub-steps:

[0063] S131: Perform semantic analysis on the B first short videos one by one to obtain B first semantic analysis information.

[0064] S132: weighting the B pieces of first semantic analysis information based on the B first viewing times to obtain B pieces of first weight information corresponding to the B first short videos.

[0065] When the first viewing number is larger, it means that the first short video matches the first user's preference more closely, and therefore the weight is higher; when the first viewing number is smaller, it means that the first short video matches the first user's preference less closely, and therefore the weight is lower.

[0066] The specific weighting rule may assign 1 to the first short video with the largest number of first views, and normalize and weight other first short videos according to their corresponding first view counts.

[0067] S133: Perform cluster analysis on the B first short videos according to the B first weight information and the B first semantic analysis information to obtain C first cluster centers.

[0068] The C first clustering centers are semantic label sets obtained by semantically clustering the B first short videos.

[0069] S134: Determine a first user portrait based on the C first cluster centers.

[0070] Each of the first cluster centers includes a number of semantic tags, and the first user portrait can be determined by extracting the semantic tags whose occurrence frequencies meet preset requirements.

[0071] The first user portrait is used to characterize the browsing preference information of the first user within a preset time period.

[0072] S2: Determine D second short videos according to the first number of views, and determine a first supplementary user portrait based on the D second short videos.

[0073] The S2 comprises the following sub-steps:

[0074] S21: Determine the first short video whose first browsing times is greater than a preset value as the second short video, thereby obtaining D second short videos.

[0075] D is a natural number smaller than C.

[0076] Usually, when student users encounter a short video that they do not understand, they will watch it repeatedly multiple times. However, the type of knowledge represented by such short videos may be what the student users need to strengthen their learning.

[0077] Therefore, the first short video whose first viewing number is greater than a preset value can be selected as the second short video, so that D of the second short videos are taken as the key analysis objects to obtain the first supplementary user portrait.

[0078] S22: For each of the second short videos, obtain D first key browsing information of the first user.

[0079] In this step, the short video browsing behavior of the first user is monitored in real time, so as to obtain the first key browsing information generated by the first user when browsing the second short video.

[0080] Specifically, the first focused browsing information includes the soundtrack, text or comments of the second short video.

[0081] When the first key browsing information is the soundtrack, the monitoring behavior to be performed on the first user includes whether the first user has performed a secondary search for the soundtrack used by the second short video. The secondary search refers to continuing to search for other short videos using the soundtrack using the soundtrack as an entry.

[0082] S23: Classify the D first key browsing information into two categories: first text browsing information and first multimedia browsing information.

[0083] Generally, the text and comments of the short videos that users follow have similar semantics, so the text and comments can be classified as text. The soundtracks that users follow usually represent the user's preference for the soundtrack style, so the short video itself and the soundtrack can be classified as multimedia.

[0084] S24: Inputting the first text browsing information and the first multimedia browsing information into a first supplementary portrait determination model and a second supplementary portrait determination model respectively to obtain a first text supplementary portrait and a first multimedia supplementary portrait.

[0085] The first supplementary portrait determination model may be based on a Transformer model to semantically summarize the user's favorite texts and comments to obtain a number of favorite semantic tags, thereby forming the first text-based supplementary portrait.

[0086] The second supplementary portrait determination model is obtained based on convolutional neural network training. In the training samples, the input is a number of multimedia soundtracks, and the output is the user's multimedia preference label. The second supplementary portrait determination model can be trained based on the above training samples.

[0087] S25: For each of the first multimedia category browsing information, a first mapping relationship is established between the first multimedia category browsing information and the first theme information of the corresponding second short video.

[0088] For example, if the first user browses a second short video with a theme of "British and American Culture" and music 1 and the number of views exceeds a preset value, the music 1 can be used as the first multimedia browsing information, and "British and American Culture" and music 1 can be saved as the first mapping relationship.

[0089] S26: Construct the first supplementary user portrait according to the first text-based supplementary portrait, the first multimedia supplementary portrait and the first mapping relationships.

[0090] The first supplementary user portrait is a supplement to the first user portrait, and represents the first user's preference tags in two dimensions: text and multimedia. In addition, the soundtrack-short video combination that the first user particularly pays attention to is also saved in the first supplementary user portrait in the form of the first mapping relationship, so that when it is predicted that the first user wants to browse the short video again, the short video can be quickly recommended.

[0091] S3: Obtain a third short video currently viewed by the first user, determine a plurality of first terms to be recommended based on the third short video and the first user portrait, and form a first term set to be recommended from the plurality of first terms to be recommended.

[0092] The S3 comprises the following sub-steps:

[0093] S31: Perform semantic expansion based on the third short video to obtain a number of first candidate terms.

[0094] In this step, with respect to the third short video currently being browsed by the first user, semantic recognition and expansion are performed on the video content of the third short video to obtain a plurality of the first candidate terms.

[0095] For example, if the third short video is about "Pan Zhanle's swimming skills", some of the first candidate terms may be Pan Zhanle, Pan Zhanle's arm swing, Pan Zhanle's turn, etc.

[0096] S32: Filtering the first candidate terms based on the first user portrait to obtain a plurality of second candidate terms.

[0097] In this step, the first user portrait obtained in S1 is used to perform similarity matching with several first candidate terms one by one, and the first candidate terms whose similarities meet preset requirements are determined as second candidate terms, thereby obtaining several second candidate terms.

[0098] S33: Select a preset number of the second candidate terms as first terms to be recommended, and form a first term set to be recommended from the plurality of the first terms to be recommended.

[0099] The preset number can be set according to actual conditions, and can also be adaptively set according to actual needs of users.

[0100] S4: Filter the first set of terms to be recommended using the first supplementary user portrait to determine a plurality of second terms to be recommended and a plurality of first short videos to be recommended, and form a second set of terms to be recommended from the plurality of second terms to be recommended.

[0101] The S4 comprises the following sub-steps:

[0102] S41: Obtain the first text-based supplementary portrait, the first multimedia supplementary portrait, and the plurality of first mapping relationships from the first supplementary user portrait.

[0103] S42: performing similarity matching between the first text-based supplementary portrait and the plurality of first terms to be recommended one by one, and taking the first terms to be recommended whose similarities meet preset requirements as second terms to be recommended.

[0104] The similarity matching adopts a cosine similarity matching algorithm.

[0105] S43: Filter the plurality of the second terms to be recommended using the first term blacklist to obtain the second term set to be recommended.

[0106] The first term blacklist refers to a list of terms that are ignored or excluded by the first user. The terms in the first term blacklist have a low relevance to the first user's preferences.

[0107] S44: Extract second multimedia browsing information and second theme information from the third short video, and traverse multiple first mapping relationships according to the second multimedia browsing information and the second theme information. If a traversal result exists, determine the second short video corresponding to the traversal result as the first short video to be recommended.

[0108] S5: In response to the first operation issued by the first user, a first number of second terms to be recommended are selected from the second set of terms to be recommended and displayed on the first interface, and a second number of first short videos to be recommended are selected from the multiple first short videos to be recommended and displayed on the first interface.

[0109] The first operation refers to a return operation issued by the user during the process of browsing the short video, which usually indicates that the user has a search demand for a certain knowledge point during the process of browsing the short video. Therefore, it is necessary to use the first operation as a trigger condition to recommend entries and short videos to the user.

[0110] The first quantity and the second quantity can be set according to actual needs, and can be specifically set according to the display space of the first interface.

[0111] The first interface is Figure 2 As shown, the first interface is divided into an entry display area and a short video display area. The second entry to be recommended is displayed in the entry display area, and the first short video to be recommended is displayed in the short video display area.

[0112] S6: In response to the second operation issued by the first user, the content displayed on the first interface is refreshed, and the first entry blacklist is updated.

[0113] The S6 comprises the following sub-steps:

[0114] S61: In response to a click operation of the first user on a first control, refreshing content displayed on the first interface.

[0115] The first control refers to a "change batch" control, which is used to indicate that the content displayed on the first interface does not meet user needs, so the first user triggers the "change batch" control to re-display new content on the first interface.

[0116] The new content may preferably be an entry that has not been displayed in the second set of entries to be recommended and the first short video to be recommended that has not been displayed during the first display.

[0117] S62: updating the first term blacklist according to the replaced second term to be recommended and the first short video to be recommended.

[0118] When the first user chooses to replace the current content on the first interface, it means that the content on the first interface does not meet the user's needs, so the content on the first interface can be added to the first entry blacklist.

[0119] The present application also proposes a user behavior prediction system for executing the above-mentioned user behavior prediction method.

[0120] The present application proposes a method and device for predicting user behavior, which generates a first user portrait based on the user's short video browsing history, and generates a first supplementary user portrait based on the user's browsing behavior for short video soundtracks, texts, and comments. When the user generates a knowledge search demand while browsing short videos, the first user portrait and the first supplementary user portrait are used to generate recommended terms and recommended short videos for the user, and the term blacklist can be updated based on the user's feedback. Through the technical solution of the present application, search behavior recommendations can be made for users from multiple dimensions such as user historical preferences, short video semantics, texts, comments, and soundtracks, and the recommendation results are more in line with the actual needs of users.

[0121] The above description is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A user behavior prediction method, characterized in that: The method includes: S1: Determine a first user portrait according to a short video browsing record of the first user within a preset time period; The short video browsing record includes the first browsing times of each short video by the first user; S2: Determine D second short videos according to the first number of views, and determine a first supplementary user portrait based on the D second short videos; S3: obtaining a third short video currently browsed by the first user, determining a plurality of first terms to be recommended according to the third short video and the first user portrait, and forming a first term set to be recommended from the plurality of first terms to be recommended; The S3 comprises the following sub-steps: S31: performing semantic expansion according to the third short video to obtain a plurality of first candidate terms; S32: filtering the first candidate terms based on the first user portrait to obtain a plurality of second candidate terms; S33: selecting a preset number of the second candidate terms as first terms to be recommended, and forming a first term set to be recommended from the plurality of the first terms to be recommended; S4: filtering the first set of terms to be recommended by using the first supplementary user portrait to determine a plurality of second terms to be recommended and a plurality of first short videos to be recommended, and forming a second set of terms to be recommended by the plurality of second terms to be recommended; S5: In response to the first operation issued by the first user, a first number of second terms to be recommended are selected from the second set of terms to be recommended and displayed on the first interface, and a second number of first short videos to be recommended are selected from the plurality of first short videos to be recommended and displayed on the first interface; S6: In response to the second operation issued by the first user, the content displayed on the first interface is refreshed, and the first entry blacklist is updated.

2. A user behavior prediction method according to claim 1, characterized in that: The S1 comprises the following sub-steps: S11: Acquire A first short videos browsed by the first user within a preset time period, wherein each of the first short videos corresponds one-to-one to a first tag; S12: Performing a first frequency analysis on the A first short videos using the first tag to obtain B first short videos and B first viewing times; S13: performing cluster analysis on the B first short videos according to the B first viewing times to obtain C first cluster centers; Among them, A, B and C are all natural numbers greater than 0, and A>B>C.

3. A user behavior prediction method according to claim 2, characterized in that: The S13 comprises the following sub-steps: S131: performing semantic analysis on the B first short videos one by one to obtain B first semantic analysis information; S132: weighting the B pieces of first semantic analysis information based on the B first viewing times to obtain B pieces of first weight information corresponding to the B first short videos; S133: performing cluster analysis on the B first short videos according to the B first weight information and the B first semantic analysis information to obtain C first cluster centers; S134: Determine a first user portrait based on the C first cluster centers.

4. A user behavior prediction method according to claim 3, characterized in that: The S2 comprises the following sub-steps: S21: Determine the first short video whose first number of views is greater than a preset value as a second short video, thereby obtaining D second short videos; wherein D is a natural number less than C; S22: For each of the second short videos, obtain D first key browsing information of the first user; S23: classifying the D first key browsing information into two categories: first text browsing information and first multimedia browsing information; S24: inputting the first text browsing information and the first multimedia browsing information into a first supplementary portrait determination model and a second supplementary portrait determination model respectively, so as to obtain a first text supplementary portrait and a first multimedia supplementary portrait; S25: For each piece of the first multimedia browsing information, establish a first mapping relationship between the first multimedia browsing information and the first theme information of the corresponding second short video; S26: Construct the first supplementary user portrait according to the first text-based supplementary portrait, the first multimedia supplementary portrait and the first mapping relationships.

5. A user behavior prediction method according to claim 4, characterized in that: The S4 comprises the following sub-steps: S41: Acquire the first text-based supplementary portrait, the first multimedia supplementary portrait, and the plurality of first mapping relationships from the first supplementary user portrait; S42: performing similarity matching between the first text-based supplementary portrait and the plurality of first terms to be recommended one by one, and taking the first terms to be recommended whose similarity meets preset requirements as second terms to be recommended; S43: Filtering the plurality of second terms to be recommended by using the first term blacklist to obtain a set of second terms to be recommended; S44: Extract second multimedia browsing information and second theme information from the third short video, and traverse multiple first mapping relationships according to the second multimedia browsing information and the second theme information. If a traversal result exists, determine the second short video corresponding to the traversal result as the first short video to be recommended.

6. A user behavior prediction method according to claim 5, characterized in that: The S6 comprises the following sub-steps: S61: Refreshing the content displayed on the first interface in response to a click operation of the first user on the first control; S62: updating the first term blacklist according to the replaced second term to be recommended and the first short video to be recommended.

7. A user behavior prediction system, used to implement a user behavior prediction method in claims 1-6 above.

Images