A short video automatic recommendation method and system

Abstract

The present invention discloses a method and system for automatic recommendation of short videos, the method comprising: parsing user activity logs and preprocessing, generating user event information and video event information respectively; extracting user features according to user event information, using a neural network deep model based on feature intersection to generate user portraits and preferences; dynamically calculating video scores according to video event information, dynamically grouping video data according to video scores and generating multiple popular data sources; loading background configuration operation strategies, recalling qualified videos from different data sources according to user portraits and preferences, and outputting them to the front end for display. The present invention learns user preferences according to user logs, generates multiple video data sources according to a grouping algorithm, and recalls qualified videos from different data sources through background configuration operation strategies to achieve automatic operation of the machine.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence video recommendation, and in particular discloses a short video machine automatic recommendation method and system. Background Art

[0002] With the popularity of terminal devices such as mobile phones and TVs, mobile phones and TVs are increasingly replacing computers. People can easily browse videos on the Internet through mobile phones and TVs. When a user group browses videos, the user group's behavior data will be recorded for video recommendations.

[0003] In today's intelligent era, for short video companies, there is huge commercial value behind the huge data traffic generated by the user group. How to effectively use the traffic to make accurate recommendations is the core of the recommendation system. Traditional recommendation methods inevitably cause traffic waste, and require a lot of manual operations to change the recommended content and layout. It is impossible to explore and utilize user interests and hobbies and recommend appropriate content to users in a personalized way. Summary of the invention

[0004] In view of this, the present invention proposes a short video machine automatic recommendation method and system, which are used to solve the problem that short videos cannot be automatically recommended.

[0005] In a first aspect, the present invention discloses a method for automatically recommending short videos, the method comprising:

[0006] Parse and preprocess user activity logs to generate user event information and video event information respectively;

[0007] Extract user features based on user event information, and use a neural network deep model based on feature intersection to generate user profiles and preferences;

[0008] Dynamically calculate video scores based on video event information, dynamically group video data based on video scores, and generate multiple popular data sources;

[0009] Load the backend configuration operation strategy, recall qualified videos from different data sources based on user portraits and preferences, and output them to the front end for display.

[0010] Preferably, the parsing of user activity logs and preprocessing specifically includes:

[0011] Obtain user logs through the log parsing program, aggregate the click, play, and exposure information of a video within a time window, and generate different focused event streams.

[0012] Preferably, extracting user features according to user event information and using a neural network deep model based on feature intersection to generate user portraits and preferences specifically includes:

[0013] Extracting user dimension features and video dimension features; the user dimension features include but are not limited to the user's region and viewing time period features, and the video dimension features include but are not limited to the tags, channels, hosts, and viewing time features of the videos watched by the user;

[0014] The dual-tower model is used to cross-feature the two dimensions of user and video to obtain the user's portrait and preference.

[0015] Preferably, the specific method of dynamically calculating the video score according to the video CTR is:

[0016] Take the video exposure, click and play data to calculate the video CTR;

[0017] The video score S is dynamically calculated based on the video CTR, the video creation time createDay, and the ranking time rankDay. The calculation rules are:

[0018] S=(T-createDay)*weight_1+(T-rankDay)*weight_2+CTR

[0019] Among them, T is the recommendation period, weight_1 and weight_2 are weight parameters set according to specific business, and the weight will decrease as time increases.

[0020] Preferably, the dynamically grouping the video data according to the video scores and generating the hot data source specifically comprises:

[0021] Arrange the video scores from high to low, and assign different groups according to the proportion of the video scores in the global top; the groups include new video group, first-level hot pool, second-level hot pool, and top hot pool. Different groups have different exposure ratios;

[0022] Different groups are further subdivided according to the video channel, anchor or tag;

[0023] The grouped data sources are stored in accessible storage media and stored separately according to different time scales to form real-time hot data sources.

[0024] Preferably, the background configuration operation strategy includes configuring how many floors there are under a TAB page, the number of videos required for each floor, and adding limiting conditions, which include the number of floors, anchor diversity and channel diversity.

[0025] Preferably, the loading background configuration operation strategy, based on the user's profile and preferences, recalls qualified videos from different data sources, specifically including:

[0026] Translate the configuration operation policy into recall rules and filtering rules through parsing programs;

[0027] Under the specified recall rules, the user's profile and preferences are used to recall videos that meet the user's preferences from the data source; the data source includes offline data sources, popular data sources, and recently updated data sources, which are recalled from top to bottom;

[0028] Reorder the recalled data according to the video scores and output the data that meets the requirements according to the filtering rules;

[0029] Finally, assemble all floors and output JSON data and return it to the caller.

[0030] In a second aspect, the present invention discloses a short video automatic recommendation system, the system comprising:

[0031] Log parsing module: used to parse and pre-process user activity logs to generate user event information and video event information respectively;

[0032] Portrait generation module: used to extract user features based on user event information and use a neural network deep model based on feature intersection to generate user portraits and preferences;

[0033] Dynamic grouping module: dynamically calculates video scores based on video event information, dynamically groups video data based on video scores, and generates multiple popular data sources;

[0034] Data recall module: loads the background configuration operation strategy, recalls qualified videos from different data sources based on user portraits and preferences, and outputs them to the front-end for display.

[0035] In a third aspect of the present invention, an electronic device is disclosed, comprising: at least one processor, at least one memory, a communication interface and a bus;

[0036] Wherein, the processor, memory, and communication interface communicate with each other via the bus;

[0037] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the method described in the first aspect of the present invention.

[0038] According to a fourth aspect of the present invention, a computer-readable storage medium is disclosed, wherein the computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to implement the method described in the first aspect of the present invention.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1) Extract user features based on user logs, use a deep neural network model based on feature crossover to generate user portraits and preferences, deeply mine user preferences, dynamically calculate video scores based on video event information, dynamically group video data based on video scores and generate multiple popular data sources, while achieving personalized recommendations for users, combined with a video dynamic grouping algorithm to achieve a balance between utilization and exploration, and improve user stickiness;

[0041] 2) The backend configures the operation strategy to generate data recall rules and filtering rules. Based on the user's portrait and preferences, qualified videos are recalled from different data sources, assembled on all floors, and the machine is operated automatically, with automatic typesetting and duplication elimination, eliminating the tedious manual operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 This is a schematic diagram of the process flow of the short video automatic recommendation method of the present invention;

[0044] Figure 2 Schematic diagram of the structure of the short video automatic recommendation system of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with 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.

[0046] See also Figure 1 , a flow chart of the automatic short video recommendation method of the present invention; the automatic short video recommendation method proposed by the present invention comprises the following steps:

[0047] S1, parse and pre-process the user activity log to generate user event information and video event information respectively;

[0048] Set the server log address to be obtained, monitor the server log in real time by installing a log parsing program, and when the log starts to generate new entries, obtain its data and send it to the log parsing program.

[0049] The logs obtained by the log parsing program are used to extract video information and remove other information impurities to collect the user's video viewing history and browsing behavior. The click, play, and exposure information of a video are aggregated within a time window to generate different focused event streams. For example: the user information time stream, the exposure, click, and play information stream of a certain user on a certain floor under a certain TAB of a certain channel, aggregate their viewing history according to the user, connect their context information, generate user event streams, and finally generate user event information. Similarly, video event information is generated based on the click, play, and exposure information of the video.

[0050] S2. Extract user features based on user event information and use a neural network deep model based on feature intersection to generate user profiles and preferences;

[0051] Extracting user dimension features and video dimension features from the user event information; the user dimension features include features such as the user's region and viewing time period, and the video dimension features include features such as the tag, channel, anchor, and viewing time of the video watched by the user;

[0052] The dual-tower model is used to perform feature cross-fertilization on the two dimensions of user and video to obtain the user's portrait and preferences, including channel, anchor and tag preferences.

[0053] S3. Dynamically calculate the video score according to the video event information, dynamically group the video data according to the video score and generate multiple hot data sources; this step generates the hot data source through a grouping algorithm. The grouping algorithm is an algorithm for distinguishing the quality of videos, and classifies the videos from good to bad. The present invention distinguishes the quality of videos by estimating the video score.

[0054] Get the user's activity log in the server;

[0055] Extract the video information and remove other information impurities;

[0056] Aggregate video exposure, click, and play data according to the time window, and calculate the video CTR (Click Through Rate) based on the video exposure, click, and play data;

[0057] The data is then grouped and processed according to multiple dimensions such as video creation time, CTR, and ranking retention time. Specifically, the video score S is dynamically calculated based on the video CTR, video creation time createDay, and ranking time rankDay. The calculation rule is:

[0058] S=(T-createDay)*weight_1+(T-rankDay)*weight_2+CTR

[0059] Among them, T is the recommendation period, weight_1 and weight_2 are weight parameters set according to specific business, and the weight will decrease as time increases.

[0060] The way to calculate video ratings is dynamic, and the mobility of videos in different groups is changed based on user interaction behavior. Arrange video ratings from high to low, and group them according to the proportion of video ratings in the global top. Groups generally include new video groups, secondary hot pools, primary hot pools, and top hot pools. Different groups have different exposure ratios, which are defined by themselves. The higher the ratio, the more times the video is recommended. For example, after estimating the ratings of each video, first arrange the video ratings from high to low. If the estimated video rating of a video ranks high and belongs to the top 10% (top 10%) of the global ranking, it can be divided into the top hot pool. If the estimated video rating of another video belongs to the top 10% to 30%, it will be divided into the first hot pool, and so on. Different groups are further subdivided according to the video channel, anchor or label.

[0061] The grouped data sources are stored in accessible storage media and stored separately according to different time scales such as hours and days to form real-time hot data sources.

[0062] The present invention extracts user features according to user logs, uses a neural network deep model based on feature intersection to generate user portraits and preferences, deeply mines user preferences, dynamically calculates video scores according to video event information, dynamically groups video data according to video scores and generates multiple popular data sources, thereby achieving personalized recommendations for users and combining video dynamic grouping algorithms to achieve a balance between utilization and exploration, thereby improving user stickiness.

[0063] S4. Load the backend configuration operation strategy, recall qualified videos from different data sources based on user portraits and preferences, and output them to the front end for display.

[0064] By configuring the operation strategy, the machine automation operation effect can be achieved. For example, configure how many floors there are under a TAB page, the number of videos required for each floor, and add restrictions, such as channel restrictions, anchor restrictions, or diversity restrictions. Then, the configured operation strategy is translated into recall rules and filtering rules through the parser. Under the specified recall rules, the user's portrait and preferences are used to recall channels, anchors, tags and other videos that meet the user's preferences from the data source. The data sources include offline data sources, popular data sources, and recently updated data sources. They are recalled in the order of offline data sources, popular data sources, and recently updated data sources. The data in each data source is sorted by CTR, and each data source is recalled from top to bottom. The recalled data is reordered according to the video score, and the data that meets the requirements is output according to the filtering rules; finally, all floors are assembled, the data is filled into the structure of the specified floor, and the JSON data is output and returned to the caller to recommend videos that meet the user's preferences.

[0065] The present invention loads the background configuration operation strategy and parses the recall rules and filtering rules of the generated data, combines the user's portrait and preferences, recalls qualified videos from different data sources, assembles all floors, realizes the automated operation of the machine, realizes automatic typesetting and duplication elimination, and eliminates the tedious manual operation.

[0066] Corresponding to the method embodiment, the present invention also discloses a short video automatic recommendation system, please refer to Figure 2 , the system comprising:

[0067] Log parsing module 10: used to parse and pre-process user activity logs to generate user event information and video event information respectively;

[0068] Portrait generation module 20: used to extract user features based on user event information and generate user portraits and preferences using a neural network deep model based on feature intersection;

[0069] Dynamic grouping module 30: dynamically calculates video scores according to video event information, dynamically groups video data according to the video scores and generates multiple popular data sources;

[0070] Data recall module 40: loads the background configuration operation strategy, recalls qualified videos from different data sources based on the user's profile and preferences, and outputs them to the front-end for display.

[0071] The above system embodiments correspond one-to-one to the method embodiments. For a brief description of the system embodiments, please refer to the method embodiments.

[0072] The present invention collects the user's video viewing history and browsing behavior, performs preprocessing operations on these data, and inputs the processed data into the training model to predict the user's portrait and user preferences; collects video exposure, click and play data and inputs them into the grouping algorithm to estimate the video score, and groups all videos according to the video score. Different groups have different exposure granularity. The higher the video score, the greater the exposure possibility, and the weight increases or decreases with the exposure, click and play data; finally, through the recommendation background unified configuration framework, the recall conditions of different sections and floors are specified, and according to the user's portrait and preference, the qualified videos are recalled from different grouping data sources, and output to the front-end display. The present invention realizes personalized recommendations for users by deeply mining user preferences. On the one hand, it achieves a balance between utilization and exploration and improves user stickiness. On the other hand, the machine realizes the automatic operation of the machine, which can automatically layout and remove duplicates, and reduce the cumbersome manual operation. It can be applied to the automatic recommendation of terminal devices such as mobile phones, TVs, computers or tablets.

[0073] The present invention also discloses an electronic device, comprising: at least one processor, at least one memory, a communication interface and a bus; wherein the processor, memory, and communication interface communicate with each other through the bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement all or part of the steps of the aforementioned short video automatic recommendation method of the present invention.

[0074] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, wherein the computer instructions enable the computer to implement all or part of the steps of the short video automatic recommendation method described in the embodiment of the present invention. The storage medium includes: a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk, and other media that can store program codes.

[0075] The system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A short video automatic recommendation method, characterized in that: The method comprises: Parse and pre-process user activity logs to generate user event information and video event information respectively; Extract user features based on user event information, and use a neural network deep model based on feature intersection to generate user profiles and preferences; Dynamically calculate video scores based on video event information, dynamically group video data based on video scores, and generate multiple popular data sources; Load the backend configuration operation strategy, recall qualified videos from different data sources based on user profiles and preferences, and output them to the front end for display; The parsing of user activity logs and preprocessing specifically include: Obtain user logs through the log parsing program, aggregate the click, play, and exposure information of a video within a time window, and generate different focused event streams; The method of extracting user features according to user event information and using a neural network deep model based on feature intersection to generate user portraits and preferences specifically includes: Extract user dimension features and video dimension features from the user event information; the user dimension features include the user's region and viewing time period features, and the video dimension features include the tag, channel, anchor, and viewing time features of the video watched by the user; Use the dual-tower model to cross-feature the user and video dimensions to obtain the user's portrait and preferences; The specific method of dynamically calculating the video score based on the video CTR is: Take the video exposure, click and play data to calculate the video CTR; The video score S is dynamically calculated based on the video CTR, the video creation time createDay, and the ranking time rankDay. The calculation rules are: S=(T-createDay)*weight_1+(T-rankDay)*weight_2+CTR Among them, T is the recommendation period, weight_1 and weight_2 are weight parameters set according to specific business, and the weight will decrease as time increases; The dynamic grouping of video data according to video scores and generation of hot data sources are specifically as follows: Arrange the video scores from high to low, and assign different groups according to the proportion of the video scores in the global top; the groups include new video group, first-level hot pool, second-level hot pool, and top hot pool. Different groups have different exposure ratios; Different groups are further subdivided according to the video channel, anchor or tag; Store the grouped data sources in accessible storage media, store them separately according to different time scales, and form real-time hot data sources; The backend configuration operation strategy includes configuring how many floors there are under a TAB page, the number of videos required for each floor, and adding limiting conditions, which include the number of floors, anchor diversity, and channel diversity; The loading backend configuration operation strategy recalls qualified videos from different data sources based on user profiles and preferences, specifically including: Translate the configuration operation policy into recall rules and filtering rules through parsing programs; Under the specified recall rules, the user's profile and preferences are used to recall videos that meet the user's preferences from the data source; the data source includes offline data sources, popular data sources, and recently updated data sources, which are recalled from top to bottom; Reorder the recalled data according to the video scores and output the data that meets the requirements according to the filtering rules; Finally, assemble all floors and output JSON data to return to the caller.

2. A short video automatic recommendation system using the method of claim 1, characterized in that: The system comprises: Log parsing module: used to parse and pre-process user activity logs to generate user event information and video event information respectively; Portrait generation module: used to extract user features based on user event information and use a neural network deep model based on feature intersection to generate user portraits and preferences; Dynamic grouping module: dynamically calculates video scores based on video event information, dynamically groups video data based on video scores, and generates multiple popular data sources; Data recall module: loads the background configuration operation strategy, recalls qualified videos from different data sources based on user portraits and preferences, and outputs them to the front-end for display.

3. An electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus; Wherein, the processor, memory, and communication interface communicate with each other via the bus; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the method of claim 1. 4 . A computer-readable storage medium storing computer instructions, wherein the computer instructions enable a computer to implement the method of claim 1 .

Images