Video recommendation method, apparatus, device, and computer program product

Abstract

The application relates to the technical field of artificial intelligence, and provides a video recommendation method, device, equipment and computer program product. The method determines a photographing user through obtained social network data, extracts communication data and behavior data from a personal terminal of the photographing user, constructs a family group of the photographing user based on the communication data, extracts behavior data of each family member from a public terminal of the family group, determines preference information of each family member based on the extracted behavior data, obtains position signaling data of each family member, determines a target member currently using the public terminal according to the position signaling data, determines a target video according to the preference information of the target member, and recommends the target video in the public terminal. The video recommendation method provided in the application can identify the preference information of a hidden user, identify a watching user through position signaling data, recommend a video meeting the preference of the watching user to the watching user, and realize accurate personalized recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a video recommendation method, apparatus, device, and computer program product. Background Technology

[0002] With the rapid development of the internet and artificial intelligence, user video preference recognition and personalized video recommendations based on user preferences have become key functions of multimedia platforms. For platforms providing content services, the crucial question is how to deliver accurate content. Currently, mainstream video recommendations are based on a user's viewing behavior under their logged-in account, combined with the user's personalized characteristics. However, for shared devices like televisions, different family members may exhibit drastically different viewing behaviors under the same account. Conventional recommendation methods cannot identify the preferences of different family members based on the same account, thus failing to provide accurate personalized recommendations. Summary of the Invention

[0003] This application provides a video recommendation method, apparatus, device, and computer program product to solve the technical problem that it is impossible to provide accurate personalized recommendation services to different family members in a shared terminal device in a family group.

[0004] In a first aspect, embodiments of this application provide a video recommendation method, including:

[0005] Acquire social network data from the communication network, and determine the user taking the photo based on the social network data;

[0006] Extract the user's communication data and first behavior data from the user's personal terminal, and construct the user's family group based on the communication data;

[0007] Based on the user taking the photo, determine the public terminal of the family group, and extract the second behavior data corresponding to the family group from the public terminal;

[0008] Based on the first behavioral data and the second behavioral data, the preference information of each family member in the family group is determined, and the location signaling data of each family member is obtained. Based on the location signaling data, the target member currently using the public terminal is determined.

[0009] The target video is determined based on the preference information of the target members and recommended in the public terminal.

[0010] In one embodiment, the step of constructing the family group of the user taking the photo based on the communication data includes:

[0011] Based on the communication data, construct a set of positive and negative samples that have a family relationship with the user who took the photo;

[0012] The positive and negative sample sets are input into a pre-trained logistic regression model to obtain the probability values ​​of the binary relation pairs corresponding to the users who take photos.

[0013] Based on the probability values ​​of the binary relation pairs, determine each family member who has a family relationship with the user taking the photo;

[0014] Obtain the basic information of the user taking the photo and each of the family members, and calculate the information difference between the basic information of the user taking the photo and the basic information of each of the family members.

[0015] The family relationship between each family member and the user taking the photo is determined based on the information difference, and the family group of the user taking the photo is constructed.

[0016] In one embodiment, the step of determining the preference information of each family member in the family group based on the first behavioral data and the second behavioral data includes:

[0017] Based on the first behavioral data, the preference information of the user taking the photo is determined; based on the second behavioral data, a preference set containing the preference information of each family member in the family group is determined.

[0018] The preference information in the preference set is filtered using the preference information of the user taking the photo, to determine the preference information of the hidden user of the public terminal in the family group;

[0019] The hidden user is identified based on the family relationship between each family member in the family group and the user taking the photo, so as to determine the preference information of each family member.

[0020] In one embodiment, the step of determining the target video based on the target member's preference information and recommending it on the public terminal includes:

[0021] Based on the preference information of the target members, preference propagation is performed in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target members;

[0022] Based on the second behavioral data, an attention mechanism is used to score the videos in the video recommendation candidate set to determine the target video with the highest score and recommend it in the public terminal.

[0023] In one embodiment, the step of determining the video recommendation candidate set corresponding to the target member by performing preference propagation in a preset knowledge graph based on the target member's preference information includes:

[0024] Collect the historical behavior data corresponding to the target member in the first behavior data and the second behavior data;

[0025] Construct an interaction matrix between the target member and the video content based on the historical behavior data;

[0026] The interaction matrix is ​​used to propagate the preference information of the target member in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target member.

[0027] In one embodiment, the step of constructing the interaction matrix between the target member and the video content based on the historical behavior data includes:

[0028] Obtain preset behavioral factor weight coefficients, and calculate the preference score list of the target member based on the behavioral factor weight coefficients. The preference score list includes user identifiers and video identifiers.

[0029] A user set is constructed based on the user identifier, and a content set is constructed based on the video identifier;

[0030] Based on the preference score information of the target members in the preference score list, the user set and the content set are divided into positive and negative samples to obtain the interaction matrix between the target members and the video content.

[0031] In one embodiment, before the step of determining the video recommendation candidate set corresponding to the target member by performing preference propagation in a preset video knowledge graph based on the target member's preference information, the method further includes:

[0032] Raw video data from different applications is acquired and cleaned to obtain basic video data;

[0033] Entity information is extracted from the basic video data, a unified content library is generated based on the entity information, and knowledge reasoning and knowledge discovery are performed based on the unified content library to construct an entity relationship graph of the entity information.

[0034] A video knowledge graph is constructed based on the entity relationship graph, and the video knowledge graph includes multiple entity relationship tuples.

[0035] Secondly, embodiments of this application provide a video recommendation device, comprising:

[0036] The user filtering module is used to acquire social network data from the communication network and determine the user taking the photo based on the social network data.

[0037] A group building module is used to extract the photography user's communication data and first behavior data from the photography user's personal terminal, and to build the photography user's family group based on the communication data.

[0038] The data extraction module is used to determine the public terminal of the family group based on the user taking the photo, and extract the second behavior data corresponding to the family group from the public terminal;

[0039] The preference recognition module is used to determine the preference information of each family member in the family group based on the first behavior data and the second behavior data, and to obtain the location signaling data of each family member, and to determine the target member currently using the public terminal based on the location signaling data;

[0040] The video recommendation module is used to determine target videos based on the preference information of the target members and recommend them on the public terminal.

[0041] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the video recommendation method described in the first aspect.

[0042] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the video recommendation method described in the first aspect.

[0043] This application provides a video recommendation method, apparatus, device, and computer program product. By constructing a family group and utilizing the video viewing behavior data of family members' personal terminals, the video viewing behavior data of the family group's public terminals is filtered and screened to identify the preference information of hidden users on the public terminals. Based on this, combined with the location of family members when IPTV is played, the corresponding viewing users are identified, thereby enabling precise personalized video recommendations for different users. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is one of the flowcharts illustrating the video recommendation method provided in the embodiments of this application;

[0046] Figure 2This is one of the schematic diagrams illustrating the hidden user preference identification method of the video recommendation method provided in this application embodiment;

[0047] Figure 3 This is the second schematic diagram of the hidden user preference identification method provided in the embodiments of this application;

[0048] Figure 4 This is one of the structural schematic diagrams of the video recommendation device provided in the embodiments of this application;

[0049] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] It's important to note that for home-based terminal products like IPTV (Internet Protocol Television), which are typically shared by all family members, users watching videos on their own mobile phones can have their preferences identified and personalized recommendations made based on data from their individual devices, since each user and mobile phone account usually corresponds to a single individual. However, for home-based products like IPTV, the personalized recommendations shift from individual users to the entire family group. Users who don't use mobile phones to watch videos (such as the elderly and children) become hidden IPTV users.

[0052] Figure 1 This is one of the flowcharts illustrating the video recommendation method provided in this application. (Refer to...) Figure 1 The video recommendation method provided in this application embodiment may include:

[0053] Step 100: Obtain social network data from the communication network, and determine the user taking the photo based on the social network data;

[0054] In this embodiment, when providing personalized recommendation services to users, the first step is to construct family social networks for different users, identify family members in each user's family group, and form micro-social circles. Specifically, anonymized social network data from the communication network is acquired, and the user taking photos is identified based on the acquired social network data. This user is used as a reference when constructing the family social network, and family members with family relationships with the user are identified based on their social network data. When identifying users taking photos, preset filtering rules are mainly used to filter out users taking photos, thereby identifying all users taking photos across the network. The filtering rules mainly include: unregistered users; users with abnormal status (such as users whose status is suspended, canceled, or inactive); SIM (Subscriber Identity Module) card users and M2M (Machine to Machine, Internet of Things) users, such as those with 0 calls and more than 0 data usage in the past 3 months; test users, such as those with 0 visits, more than 0 SMS messages, and 0 data usage in the past 3 months; users who have been roaming outside the province for a long time, such as those whose calls in the past 3 months have all been roaming calls; and users with low-spending numbers, such as those whose average income per user in the past 6 months is less than or equal to 5 yuan. By using the above filtering rules, the social network data of all users across the network is filtered to ultimately determine the users who took photos across the entire network.

[0055] Step 200: Extract the photography user's communication data and first behavior data from the photography user's personal terminal, and construct the photography user's family group based on the communication data;

[0056] After identifying the user taking the photo, communication data and video viewing behavior data are extracted from the user's personal terminal, such as a mobile phone or tablet. These personal terminals are unique to the user; the following explanation uses a mobile phone as an example. Feature extraction is performed on the extracted communication data. For example, characteristics such as high call frequency, strong stability, and high overlap of location base stations at night, on weekends, and holidays are used to extract the user's communication features. Based on these features, other users who may be related to the user are identified, thus constructing the user's family group. This family group includes not only individual family members but also the family relationships between each family member and the user. The user's video viewing behavior data on their mobile phone is extracted from video applications installed and used on the user's phone. This application can be one or multiple applications.

[0057] Step 300: Determine the public terminal of the family group based on the user taking the photo, and extract the second behavior data corresponding to the family group from the public terminal;

[0058] Furthermore, based on the public terminal used by the user taking the photo, the family group is identified, such as IPTV. The following explanation uses IPTV as an example. When determining the IPTV used by the user taking the photo, it could be an IPTV with the same broadband data connection as the user's user ID, or an IPTV with the same broadband data connection as a family member of the user taking the photo; no specific limitation is made here. Different users' IPTVs are distinguished by device identifiers, which are then associated or bound to the user's user ID. After determining the public terminal corresponding to each user taking the photo, the video viewing behavior data of each member in the family group is extracted from this public terminal for subsequent preference identification and analysis.

[0059] Step 400: Based on the first behavioral data and the second behavioral data, determine the preference information of each family member in the family group, obtain the location signaling data of each family member, and determine the target member currently using the public terminal based on the location signaling data;

[0060] Based on video viewing behavior data extracted from the mobile phones of users who take photos, the viewing behavior data of family members extracted from public terminals is filtered and screened. The behavior of users who take photos is mapped to that of public terminals, thereby determining the preferences of hidden users on public terminals. It is known that the video viewing behavior of the same user on a mobile phone is consistent with their video viewing behavior on IPTV. Therefore, user video viewing behavior data and preferences on mobile phones can be fully or partially mapped to video viewing behavior on IPTV. By comparing and analyzing the behavior data on mobile phones and IPTV, the behavioral preferences of hidden IPTV users are filtered and screened, thus determining the preference information of each family member in the family group. When there are multiple hidden users, behavior is differentiated based on the user's own characteristics, thereby identifying the preferences of each family member. For example, hidden users in a family are generally mostly elderly people and children, whose individual differences are significant, and their video preferences also vary considerably. The elderly often prefer traditional performing arts programs, while children often prefer cartoons and educational programs.

[0061] After determining the preferences of each family member, location signaling data is obtained from their mobile phones to ascertain their location information. Based on the IPTV installation address, communication address, and mobile phone location signaling data, the target member currently using IPTV is identified. It is known that the installation location of IPTV is generally fixed, but different family members have different work and rest schedules. Members who need to work or study generally use it outside of work hours, while members who do not need to work or go to work may use it more during work hours. By obtaining the location signaling data of each family member and determining whether their location matches the IPTV installation location, it is possible to determine whether each member is currently using IPTV, thus identifying the target member currently using IPTV. This target member can be one or more.

[0062] Step 500: Determine the target video based on the target member's preference information and recommend it in the public terminal.

[0063] After identifying the target user of IPTV, videos that can be recommended are determined based on that user's preference information and then recommended on IPTV. If there is only one target user, video content matching their preferences is determined based on that user's preference information, providing a precise and personalized recommendation service. If there are multiple target users, their preference information is merged to determine the target video to be recommended and then recommended on IPTV. Furthermore, the fusion of preference information from different users can be achieved by first determining a candidate set of recommended videos for each user, and then using an attention mechanism to score the videos in each candidate set based on each user's personal characteristics, frequency of IPTV usage, currently running applications, and frequency of use of those applications, thereby determining the target video with the highest score and recommending it on IPTV. It is understandable that there can be one or more target videos, which are then displayed on IPTV for the user to choose from.

[0064] In this embodiment, by constructing a family group and using the video viewing behavior data of family members' personal terminals, the video viewing behavior data of the family group's public terminals is filtered and screened to identify the preference information of hidden users on the public terminals. Based on this, combined with the location of family members when IPTV is played, the corresponding viewing users are identified, so that accurate personalized video recommendations can be made for different users.

[0065] In one embodiment, step 200 may further include:

[0066] Step 201: Construct a set of positive and negative samples that have a family relationship with the user taking the photo, based on the communication data;

[0067] When constructing family groups, a positive and negative sample set is built based on the communication data of the users who take photos, identifying those with family relationships to the users. Specifically, firstly, information is associated between users in the family network and the users who take photos to generate a call data table for the users who take photos. This call data table includes pairs of family relationships for the users who take photos. Then, positive and negative samples are divided based on a preset sample partitioning rule. An example of such a rule is as follows:

[0068] 1. Filter positive samples: Remove family relationship pairs with multiple accounts for one person by user identifier, control the number of family network members of the same user taking photos to 2-3 people, and ensure that the family network IDs are consistent, and set the filtered family relationship to 1;

[0069] 2. Negative Sample Filtering: First, remove user-identified relationship pairs within the same network. For example, if user A is the user taking the photo, and user B is a member of user A's family network, a corresponding relationship pair A-B is generated. Similarly, if user B is the user taking the photo, user A can be identified as a member of user B's family network, leading to the generation of a corresponding relationship pair. This results in duplicate relationship pairs, which are then removed. After removing duplicate relationship pairs, negative samples are filtered. Negative samples are users with inconsistent family network IDs; one user has a family network while the other does not. The family relationships of the filtered users are set to 0.

[0070] A sample dataset was constructed based on the selected positive and negative samples. Specifically, users with family relationships exhibited characteristics such as high call frequency and strong stability, and high overlap of base stations at their locations outside of working hours. Therefore, the feature data shown in Table 1 was extracted to construct a dataset containing positive and negative samples. A positive and negative sample dataset of users with family relationships to those who took photos was constructed based on the feature data in Table 1.

[0071] Step 202: Input the positive and negative sample sets into the pre-trained logistic regression model to obtain the probability values ​​of the binary relation pairs corresponding to the users taking photos;

[0072] By using the pre-trained logistic regression model, the constructed positive and negative sample datasets are input into the model to obtain the probability values ​​of the binary family relationship pairs of the users who took the photos, as shown in Table 2.

[0073]

[0074] Table 1

[0075]

[0076] Table 2

[0077] The specific identification process is as follows:

[0078] First, the conditional probability that the users taking the photos have a family relationship is shown in Formula 1 below:

[0079] P(y i =1|x i ) = p i And P(y) i =0|x i )=1-p i (1)

[0080] but:

[0081] Where x k Let b represent the k-th feature variable in the model. k For x k The corresponding regression coefficients. Formula 2 above can be transformed to obtain:

[0082]

[0083] To solve for the regression coefficients, we need to maximize their likelihood function L(θ) over n samples:

[0084]

[0085]

[0086] Taking the partial derivative of the log-likelihood function of Formula 5 above yields k+1 nonlinear likelihood equations. These equations are then solved iteratively using the Newton-Raphson algorithm. Once the algorithm converges, the k regression coefficients of the model can be obtained, thus calculating the conditional probability that the user and the user taking the photo have a family relationship in each relationship pair.

[0087] Step 203: Determine each family member who has a family relationship with the user taking the photo based on the probability value of the binary relationship pair;

[0088] Based on the probability values ​​of binary family relationship pairs, an algorithm is used to expand and form small family social circles, ultimately generating a family group. Specifically, the fast-unfolding (community detection) algorithm is used to characterize the tightness of the community through modularity, as shown in Formula 6 below:

[0089]

[0090] Where m represents the sum of all weights in the network, A ij k represents the probability value of the binary family relationship pair, i.e., the weight between node i and node j. i =∑ j Ai,j represents the sum of the weights of the edges connected to vertex i, and c i δ(c) represents the family circle to which vertex i is assigned. i c jThis function is used to determine whether vertices i and j are in the same family circle. If they are, it returns 1; otherwise, it returns 0.

[0091] Step 204: Obtain the basic information of the user taking the photo and each of the family members, and calculate the information difference between the basic information of the user taking the photo and the basic information of each of the family members;

[0092] Step 205: Determine the family relationship between each family member and the user taking the photo based on the information difference, and construct the family group of the user taking the photo.

[0093] The fast-unfolding algorithm infers that the key person among family members is generally the head of household. By combining basic information about each member and the information differences between different members, the family relationships between other family members and the key person are determined, thus establishing corresponding family groups. The rules for determining the family relationships between other family members and the key person are as follows:

[0094] 1. Spouse: A user of this website, whose gender is different from that of the key family member, and whose absolute age difference is less than 10 years;

[0095] 2. Parental generation members: Users of this website who are at least 20 years older than the key family member;

[0096] 3. Children / Children: Users of this website who are at least 20 years younger than the key family member;

[0097] 4. Peer members: Users of this website who are of the same gender as the key family member and whose absolute age difference is within 6 years, or who are of different genders but whose first 6 digits of their ID card are the same and whose absolute age difference is less than 6 years.

[0098] As can be seen, the data in the judgment rules is only used for illustrative purposes and can be adjusted according to actual needs. The family relationships between family members can be identified through the above rules, and family groups as shown in Table 3 below can be constructed.

[0099] In this embodiment, by constructing family groups and identifying the family relationships among family members in the family groups, it is beneficial to identify hidden users in public terminals, thereby discovering the preference information of hidden users, which is conducive to providing accurate personalized recommendation services for hidden users.

[0100]

[0101] Table 3

[0102] In one embodiment, step 300 may further include:

[0103] Step 301: Based on the first behavioral data, determine the preference information of the user taking the photo; based on the second behavioral data, determine a preference set containing the preference information of each family member in the family group.

[0104] When identifying the preference information of hidden members of a public IPTV terminal in a family group, the preference information of the user taking the photo is first determined based on the first behavioral data extracted from the user's mobile phone. Then, the preference set containing the preference information of each family member is determined based on the second behavioral data extracted from IPTV.

[0105] Step 302: Filter the preference information in the preference set using the preference information of the user taking the photo, and determine the preference information of the hidden user of the public terminal in the family group;

[0106] Mapping the preferences of users who take photos to a preference set, filtering the preferences in the set to remove those of identical users, leaves only the preferences of the hidden users. (Refer to...) Figure 3 , Figure 3 This diagram illustrates the process of identifying hidden user preferences. Users A and B are family members who watch videos using their mobile phones, while user X is a hidden user who does not use a mobile phone to watch videos. Based on behavioral data extracted from the phones of users A and B, their preference information is determined and then mapped to a preference set for filtering and elimination. The remaining information represents user X's preferences. Further, when mapping the preferences of users who take photos to the preference set, the basic information of the user taking photos, their video preferences, and location signaling data during video viewing are used to determine whether the behavioral data in IPTV corresponds to the viewing behavior of the user taking photos. For example, if one or more behavioral data points match the preferences of a user taking photos, but the location signaling data of the user taking photos is not at the IPTV's location during the displayed viewing time period, then the preference information corresponding to that behavioral data does not belong to the user taking photos; otherwise, it is eliminated. By matching the user's basic information, video viewing preferences, and location signaling data with IPTV family viewing behavior and eliminating preferences that do not match the hidden user's behavior, a more accurate and in-depth analysis of the hidden user's video content preferences can be achieved.

[0107] Step 303: Identify the hidden user based on the family relationship between each family member in the family group and the user taking the photo, so as to determine the preference information of each family member.

[0108] It is understandable that there may be one or more hidden users. Based on the family relationship between each family member and the user taking the photo, the identity information of the hidden users is identified. When there are multiple hidden users, the basic information and location signaling data of the hidden users are matched to eliminate behavioral data that does not match the identity of each hidden user, so as to make the preference identification of each family member more accurate, thereby determining the video content preferences of each family member in the family group.

[0109] For example, in a family group of three, user A, the "dad," has the following video viewing behavior data as shown in Table 4. In Table 4, when playing "cartoon" videos on IPTV, the system combines information such as video playback time and location to determine whether the dad is at home. By removing the "cartoon" tag that does not match the dad's personal identity preference, the system can more accurately analyze the user's video usage preferences.

[0110]

[0111] Table 4

[0112] Furthermore, the behavioral data of user B's "mother" watching videos is shown in Table 5 below:

[0113]

[0114] Table 5

[0115] The behavioral data of user X's "child" watching videos is shown in Table 6 below:

[0116]

[0117] Table 6

[0118] Based on this, it is possible to construct, such as Figure 3 The user profile shown determines the video content preferences of each family member and makes precise personalized recommendations based on their video preferences, location signaling data, and other information. For example, if a user turns on IPTV at 6 PM, the location signaling data of users A and B's mobile phones shows that they are not at home. Combined with user X's usual preference to watch cartoons at 6 PM, precise personalized cartoon recommendations are made for user X.

[0119] In this embodiment, user behavior data on personal terminals is used to filter and screen their behavior data on public terminals. By combining user basic information, video viewing preferences, location signaling data, etc., the family viewing behavior on public terminals is matched, and preferences that do not conform to the user's personal behavior are eliminated. This allows for a more accurate and in-depth analysis of hidden user video content preferences, which is beneficial for providing accurate and personalized video recommendation services to different family members through public terminals.

[0120] In one embodiment, step 400 may further include:

[0121] Step 401: Based on the preference information of the target member, perform preference propagation in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target member;

[0122] Step 402: Based on the second behavior data, use an attention mechanism to score the videos in each video recommendation candidate set to determine the target video with the highest score and recommend it in the public terminal.

[0123] After determining the preferences of each family member, to provide accurate and personalized video recommendation services to different members, the location signaling data of each family member is first obtained. Based on the location signaling data, target members who may currently be using IPTV are identified, and recommendations are made based on the preferences of these target members to improve the accuracy of video recommendations. When there are multiple target members, the preference information of different members is combined for recommendations.

[0124] Specifically, based on the target member's preference information, preference propagation is performed in a pre-defined video knowledge graph to determine the video recommendation candidate set corresponding to the target member. When there are multiple target members, each member corresponds to a separate video recommendation candidate set. Based on behavioral data obtained from IPTV, the frequency of IPTV usage and the number of videos watched by the target member are determined. Combined with the member's own basic information, an attention mechanism is used to score the attention of each video in the video recommendation candidate set, and the target video with the highest score is determined and recommended to the target member in IPTV.

[0125] Furthermore, step 401 may also include:

[0126] Step 4011: Collect the historical behavior data corresponding to the target member in the first behavior data and the second behavior data;

[0127] Step 4012: Construct an interaction matrix between the target member and the video content based on the historical behavior data;

[0128] Step 4013: Use the interaction matrix to propagate the preference information of the target member in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target member.

[0129] When determining the video recommendation candidate set based on the target members' preference information, the behavior data of the same member is first summarized based on the behavior data extracted from mobile phones and IPTV to form historical behavior data including video log data watched by the target members.

[0130] Based on the target member's historical behavior data, an interaction matrix is ​​constructed between the member and the video content they watch, as shown in Formula 7 below:

[0131] Y = {y uv |u∈U,v∈V} (7)

[0132] Where U = {u1, u2, ..., ui} is the user set, V = {v1, v2, ..., vj} is the video content set, i represents the number of users, and j represents the number of video content items. uv This represents the user's interactive history of watching content, with values ​​of 0 and 1. A value of 0 indicates that the user has not watched the video content, while a value of 1 indicates that the user has watched the video content. The established interaction matrix is ​​used as a seed to propagate preferences within a pre-defined video knowledge graph, thereby obtaining a candidate set of video recommendations for the target member.

[0133] The preset video knowledge graph includes entity relation tuple data of video content. In this embodiment, taking triple data as an example, the constructed interaction matrix is ​​used as the seed of interest for the target member. The RippleNet algorithm is used to propagate preferences in the knowledge graph to predict the probability that the user will watch other content. If the defined triple of the video knowledge graph is G = {(h, r, t) | h ∈ ε, r ∈ R, t ∈ ε}, then the Kth associated entity of the target member is as shown in the following formula 8:

[0134]

[0135] in, Let h represent the set of video content viewed by the user in the history, r represent the entity relationship in the knowledge graph, and t represent the tail entity. The set of ripples for the user's k-th jump is:

[0136]

[0137] in, This represents the k-th hop triplet relation of the content entities that a user has viewed in the past, as extended in the knowledge graph.

[0138] Construct a ripple set of user-related k-th preference propagation. For each triple (h, r, t) in the set, multiply (h*r) by the content embedding vector to obtain the relevance score. Normalize the relevance score p using the softmax function. i The formula for calculating the relevance score is as follows:

[0139]

[0140] Among them, Ri ∈R d×d h i ∈R d They are relations r i and head node h i The embedding vector. p i For content v and entity h i In R i The similarity probability value in vector space. Obtain the probability p. i Afterwards, The relevant probability p corresponding to the middle tail i Perform a weighted summation and return the vector. The calculation formula is as follows:

[0141]

[0142] Where t i ∈R d It is the tail node t i The embedding vector, the vector This represents the relevant video content viewed in the user's history in the k-th hop of the knowledge graph.

[0143] In the RippleNet algorithm, the representations obtained through multiple interest expansions are summed to finally obtain the user embedding vector representation:

[0144]

[0145] Here, H represents the total number of hops in the knowledge graph of the RippleNet algorithm, which is a variable value that can be set according to actual needs.

[0146] The final predicted value is calculated by combining the video content's embedding representation v with the user's embedding representation u, using the following formula:

[0147]

[0148] Based on the final predicted values, a candidate set of videos for recommendation is determined. An attention mechanism is then used to score the attention of each video in the candidate set, thereby identifying the target video. Specifically, in group video recommendation, the recommendation results are merged. A weighted sum is calculated based on the age of family group members, the number of videos watched, and the predicted viewing score of the video, ultimately obtaining the target member's preference for video I. i The score pred_score(G, I) of the video recommendation candidate set (where I represents the pred_score of G, I) is calculated as follows: i The calculation formula is as follows:

[0149]

[0150] Among them, pred_score(u x I i ) represents user u x Watch video I i The predicted score, w(u) x G) represents user u x The weight value within its corresponding family group G is calculated using the following formula:

[0151]

[0152] Where β is the configured user weight coefficient, and act(u) represents the number of videos watched by user u.

[0153] In this embodiment, by propagating the preference information of target members within a video knowledge graph, candidate videos recommended to target members can be accurately obtained. The knowledge graph is integrated into video recommendation as content network topology information, and corresponding video content is recommended based on content network expansion, effectively solving the cold start and data sparsity problems. An attention mechanism is introduced into group recommendation to focus on key member groups. Different weights are used for weighted summation of users of different age groups to obtain the predicted viewing probability of candidate videos for the group, improving the accuracy of group recommendation. This allows for better provision of precise personalized recommendation services to users, increasing video content click-through rates and user stickiness.

[0154] Furthermore, step 4012 may also include:

[0155] Step 40121: Obtain preset behavioral factor weight coefficients, and calculate the preference score list of the target member based on the behavioral factor weight coefficients. The preference score list includes user identifiers and video identifiers.

[0156] Step 40122: Construct a user set based on the user identifier and a content set based on the video identifier;

[0157] Step 40123: Based on the preference score information of the target members in the preference score list, perform positive and negative sample division on the user set and the content set to obtain the interaction matrix between the target members and the video content.

[0158] When constructing the interaction matrix between users and video content, preset behavioral factor weight coefficients are obtained. These behavioral factors primarily include behavior type and behavior time. For behavior type, generally, higher costs correspond to greater weights, and weights are configured based on cost. For behavior time, the closer the behavior time is to the current time, the greater the tag weight corresponding to the behavior; that is, a current behavior carries a greater weight than a behavior that occurred a week ago. Based on this, different weight coefficients are configured for user behavior types and behavior times. By collecting, aggregating, and cleaning user behavior data from different applications, a unified content library is constructed, integrating user behavior data from different applications to ensure data consistency for subsequent processing. Through a constructed family social relationship network, behavioral data of user members within the same family group is aggregated to supplement user behavior data. Then, user behavior is quantified. Based on the collected user behavior data, user ratings for video content and user content tag ratings are generated according to factors such as the type of user interaction with video content, thus obtaining user behavior preference scores.

[0159] Specifically, based on the acquired behavioral factors, the user's behavioral preference score for the video is calculated using the following formula 16:

[0160]

[0161] Where act_weight is the user's behavior type weight coefficient, act_attenuation is the user's behavior time decay weight coefficient, and m is the number of times the user watches the same video within the statistical period.

[0162] Furthermore, the user behavior type weight coefficients are configured as shown in Table 7 below, and calculations are performed based on Table 7:

[0163] act_weight=weight*type_score (17)

[0164] The user behavior time decay weighting coefficient is:

[0165] act_attenuation=exp(-a*t) (18)

[0166] Where a is the attenuation coefficient and t is the number of days since the viewing behavior occurred.

[0167]

[0168] Table 7

[0169] For example, to statistically analyze the video viewing behavior data of all selected users over 90 days and quantify this behavior to obtain a list of user preference scores, the user behavior data is first summarized by linking it to the video information table as follows:

[0170]

[0171] Table 8

[0172] If the attenuation coefficient a = 1, then the user's score for watching the video is:

[0173] Score=1*exp(-1)+3*exp(-2)=0.77 (19)

[0174] Similarly, the score for watching video 00000001 can be calculated:

[0175] Score = 3 * exp(-3) = 0.15 (20)

[0176] The user preference score is calculated using the formula above, and an example of the corresponding preference score list is shown in List 9 below:

[0177] User ID Video Identifier User preference score 18867103731 00000000 0.77 18867103731 00000001 0.15

[0178] Table 9

[0179] This method calculates a list of preference scores for target members, constructs a user set based on the user identifiers of target members, constructs a content set of videos based on the video identifiers in the preference score list, and divides the video content into positive and negative samples based on the preference score information of target members for each video in the preference score list, thereby constructing an interaction matrix between target members and video content.

[0180] In this embodiment, user behavior is quantified based on historical video viewing data and constructed behavioral factors. Different weights are assigned to user preferences at different time periods. When providing personalized recommendation services to users, the system can adapt to changes in user preferences in a timely manner, thereby improving the accuracy of video recommendations.

[0181] In one embodiment, prior to step 401, the following may also be included:

[0182] Step 410: Obtain raw video data from different applications and clean it to obtain basic video data;

[0183] Step 420: Extract entity information from the basic video data, generate a unified content library based on the entity information, and perform knowledge reasoning and knowledge discovery based on the unified content library to construct an entity relationship graph of the entity information;

[0184] Step 430: Construct a video knowledge graph based on the entity relationship graph, wherein the video knowledge graph includes multiple entity relationship tuples.

[0185] In this embodiment, providing video recommendations to users is based on a pre-built video knowledge graph. When building the video knowledge graph, firstly, raw video data is obtained from different applications. The obtained raw video data is then cleaned to remove data containing garbled characters, missing or incomplete fields, movie and TV series trailers, behind-the-scenes footage, and clips, etc., to obtain basic video data and ensure the accuracy of the data used to build the knowledge graph.

[0186] Furthermore, entity information is extracted from the cleaned basic video data. This entity information includes content entities and relational entities, such as directors, actors, content tags, languages, and regions of movies and TV series. Based on the extracted entity information and video resources such as extracted movie and TV series episodes, a unified content library is generated. Video data obtained from different applications is stored in this unified content library, enriching the knowledge graph base. The video resources are structured data; for unstructured data, relationships are established through extracted entity information. In the generated unified content library, knowledge reasoning and knowledge discovery are used to construct relational graphs for various entities, including content entities and attribute entities, resulting in corresponding entity relation tables, such as movie-genre, movie-director, movie-actor, etc. Based on the constructed entity relation graphs, entity relation tuple data is built to generate the video knowledge graph.

[0187] In this embodiment, the cold start and data sparsity problems in video recommendation are solved by introducing a knowledge graph into the video recommendation process.

[0188] The video recommendation apparatus provided in the embodiments of this application is described below. The video recommendation apparatus described below can be referred to in correspondence with the video recommendation method described above.

[0189] Reference Figure 4 The video recommendation device provided in this application embodiment includes:

[0190] User screening module 10 is used to acquire social network data in the communication network and determine the user taking the photo based on the social network data:

[0191] Group building module 20 is used to extract the photography user's communication data and first behavior data from the photography user's personal terminal, and build the photography user's family group based on the communication data;

[0192] The data extraction module 30 is used to determine the public terminal of the family group based on the user taking the photo, and extract the second behavior data corresponding to the family group from the public terminal;

[0193] The preference recognition module 40 is used to determine the preference information of each family member in the family group based on the first behavior data and the second behavior data, and to obtain the location signaling data of each family member, and to determine the target member currently using the public terminal based on the location signaling data;

[0194] The video recommendation module 50 is used to determine target videos based on the preference information of the target members and recommend them in the public terminal.

[0195] In one embodiment, the group building module 20 is further configured to:

[0196] Based on the communication data, construct a set of positive and negative samples that have a family relationship with the user who took the photo;

[0197] The positive and negative sample sets are input into a pre-trained logistic regression model to obtain the probability values ​​of the binary relation pairs corresponding to the users who take photos.

[0198] Based on the probability values ​​of the binary relation pairs, determine each family member who has a family relationship with the user taking the photo;

[0199] Obtain the basic information of the user taking the photo and each of the family members, and calculate the information difference between the basic information of the user taking the photo and the basic information of each of the family members.

[0200] The family relationships between each family member and the user taking the photo are determined based on the information difference, and the family group of the user taking the photo is constructed.

[0201] In one embodiment, the preference recognition module 40 is further configured to:

[0202] Based on the first behavioral data, the preference information of the user taking the photo is determined; based on the second behavioral data, a preference set containing the preference information of each family member in the family group is determined.

[0203] The preference information in the preference set is filtered using the preference information of the user taking the photo, to determine the preference information of the hidden user of the public terminal in the family group;

[0204] The hidden user is identified based on the family relationship between each family member in the family group and the user taking the photo, so as to determine the preference information of each family member.

[0205] In one embodiment, the video recommendation module 50 is further configured to:

[0206] Based on the preference information of the target members, preference propagation is performed in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target members;

[0207] Based on the second behavioral data, an attention mechanism is used to score the videos in the video recommendation candidate set to determine the target video with the highest score and recommend it in the public terminal.

[0208] In one embodiment, the video recommendation module 50 is further configured to:

[0209] Collect the historical behavior data corresponding to the target member in the first behavior data and the second behavior data;

[0210] Construct an interaction matrix between the target member and the video content based on the historical behavior data;

[0211] The interaction matrix is ​​used to propagate the preference information of the target member in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target member.

[0212] In one embodiment, the video recommendation module 50 is further configured to:

[0213] Obtain preset behavioral factor weight coefficients, and calculate the preference score list of the target member based on the behavioral factor weight coefficients. The preference score list includes user identifiers and video identifiers.

[0214] A user set is constructed based on the user identifier, and a content set is constructed based on the video identifier;

[0215] Based on the preference score information of the target members in the preference score list, the user set and the content set are divided into positive and negative samples to obtain the interaction matrix between the target members and the video content.

[0216] In one embodiment, the video recommendation device further includes a knowledge graph construction module, used for:

[0217] Raw video data from different applications is acquired and cleaned to obtain basic video data;

[0218] Entity information is extracted from the basic video data, a unified content library is generated based on the entity information, and knowledge reasoning and knowledge discovery are performed based on the unified content library to construct an entity relationship graph of the entity information.

[0219] A video knowledge graph is constructed based on the entity relationship graph, and the video knowledge graph includes multiple entity relationship tuples.

[0220] Figure 5Example: A schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call a computer program in the memory 530 to execute the steps of the video recommendation method, such as including:

[0221] Acquire social network data from the communication network, and determine the user taking the photo based on the social network data;

[0222] Extract the user's communication data and first behavior data from the user's personal terminal, and construct the user's family group based on the communication data;

[0223] Based on the user taking the photo, determine the public terminal of the family group, and extract the second behavior data corresponding to the family group from the public terminal;

[0224] Based on the first behavioral data and the second behavioral data, the preference information of each family member in the family group is determined, and the location signaling data of each family member is obtained. Based on the location signaling data, the target member currently using the public terminal is determined.

[0225] The target video is determined based on the preference information of the target members and recommended in the public terminal.

[0226] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0227] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the video recommendation method provided in the above embodiments, such as including:

[0228] Acquire social network data from the communication network, and determine the user taking the photo based on the social network data;

[0229] Extract the user's communication data and first behavior data from the user's personal terminal, and construct the user's family group based on the communication data;

[0230] Based on the user taking the photo, determine the public terminal of the family group, and extract the second behavior data corresponding to the family group from the public terminal;

[0231] Based on the first behavioral data and the second behavioral data, the preference information of each family member in the family group is determined, and the location signaling data of each family member is obtained. Based on the location signaling data, the target member currently using the public terminal is determined.

[0232] The target video is determined based on the preference information of the target members and recommended in the public terminal.

[0233] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including:

[0234] Acquire social network data from the communication network, and determine the user taking the photo based on the social network data;

[0235] Extract the user's communication data and first behavior data from the user's personal terminal, and construct the user's family group based on the communication data;

[0236] Based on the user taking the photo, determine the public terminal of the family group, and extract the second behavior data corresponding to the family group from the public terminal;

[0237] Based on the first behavioral data and the second behavioral data, the preference information of each family member in the family group is determined, and the location signaling data of each family member is obtained. Based on the location signaling data, the target member currently using the public terminal is determined.

[0238] The target video is determined based on the preference information of the target members and recommended in the public terminal.

[0239] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0240] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0241] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0242] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A video recommendation method, characterized in that, Includes the following steps: Acquire social network data from the communication network, and determine the user taking the photo based on the social network data; Extract the user's communication data and first behavior data from the user's personal terminal, and construct the user's family group based on the communication data; Based on the user taking the photo, determine the public terminal of the family group, and extract the second behavior data corresponding to the family group from the public terminal; Based on the first behavioral data and the second behavioral data, the preference information of each family member in the family group is determined, and the location signaling data of each family member is obtained. Based on the location signaling data, the target member currently using the public terminal is determined. The target video is determined based on the preference information of the target members and recommended in the public terminal.

2. The video recommendation method according to claim 1, characterized in that, The step of constructing the family group of the user taking the photo based on the communication data includes: Based on the communication data, construct a set of positive and negative samples that have a family relationship with the user who took the photo; The positive and negative sample sets are input into a pre-trained logistic regression model to obtain the probability values ​​of the binary relation pairs corresponding to the users who take photos. Based on the probability values ​​of the binary relation pairs, determine each family member who has a family relationship with the user taking the photo; Obtain the basic information of the user taking the photo and each of the family members, and calculate the information difference between the basic information of the user taking the photo and the basic information of each of the family members. The family relationship between each family member and the user taking the photo is determined based on the information difference, and the family group of the user taking the photo is constructed.

3. The video recommendation method according to claim 2, characterized in that, The step of determining the preference information of each family member in the family group based on the first behavioral data and the second behavioral data includes: Based on the first behavioral data, the preference information of the user taking the photo is determined; based on the second behavioral data, a preference set containing the preference information of each family member in the family group is determined. The preference information in the preference set is filtered using the preference information of the user taking the photo, to determine the preference information of the hidden user of the public terminal in the family group; The hidden user is identified based on the family relationship between each family member in the family group and the user taking the photo, so as to determine the preference information of each family member.

4. The video recommendation method according to claim 1, characterized in that, The step of determining the target video based on the target members' preference information and recommending it on the public terminal includes: Based on the preference information of the target members, preference propagation is performed in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target members; Based on the second behavioral data, an attention mechanism is used to score the videos in the video recommendation candidate set to determine the target video with the highest score and recommend it in the public terminal.

5. The video recommendation method according to claim 4, characterized in that, The step of determining the video recommendation candidate set corresponding to the target member by performing preference propagation in a preset knowledge graph based on the target member's preference information includes: Collect the historical behavior data corresponding to the target member in the first behavior data and the second behavior data; Construct an interaction matrix between the target member and the video content based on the historical behavior data; The interaction matrix is ​​used to propagate the preference information of the target member in a preset video knowledge graph to determine the video recommendation candidate set corresponding to the target member.

6. The video recommendation method according to claim 5, characterized in that, The step of constructing the interaction matrix between the target member and the video content based on the historical behavior data includes: Obtain preset behavioral factor weight coefficients, and calculate the preference score list of the target member based on the behavioral factor weight coefficients. The preference score list includes user identifiers and video identifiers. A user set is constructed based on the user identifier, and a content set is constructed based on the video identifier; Based on the preference score information of the target members in the preference score list, the user set and the content set are divided into positive and negative samples to obtain the interaction matrix between the target members and the video content.

7. The video recommendation method according to claim 4, characterized in that, Before the step of determining the video recommendation candidate set corresponding to the target member by performing preference propagation in a preset video knowledge graph based on the target member's preference information, the method further includes: Raw video data from different applications is acquired and cleaned to obtain basic video data; Entity information is extracted from the basic video data, a unified content library is generated based on the entity information, and knowledge reasoning and knowledge discovery are performed based on the unified content library to construct an entity relationship graph of the entity information. A video knowledge graph is constructed based on the entity relationship graph, and the video knowledge graph includes multiple entity relationship tuples.

8. A video recommendation device, characterized in that, include: The user screening module is used to acquire social network data in the communication network and determine the user taking the photo based on the social network data. A group building module is used to extract the photography user's communication data and first behavior data from the photography user's personal terminal, and to build the photography user's family group based on the communication data. The data extraction module is used to determine the public terminal of the family group based on the user taking the photo, and extract the second behavior data corresponding to the family group from the public terminal; The preference recognition module is used to determine the preference information of each family member in the family group based on the first behavior data and the second behavior data, and to obtain the location signaling data of each family member, and to determine the target member currently using the public terminal based on the location signaling data; The video recommendation module is used to determine target videos based on the preference information of the target members and recommend them on the public terminal.

9. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the video recommendation method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the video recommendation method according to any one of claims 1 to 7.

Images