A method and system for predicting single-point pay-per-view video purchased by users of smart TV platform

Abstract

The present invention discloses a method and system for predicting the purchase of single-point pay-per-view videos by users of a smart TV platform, including: constructing an adjacency graph for video viewing records; constructing a video embedding vector for each video node in the adjacency graph; for the target user's video viewing records, obtaining a viewing sequence according to the time embedding vector and the video embedding vector of the viewing moment, performing attention weighting on each video in the viewing sequence to obtain viewing interest features; for the target user's video purchase records, obtaining a purchase preference feature according to the video embedding vector of the purchased video and the purchase amount; predicting and recommending recommended videos according to the viewing interest feature and the purchase preference feature. A graph is constructed for all user video viewing records to solve the problem of insufficient expression of the embedding vector due to the small number of single-point pay-per-view video records. The time factor is considered in the modeling of the video embedding vector, the viewing interest feature, and the purchase preference feature, so that the prediction result is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of video recommendation, and in particular to a method and system for predicting a video to be purchased by a user of a smart TV platform. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] A la carte videos are paid videos that users need to purchase separately and watch within a specified time frame. If they expire, they need to be purchased again. Understanding users' purchasing intentions is very important for both service providers and users. From the perspective of service providers, understanding users' current purchasing interests can increase the platform's revenue. Based on users' purchasing intentions, various preferential activities can be carried out, such as issuing limited-time coupons, to promote user consumption, increase revenue, and improve user stickiness. From the user's perspective, the platform can greatly improve the user experience by recommending videos based on users' viewing intentions.

[0004] Due to the limitations of current Internet smart TV hardware and network communications, the data transmitted to the backend by the device is mainly user viewing and purchase records with timestamps, as well as stored video metadata and user metadata. With the support of the above data, the purchase probability prediction of single-point paid videos and the recommendation of videos that users are interested in mainly have the following problems:

[0005] First, among all the available data, users’ viewing records reflect their viewing interests, and their purchase records reflect their purchase preferences. For single-point paid videos, the proportion of relevant viewing records is much smaller than that of non-paid videos, resulting in insufficient video embedding representation and inaccurate prediction when using traditional video embedding representation methods.

[0006] Second, existing video embedding representations or modeling of users’ viewing interests and purchasing preferences ignore the temporal factor.

[0007] Finally, the viewing interests reflected in the user's viewing history and the purchasing preferences reflected in the user's purchasing history are sometimes inconsistent. The videos the user usually watches reflect one interest, and the à la carte paid videos purchased reflect another interest. Traditional methods regard this as one of the user's interests, and the prediction is inaccurate. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a method and system for predicting the single-point pay-per-view video purchased by users on a smart TV platform, which constructs a graph of all user video viewing records to solve the problem of insufficient embedding vector expression caused by fewer single-point pay-per-view video records. The time factor is considered in the modeling of video embedding vectors, viewing interest features, and purchase preference features, making the prediction results more accurate.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for predicting a video purchased by a user of a smart TV platform, comprising:

[0011] Construct an adjacency graph based on the time sequence of videos watched by users in the video viewing history;

[0012] For each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes the time embedding vector of the viewing moment and the edges with the neighboring nodes;

[0013] For the target user’s video viewing history, the viewing sequence is obtained based on the time embedding vector and video embedding vector of the viewing moment, and attention is weighted for each video in the viewing sequence to obtain the viewing interest feature;

[0014] For the target user’s video purchase records, the purchase preference features are obtained based on the video embedding vectors of the purchased videos and their purchase amounts;

[0015] The recommended videos are predicted based on viewing interest characteristics and purchase preference characteristics, and videos are recommended to target users based on the obtained recommendation probability.

[0016] The present invention provides a method for predicting a video purchased by a user of a smart TV platform, and also provides a method for recommending a video, which can more accurately know the viewing interest of the user based on the viewing record and purchase record of the user.

[0017] As an optional implementation, the time embedding vector of the viewing moment is:

[0018]

[0019] Among them, t2v(t) represents the time embedding vector corresponding to time t, t2v(t)[i] represents the i-th element in the time embedding vector, with a total of k elements; Φ represents the periodic function, ω i and are the frequency and phase shift of the periodic function respectively.

[0020] As an optional implementation, the adjacency graph uses videos at the viewing time as nodes, uses two videos in a chronological order as edges, and assigns weights to the edges according to the interval time between the two videos.

[0021] As an optional implementation, the process of constructing a video embedding vector includes: using a multi-head attention component to extract multi-head attention features from its own feature splicing vector and the feature splicing vectors of its neighboring nodes, splicing the extracted head features, and after a linear transformation, using a multi-layer feedforward neural network to obtain the video embedding vector.

[0022] As an optional implementation, the process of obtaining viewing interest features includes: using a recurrent neural network and an attention mechanism to perform serialization modeling and attention weighting on the user's video viewing records to obtain viewing interest features.

[0023] As an optional implementation, the process of obtaining purchase preference features includes: forming a first matrix by chronologically embedding the video embedding vectors of the purchased videos, and forming a second matrix by copying the purchase amount according to the number of videos in the video purchase record; obtaining a vector representation matrix that integrates the purchase preference information based on the first matrix and the second matrix, and obtaining the purchase preference features using a multi-head attention network on the vector representation matrix.

[0024] As an optional implementation, when predicting the video to be recommended, the video embedding vector and user basic information of the video to be recommended are added as prediction inputs.

[0025] In a second aspect, the present invention provides a system for predicting a video purchased by a user of a smart TV platform, comprising:

[0026] A graph construction module configured to construct an adjacency graph based on a time sequence of videos watched by a user in a video viewing record;

[0027] A video embedding vector construction module is configured to construct a video embedding vector for each video node in the adjacency graph according to its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes a time embedding vector of the viewing moment and edges with neighboring nodes;

[0028] The viewing interest feature extraction module is configured to obtain a viewing sequence based on the time embedding vector and the video embedding vector of the viewing moment for the target user's video viewing record, and perform attention weighting on each video in the viewing sequence to obtain a viewing interest feature;

[0029] A purchase preference feature extraction module is configured to obtain purchase preference features based on the video embedding vectors of the purchased videos and the purchase amounts of the target users' video purchase records;

[0030] The prediction module is configured to predict the recommended videos according to the viewing interest characteristics and the purchase preference characteristics, and recommend videos to the target users according to the obtained recommendation probability.

[0031] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0032] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention proposes a method for predicting the single-point paid video purchased by users of a smart TV platform, or a paid video recommendation method, or a video recommendation method, which constructs a graph of all user video viewing records, and then calculates the video embedding vector of each video to solve the problem of insufficient embedding vector expression caused by fewer single-point paid video records. Secondly, the user's video viewing record and video purchase record are modeled separately to obtain the user's viewing interest characteristics and purchase preference characteristics respectively, accurately capture and distinguish the user's viewing interest and purchase preference, and in the modeling of the video embedding vector, viewing interest characteristics, and purchase preference characteristics, the time factor is considered, so that the prediction result is more accurate.

[0035] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 A block diagram of a method for predicting a single-point pay-per-view video purchased by a user of a smart TV platform provided in Embodiment 1 of the present invention;

[0038] Figure 2 for Figure 1 Zoomed-in view of the adjacency graph during the video embedding representation stage;

[0039] Figure 3 A schematic diagram of constructing a video embedding vector provided in Example 1 of the present invention;

[0040] Figure 4 A schematic diagram of extracting purchase preference features provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0044] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0045] Example 1

[0046] like Figure 1 As shown, this embodiment provides a method for predicting a single-point pay-per-view video purchased by a user of a smart TV platform, comprising:

[0047] Obtain the user's video viewing history and video purchase history;

[0048] Construct an adjacency graph based on the time sequence of videos watched by users in the video viewing history;

[0049] For each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes the time embedding vector of the viewing moment and the edges with the neighboring nodes;

[0050] For the target user’s video viewing history, the viewing sequence is obtained based on the time embedding vector and video embedding vector of the viewing moment, and attention is weighted for each video in the viewing sequence to obtain the viewing interest feature;

[0051] For the target user’s video purchase records, the purchase preference features are obtained based on the video embedding vectors of the purchased videos and their purchase amounts;

[0052] The paid videos to be predicted are predicted based on viewing interest characteristics and purchase preference characteristics, and videos are recommended to target users based on the obtained purchase probability.

[0053] The method of this embodiment can also be used to recommend paid videos of interest, predict the paid videos to be recommended based on the viewing interest characteristics and purchase preference characteristics of the target user, and recommend paid videos of interest to the target user based on the probability.

[0054] Specifically, in more embodiments, a method for recommending a single-point pay-per-view video on a smart TV platform is provided, including:

[0055] Obtain the user's video viewing history and video purchase history;

[0056] Construct an adjacency graph based on the time sequence of videos watched by users in the video viewing history;

[0057] For each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes the time embedding vector of the viewing moment and the edges with the neighboring nodes;

[0058] For the target user’s video viewing history, the viewing sequence is obtained based on the time embedding vector and video embedding vector of the viewing moment, and attention is weighted for each video in the viewing sequence to obtain the viewing interest feature;

[0059] For the target user’s video purchase records, the purchase preference features are obtained based on the video embedding vectors of the purchased videos and their purchase amounts;

[0060] The recommended paid videos are predicted based on viewing interest characteristics and purchase preference characteristics, and videos are recommended to target users based on the obtained recommendation probability.

[0061] It is understandable that the above method can be used for interest recommendations for ordinary videos in addition to paid videos.

[0062] Specifically, in more embodiments, a smart TV platform video recommendation method is also provided, including:

[0063] Obtain the user's video viewing history and video purchase history;

[0064] Construct an adjacency graph based on the time sequence of videos watched by users in the video viewing history;

[0065] For each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes the time embedding vector of the viewing moment and the edges with the neighboring nodes;

[0066] For the target user’s video viewing history, the viewing sequence is obtained based on the time embedding vector and video embedding vector of the viewing moment, and attention is weighted for each video in the viewing sequence to obtain the viewing interest feature;

[0067] For the target user’s video purchase records, the purchase preference features are obtained based on the video embedding vectors of the purchased videos and their purchase amounts;

[0068] The recommended videos are predicted based on viewing interest characteristics and purchase preference characteristics, and videos are recommended to target users based on the obtained recommendation probability.

[0069] The method provided in the above embodiment is described in detail below.

[0070] It should be noted that in this embodiment, all data is obtained in compliance with laws and regulations and based on user consent, and the data is used legally.

[0071] Time is a very important factor for expressing user interests or for expressing video embedding vectors. From the user's perspective, for multi-person usage scenarios of Internet smart TVs, there will be different viewing patterns on weekdays, weekends, and even at different times of the day. Through the analysis of a large number of user viewing records, users' viewing behavior on weekdays is more intensive than on weekends, and there are also obvious differences in viewing preferences in the morning, afternoon, and evening time periods of the day.

[0072] From the perspective of videos, especially paid videos, movies account for a large proportion. When these videos are first shown on the platform, the number of views and purchases will peak in a short period of time, and then gradually decline, showing a linear change over time. However, videos such as TV series are updated at a fixed time every week, and the number of views and purchases will show cyclical changes.

[0073] Therefore, in order to capture the linear or periodic changing time characteristics of the user's video viewing history and video purchase history, this embodiment uses the following time embedding method to obtain the time embedding vector of the viewing moment:

[0074]

[0075] Where t2v(t) represents the time embedding vector corresponding to time t, t2v(t)[i] represents the i-th element in the time embedding vector, Φ represents the periodic activation function, and the commonly used periodic activation functions are sin or cos function, ω i and are learnable parameters, representing the frequency and phase shift of the periodic function respectively.

[0076] Since the periodic function The period is So for time points t and have the same value. Therefore, the periodicity can be captured simply by using periodic functions without explicit feature engineering; at the same time, using Linear terms are used to capture non-periodic patterns that depend on time.

[0077] The temporal embedding vector obtained above is used in the process of video embedding vector representation and modeling of users' viewing interests and purchasing preferences.

[0078] In order to better learn at a specific moment t i Video i The embedding vector representation of t is used in this embodiment to construct an adjacency graph based on the video viewing records of all users based on the graph attention network; i Videos watched all the time i as nodes, and two videos in chronological order as edges i .

[0079] For example, setting t 0 is the current time, a user at time t 1 (t 1 <t 0 )Watch the video 1 , another user at time t 2 (t 2 <t 0 )Watch the video 2 , the third user at time t 3 (t 3 <t 0 )Watch the video 3 , and these three users are respectively 1 -t 0 , t 2 -t 0 , t 3 -t 0 Watched the video in these three time periods 0 , so in the video v 0 and video v 1 、v 2 、v 3 Set the edge e between 10 、e 20 、e 30 ,like Figure 2 shown.

[0080] Among them, edge e ijRepresented by a vector, each bit in the vector has a specific meaning. Some bits can indicate whether the two videos connected by this edge belong to the same primary classification, and some bits can indicate whether the two videos connected by this edge belong to the same secondary classification. If the attributes of the two videos are the same, the bit is set to 1, otherwise it is set to 0, and the interval time between the two videos is used as the weight to weight the vector. ij The specific settings can be changed according to the experimental results.

[0081] In this embodiment, for each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes, such as Figure 3 As shown, for the central video node v 0 , represents the video embedding vector according to the neighbor nodes and combined with specific aggregation operations; for the video at time t 0 Watched video node v 0 , its neighbors are defined as:

[0082] N 0 = {v 1 (t 1 ),v 2 (t 2 ),…,v n (t n )}.

[0083] For each video node v in the adjacency graph i , we need to calculate its video embedding vector. The feature concatenation vector of the video node consists of three parts, one of which is defined as in, Indicates video v i The output of the first-order neighbor node in the l-1 layer is the embedding vector obtained by the neighbor node in the previous iteration, which is also obtained by its first-order neighbor through the ET-GAT method. If l is 1, then It's video v i The original features or randomly initialized features.

[0084] Part of it is t′ i The time embedding vector t2v(t′) corresponding to the moment i ), t′ i is the time t when the video viewing behavior occurs i With the time origin t origin The time interval, time origin t origin Defined as the earliest time point in the user's viewing and purchasing records in the system.

[0085] One part is node v i and vj The edge between ij , edge ij The contained information can also be integrated into the hidden layer representation of the target video node.

[0086] The above three steps are combined to obtain the video node v i The feature concatenation vector H of the lth layer i (t i );

[0087]

[0088] The feature concatenation vectors of the neighbor nodes and the video node are used together to extract the multi-head attention features using the multi-head attention component, where the multi-head attention component is:

[0089] Q i =H i W Q ,K i =H i W K ,V i =H i W V

[0090]

[0091] Among them, Q, K, and V are the parameters in the multi-head attention mechanism; d k K i The dimension of the vector is used to prevent the gradient from disappearing.

[0092] Calculate the matrix Q i With K i , thereby obtaining the weight coefficient of the attribute value corresponding to each key, and after normalization through softmax, the weight and the corresponding value are weighted summed to obtain the Attention score in the head.

[0093] MultiHead(Q,K,V)=Concat(head 1 ,head 2 ,…,head n )W O

[0094]

[0095] The above two formulas show that the parameters between multiple heads are not shared, and the parameter W for linear transformation of Q, K, and V inside the head is also different, so modeling can be performed from multiple subspaces.

[0096] After splicing the extracted head features, a linear transformation is performed, and then a multi-layer feedforward neural network is used to obtain the multi-head attention mechanism result. That is, the video node v i Video embedding vector at level l.

[0097] At this stage, this embodiment sets a connection prediction task to train the parameters of the model to obtain a more accurate video embedding vector representation.

[0098]

[0099] Among them, the loss function L takes the v existing in the user's video viewing record as i and v j Considered as a positive sample set S p , v i and v j There is an edge between them, and an equal number of negative sample sets S are randomly generated according to the number of positive samples n , in the negative sample set v i and v j There is no edge between them; σ is the sigmoid activation function. After multiple iterations, the obtained video embedding vector can be used in downstream tasks.

[0100] Since it can be trained independently, the video embedding vector obtained in this stage is used as the input of the next stage. It can also complete some functions independently in the system, such as completing the function of recommending similar videos by calculating the similarity between video embedding vectors.

[0101] In this embodiment, different model structures are used to model the user's viewing interests and purchasing preferences according to the user's video viewing records and video purchase records, mainly based on the different characteristics of the two sequences. The user's video viewing records are relatively dense, the time interval between watching videos is small, and the dependence before and after is strong. Considering the performance of the smart TV platform system, the user's video viewing records are modeled by combining a recurrent neural network and an attention mechanism.

[0102] A fixed number of video viewing records are used, and GRU is used as the basic unit to avoid the "gradient disappearance" problem. At the same time, GRU has a similar effect to LSTM, but is slightly faster.

[0103] The target user’s video viewing records are serialized according to the time embedding vector and video embedding vector of the viewing moment, and the modeling is as follows:

[0104] v′ t =v t||t2v(t′)

[0105] r t =σ(W ur v t +W hr h t-1 +b r )

[0106] Among them, v′ t represents the result of concatenating the video embedding vector and the time embedding vector of the viewing moment after being processed by the time embedding layer, r t is the reset gate in the GRU unit, h t-1 represents the result of the previous GRU unit, σ is the sigmoid function, and the result is used as the gating signal, W ur , W hr 、b r These are all trainable parameters.

[0107] After getting the reset gate signal, the previous hidden layer signal is reset, which is equivalent to selectively memorizing the state at the previous moment;

[0108] h′=tanh(W uh v t +W hh (r t ⊙h t-1 )+b h )

[0109] Among them, ⊙ represents the Hadamard product, that is, the matrix element multiplication operation.

[0110] z t =σ(W uz v t +W hz h t-1 +b z )

[0111] h t =(1-z t )⊙h t-1 +z t ⊙h′

[0112] Among them, z t Indicates the update gate signal, the value range is between 0 and 1. The closer it is to 1, the more data is "memorized".

[0113] Finally, we get the result h of the current GRU unit t , the result not only reflects the current input u t , and it is also the modeling result of the current sequence;

[0114] h=h1 ,h 2 ,…,h n ∈R n×d

[0115] Where n represents the number of videos in the sequence or the number of GRU units, and d represents the number of dimensions of each hidden layer vector, which is the same as the dimension of the video embedding vector.

[0116] After serialization modeling, a viewing sequence is obtained. In order to capture the correlation between the videos in the viewing sequence and the video to be predicted, this embodiment adds an attention unit on top of the GRU unit to calculate the attention score of each video in the sequence, and then integrates the attention score into the viewing sequence;

[0117]

[0118]

[0119] Among them, v c represents the embedding vector representation of the video to be recommended, h t represents the hidden layer output of the i-th GRU unit, a i represents the attention score of each video in the viewing sequence, It represents the viewing interest characteristics of the target users.

[0120] Compared with the user's video viewing record, the video purchase record contains the amount paid when purchasing the video. For the same user, the more the amount paid, the more interested the user is in the video. Therefore, the amount can be used as the purchase weight of the video.

[0121] In this embodiment, the amount is first normalized. The purpose of the normalization operation is that the consumption capacity of different users may be different, and the normalization operation can eliminate the deviation of the consumption capacity between different users;

[0122]

[0123] Among them, x is a vector composed of the amount of money in the user's video purchase record (in addition, not only the amount, but also multiple attributes such as release time can be used as weights), Min(x) represents the minimum consumption amount in each user's video purchase record, and Max(x) represents the maximum consumption amount in each user's video purchase record.

[0124] like Figure 4 As shown, the video embedding vectors in the user's video purchase record are arranged in chronological order to form a first matrix C, and the normalized purchase amount vector x is copied n times to form a second matrix X, where n represents the number of videos in the user's purchase record;

[0125] F=X T V

[0126] Among them, F is the video embedding vector representation matrix that integrates user purchase preference information.

[0127] Then, the matrix is ​​used in a multi-head attention network. Due to the structure of the multi-head attention network, it can establish multiple subspaces, that is, use multiple parameter matrices to model the user's purchase history, which can explicitly distinguish the viewing interests and purchase preferences contained in the user's purchase history. Finally, the purchase preference features that represent the user's long-term interests and consumption preferences are obtained.

[0128] Therefore, the video to be recommended is predicted based on the viewing interest characteristics and purchase preference characteristics; in addition, the video embedding vector and user basic information of the video to be recommended can be added in the prediction process, and the final prediction probability can be obtained by using multiple fully connected layers, as shown in the following formula:

[0129]

[0130] p(Z)=sigmoid(W z Z+b)

[0131] Among them, p(Z) represents the predicted probability, u profile , v c They represent the user's basic information, viewing interest features, purchase preference features, and the video embedding vectors of the video to be recommended.

[0132] The above problem is a binary classification problem, and the loss function can be defined as the binary cross entropy loss function:

[0133]

[0134] Among them, the sample label y i ∈{0,1}, 1 indicates that the user will purchase the item in the next week, and 0 indicates the opposite; N represents the number of all training samples.

[0135] Finally, the probability of the recommended videos is predicted, and videos are recommended to target users based on the probability.

[0136] Example 2

[0137] This embodiment provides a system for predicting a video purchased by a user of a smart TV platform, including:

[0138] A graph construction module configured to construct an adjacency graph based on a time sequence of videos watched by a user in a video viewing record;

[0139] A video embedding vector construction module is configured to construct a video embedding vector for each video node in the adjacency graph according to its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes a time embedding vector of the viewing moment and edges with neighboring nodes;

[0140] The viewing interest feature extraction module is configured to obtain a viewing sequence based on the time embedding vector and the video embedding vector of the viewing moment for the target user's video viewing record, and perform attention weighting on each video in the viewing sequence to obtain a viewing interest feature;

[0141] A purchase preference feature extraction module is configured to obtain purchase preference features based on the video embedding vectors of the purchased videos and the purchase amounts of the target users' video purchase records;

[0142] The prediction module is configured to predict the recommended videos according to the viewing interest characteristics and the purchase preference characteristics, and recommend videos to the target users according to the obtained recommendation probability.

[0143] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0144] In further embodiments, there is also provided:

[0145] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0146] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0147] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0148] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0149] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0150] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0151] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting the purchase of a single-point pay-per-view video by a user of a smart TV platform, characterized in that: include: An adjacency graph is constructed based on the time sequence of videos watched by the user in the video viewing record; the adjacency graph uses the videos at the viewing time as nodes, uses the edges between two videos with a time sequence, and assigns weights to the edges according to the interval time between the two videos; For each video node in the adjacency graph, a video embedding vector is constructed based on its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes the time embedding vector of the viewing moment and the edges with the neighboring nodes; For the target user’s video viewing history, the viewing sequence is obtained based on the time embedding vector and video embedding vector of the viewing moment, and attention is weighted for each video in the viewing sequence to obtain the viewing interest feature; For the target user’s video purchase records, the purchase preference features are obtained based on the video embedding vectors of the purchased videos and their purchase amounts; The recommended videos are predicted based on viewing interest characteristics and purchase preference characteristics, and videos are recommended to target users based on the obtained recommendation probability.

2. A method for predicting a video purchased by a user of a smart TV platform as claimed in claim 1, characterized in that: The time embedding vector of the viewing moment is: in, express The time embedding vector corresponding to the moment, represents the first elements, a total of k elements; represents a periodic function, and are the frequency and phase shift of the periodic function respectively.

3. A method for predicting a video purchased by a user of a smart TV platform as claimed in claim 1, characterized in that: The construction process of the video embedding vector includes: using the multi-head attention component to extract multi-head attention features from its own feature splicing vector and the feature splicing vector of its neighboring nodes, splicing the extracted head features, and after a linear transformation, using a multi-layer feedforward neural network to obtain the video embedding vector.

4. A method for predicting a video purchased by a user of a smart TV platform as claimed in claim 1, characterized in that: The process of obtaining viewing interest features includes: using a recurrent neural network and an attention mechanism to perform serial modeling and attention weighting on the user's video viewing records to obtain viewing interest features.

5. The method for predicting a video purchased by a user of a smart TV platform according to claim 1, characterized in that: The process of obtaining purchase preference features includes: forming a first matrix by arranging the video embedding vectors of purchased videos in chronological order, and forming a second matrix by copying the purchase amount according to the number of videos in the video purchase record; obtaining a vector representation matrix that integrates purchase preference information based on the first matrix and the second matrix, and obtaining purchase preference features using a multi-head attention network on the vector representation matrix.

6. A method for predicting a video purchased by a user of a smart TV platform as claimed in claim 1, characterized in that: When predicting the recommended video, the video embedding vector and user basic information of the recommended video are added as prediction input.

7. A system for predicting the purchase of single-point pay-per-view videos by users of a smart TV platform, characterized in that: include: A graph construction module is configured to construct an adjacency graph based on the time sequence of videos watched by the user in the video viewing record; the adjacency graph uses the videos at the viewing time as nodes, uses the edges between two videos with a time sequence, and assigns weights to the edges according to the interval time between the two videos; A video embedding vector construction module is configured to construct a video embedding vector for each video node in the adjacency graph according to its own feature splicing vector and the feature splicing vectors of its neighboring nodes; the feature splicing vector includes a time embedding vector of the viewing moment and edges with neighboring nodes; The viewing interest feature extraction module is configured to obtain a viewing sequence based on the time embedding vector and the video embedding vector of the viewing moment for the target user's video viewing record, and perform attention weighting on each video in the viewing sequence to obtain a viewing interest feature; A purchase preference feature extraction module is configured to obtain purchase preference features based on the video embedding vectors of the purchased videos and the purchase amounts of the target users' video purchase records; The prediction module is configured to predict the recommended videos according to the viewing interest characteristics and the purchase preference characteristics, and recommend videos to the target users according to the obtained recommendation probability.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

Images