Program sequence recommendation method and system based on group interest multi-level decoupling

Abstract

The application provides a program sequence recommendation method and system based on group interest multi-level decoupling, and belongs to the technical field of personalized recommendation. The method comprises the following steps: a group sequence encoder is used to encode a group interaction sequence processed by an embedding layer to obtain a group representation; a group-level interest decoupling module is used to decouple the interest representations of different members from the group representation; a group variational autoencoder in a member-level interest decoupling module is used to further decouple the interest representations of different members decoupled by the group-level interest decoupling module to obtain decoupled representations of different members and a group consensus representation; the decoupled representations of different members and the group consensus representation are used for next item prediction to obtain candidate item scores of the group at the next moment, and the group is recommended. By using the application, the problem of poor IPTV program recommendation accuracy and robustness of the existing family group can be solved.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and more specifically, to a program sequence recommendation method, system, and computer-readable storage medium based on multi-level decoupling of group interests. Background Technology

[0002] With the rise of IPTV, the importance and necessity of personalized IPTV recommendation methods have become increasingly prominent in response to the massive program resources and diverse user interests. Unlike other recommendation scenarios, IPTV recommendation is aimed at family groups rather than individual users. Family groups have the following characteristics: (1) Unknown member attributes and behaviors: Each member shares an account as a family unit. The generated clicks and viewing behaviors are recorded in the shared account. The terminal cannot distinguish which members are in the group or which member generated each behavior. That is, a single account carries the behavior of multiple members, and the data noise and sparsity are far greater than those of general groups; (2) Unique and dynamic member interests: Each member has its own unique interests, which will change dynamically over time; (3) Group consensus: At certain times, some members of the family group will reach an agreement and achieve consensus. That is, the interactive behavior of the family group is driven by the intertwined group consensus and the interests of each member. There is a complex relationship and interaction between the group consensus and the interests of each member, which makes the representation learned under the condition of ignoring the characteristics of the family group redundant and noisy. Therefore, existing IPTV recommendation methods ignore the characteristics of family groups in IPTV scenarios, making it difficult to effectively distinguish the individual interests and family consensus of each member, thus affecting the accuracy and robustness of IPTV program recommendations for family groups. Summary of the Invention

[0003] In view of the above problems, the purpose of this invention is to provide a program sequence recommendation method and system based on multi-level decoupling of group interests, so as to solve the problems of poor accuracy and robustness of existing IPTV program recommendations for family groups.

[0004] This invention provides a program sequence recommendation method based on multi-level decoupling of group interests, comprising: encoding a group interaction sequence processed by an embedding layer through a group sequence encoder to obtain a group representation;

[0005] The interest representations of different members are decoupled from the group representation through the group-level interest decoupling module;

[0006] The interest representations of different members decoupled by the group-level interest decoupling module are further decoupled by the group variational autoencoder in the member-level interest decoupling module to obtain the decoupled representations of different members and the group consensus representation.

[0007] The decoupled representations of the different members and the group consensus representation are used to make the next prediction, obtain the candidate item scores of the group at the next time step, and make recommendations to the group.

[0008] Preferably, the embedding layer processes the group interaction sequence using the following formula:

[0009] e g =Dropout(LayerNorm(m g +p g )),

[0010] in, This indicates the embedding initialization of the group interaction sequence;

[0011] m g ∈R n×d The item embedding represents the group interaction sequence;

[0012] p g ∈R n×d This represents the positional embedding of the group interaction sequence, where n represents the maximum sequence length;

[0013] Dropout(·) and LayerNorm(·) represent the Dropout operation and the normalization layer, respectively.

[0014] Preferably, the group sequence encoder encodes the group interaction sequence processed by the embedding layer using the following formula:

[0015] Among them, H g ∈R n×d The output of the group representation indicates all times;

[0016] Ψ represents the learnable parameters in the group sequence encoder;

[0017] This represents the interest representation of the group interaction sequence at time t;

[0018] During the encoding of the group interaction sequence processed by the embedding layer using a group sequence encoder, an auxiliary loss function is used to maximize the log-likelihood function of the next prediction of the N group interaction sequences at all time steps. The formula for the auxiliary loss function is as follows:

[0019]

[0020] in, and Let represent the embedding of the target item and the embedding of the negative sampled item, respectively, and σ(·) represent the sigmoid function.

[0021] Preferably, the step of decoupling the interest representations of different members from the group representation through the group-level interest decoupling module includes:

[0022] The embedding representation of the group interaction sequence at each time step is processed by a deep model, and the corresponding member prototypes are activated and assigned relevance weights.

[0023] The importance of the group interaction sequence to the prediction of the next item of the group is obtained by using attention weights and the embedding representation of the group interaction sequence at each time step.

[0024] Based on the member prototypes and relevance weights, and the importance of the next prediction for the group, the interest representations of different members are decoupled.

[0025] Preferably, the formula for calculating the prototype weights of members of a group at time step i is:

[0026]

[0027] Where k = 1, 2, ..., K; p k|i The group representation at time step i represents the correlation with the prototype of the k-th member. c represents the group representation at the i-th time step. k Represents the prototype of the k-th member; LN i This represents different LayerNorm layers.

[0028] Preferably, the formula for the attention weight is:

[0029]

[0030] Where i = 1, 2, ..., n; W ∈ R d×d , ρ∈R d as well as All represent learnable parameters;

[0031] p i This indicates attention weight; query indicates a query. Represents the group representation at the i-th time step; Key; i This indicates residual connection and nonlinearization of the bond; W T d represents the transpose of W; d represents the embedding dimension.

[0032] Preferably, the formula for decoupling the interest representations of different members is:

[0033]

[0034] Where K = 1, 2, ..., K; i = 1, 2, ..., n;

[0035] Indicates the interests of different members; Let represent the interest representation of member k at time i in group g.

[0036] In the process of decoupling the interest representations of different members from the group representation through the group-level interest decoupling module, the interest representation is optimized using a member prototype loss function, the formula of which is:

[0037]

[0038] Where k = 1, 2, ..., K; i and j represent the i-th and j-th time steps; and Represents the member prototype c i and c j L2 norm normalization; The group at time step i is represented by h. i L2 norm normalization.

[0039] Preferably, the further decoupling of the interest representations of different members decoupled by the group-level interest decoupling module through the group variational autoencoder in the member-level interest decoupling module to obtain the decoupled representations of different members and the group consensus representation includes:

[0040] The interest representations of different members in the same group are further decoupled by a member encoder to obtain decoupled representations of different members;

[0041] The interest representations among different members of the same group are processed by a group-shared encoder to obtain a group consensus representation;

[0042] The decoupled representations of the different members and the group consensus representation are processed by the decoder to obtain the decoupled representations of the different members and the group consensus representation.

[0043] Preferably, the formula for the member encoder is:

[0044]

[0045] Block i (X) = Dropout(LayerNorm(FFN) i (X))),

[0046] FFN i (X)=Tanh(XW i +b i ),

[0047] in, Decoupled representation of different members;

[0048] k = 1, 2, ..., K; Dropout and LayerNorm represent the Dropout operation and layer normalization operation, respectively, W i and b i The parameters to be learned are represented, i = 1, 2; Tanh represents the activation function; FFN i (·) represents the i-th feedforward neural network.

[0049] Preferably, the formula used by the group-shared encoder is:

[0050]

[0051] in, This represents the group consensus representation; G-Encoder(·) shared This indicates a group-shared variational encoder.

[0052] Preferably, the decoder uses the following formula:

[0053] Among them, O g ∈R n×d =[o1,o2,…,o n ] represents the predicted embedding; FFN(·) represents the feedforward neural network; This indicates a group consensus representation. Z, a decoupled representation of the interests of group members g .

[0054] In the process of processing the decoupled representations of the different members and the group consensus representation concatenated together by the decoder, the variational autoencoder is trained using the ELBO loss function, the formula of which is:

[0055]

[0056] in, Represent the prior distribution; φ represents the variational approximation function of the posterior distribution with parameter λ in the k-th member encoder; φ represents the P-value in the decoder. φ The parameters of the (x|z) function; β represents the Kullback-Leibler regularization coefficient, and the constraint member interest represents the correlation of each dimension.

[0057] Preferably, matrix multiplication is used to obtain the candidate item scores of the group at the next time step n+1. The following formula is used:

[0058]

[0059] Where E represents the item embedding matrix; S g This represents the sequence of item interactions in group g; Let P represent the next prediction result for group g, and let P represent the probability score.

[0060] This invention also provides a program sequence recommendation system based on multi-level decoupling of group interests, comprising:

[0061] The group representation acquisition module is used to encode the group interaction sequence processed by the embedding layer through a group sequence encoder to obtain the group representation;

[0062] A group-level interest decoupling module is used to decouple the interest representations of different members from the group representation;

[0063] The member-level interest decoupling module is used to further decouple the interest representations of different members decoupled by the group-level interest decoupling module, and obtain the decoupled representations of different members and the group consensus representation;

[0064] The prediction module is used to make the next prediction on the decoupled representations of the different members and the group consensus representation, obtain the candidate item scores of the group at the next time step, and make recommendations to the group.

[0065] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described program sequence recommendation method based on multi-level decoupling of group interests.

[0066] As can be seen from the above technical solution, the program sequence recommendation method and system based on multi-level decoupling of group interests provided by this invention uses a Transformer-based group sequence encoder to capture the potential temporal dependencies between interactive items in the temporal domain, and introduces an auxiliary loss function to improve the stability of the group sequence encoder and accelerate convergence. Then, a two-stage interest decoupling module is used, including a group-level interest decoupling module and a member-level interest decoupling module. The group-level interest decoupling module identifies and separates independent member prototypes in an unsupervised manner through an end-to-end prototype learning method, thereby decoupling the unique and dynamic interest representations of different members from the group representation. To accelerate the convergence of member prototypes during training and enhance the group-level decoupling effect, this invention also introduces a simple and effective member prototype loss function. The member-level interest decoupling module further decouples the interests of each member within the group through a group variational autoencoder, ensuring that each dimension of the member's interest representation is relatively independent and more representative, thereby enhancing the interest representation of each member. While achieving member-level interest decoupling, it also models a deep-level group consensus; thus solving the problems of poor accuracy and robustness of existing IPTV program recommendations for family groups.

[0067] To achieve the foregoing and related objectives, one or more aspects of the invention include the features that will be described in detail below. The following description and accompanying drawings illustrate certain exemplary aspects of the invention. However, these aspects indicate only a few of the various ways in which the principles of the invention can be used. Furthermore, the invention is intended to encompass all such aspects and their equivalents. Attached Figure Description

[0068] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0069] Figure 1 This is a schematic diagram of the program sequence recommendation method based on multi-level decoupling of group interests according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the logical structure of a program sequence recommendation system based on multi-level decoupling of group interests according to an embodiment of the present invention.

[0071] In all the accompanying drawings, the same reference numerals indicate similar or corresponding features or functions. Detailed Implementation

[0072] In the following description, numerous specific details are set forth for illustrative purposes and to provide a thorough understanding of one or more embodiments. However, it will be apparent that these embodiments may also be implemented without these specific details. In other instances, well-known structures and devices are shown in block diagram form for ease of description of one or more embodiments.

[0073] To address the aforementioned issues of poor accuracy and robustness in existing IPTV program recommendations for family groups, this invention provides a program sequence recommendation method and system based on multi-level decoupling of group interests.

[0074] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0075] To illustrate the program sequence recommendation method based on multi-level decoupling of group interests provided by this invention, Figure 1 The flowchart of a program sequence recommendation method based on multi-level decoupling of group interests according to an embodiment of the present invention is shown.

[0076] like Figure 1 As shown, the program sequence recommendation method based on multi-level decoupling of group interests provided by this invention includes:

[0077] S1: Encode the group interaction sequence processed by the embedding layer using a group sequence encoder to obtain the group representation;

[0078] S2: Decouple the interest representations of different members from the group representation using the group-level interest decoupling module;

[0079] S3: The interest representations of different members decoupled by the group-level interest decoupling module are further decoupled by the group variational autoencoder in the member-level interest decoupling module to obtain the decoupled representations of different members and the group consensus representation;

[0080] S4: Make the next prediction on the decoupled representation of the different members and the group consensus representation, obtain the candidate item score of the group at the next moment, and recommend it to the group.

[0081] In embodiments of this invention, the aim is to improve the accuracy and robustness of IPTV program recommendations by hierarchically decoupling member interests and modeling group consensus based on the interaction sequences of family groups, utilizing unsupervised prototype learning and group variational autoencoders. The algorithm mainly comprises a group encoder and a two-stage interest decoupling module, which includes a group-level interest decoupling module and a member-level interest decoupling module. Specifically, in the group sequence encoder, a Transformer-based group sequence encoder is used to capture the potential temporal dependencies between interactive items in the temporal domain, and an auxiliary loss function is introduced to improve the stability of the group sequence encoder and accelerate convergence. Then, using the two-stage interest decoupling module, group-level and member-level interest decoupling and group consensus modeling are achieved sequentially through end-to-end prototype learning and variational encoders, thereby hierarchically decoupling personalized member interests and group consensus, thus improving the accuracy and robustness of IPTV recommendations.

[0082] The invention includes a group sequence encoder and a two-stage interest decoupling module, comprising a group-level interest decoupling module and a member-level interest decoupling module. In the group sequence encoder, a Transformer-based group sequence encoder is used to capture the potential temporal dependencies between interactive items in the temporal domain. An auxiliary loss function is used to improve the stability and accelerate convergence of the group sequence encoder. The group-level interest decoupling module uses an end-to-end prototype learning method to identify and separate independent member prototypes in an unsupervised manner, thereby decoupling unique and dynamic interest representations of different members from the group representation. To accelerate the convergence of member prototypes during training and enhance the group-level decoupling effect, the invention also introduces a member prototype loss function. The member-level interest decoupling module further decouples the interests of each member within the group through a group variational autoencoder, ensuring that each dimension of the member interest representation is relatively independent and more representative, thus enhancing the interest representation of each member. While achieving member-level interest decoupling, it also models a deep-level group consensus.

[0083] Specifically, this invention decouples the interest representations of different members in a hierarchical manner and mines the group consensus representation by simulating group decision-making to predict the next program that the entire family group will interact with.

[0084] Suppose there is a group of family groups and programs, respectively using and This indicates that each group g has a chronological sequence of program interactions. Among them, |S g | refers to the length of the program interaction sequence for family group g. This refers to the programs that group g interacts with at time t, such as watching a movie or clicking on a product. This paper sets the maximum sequence length to n. If the sequence length is greater than n, the most recent n interactive items are considered; if the sequence length is less than n, padding "0" is added to the left until the total interaction sequence length is n.

[0085] For each family group g, the objective of this invention is to perform group interaction sequences S without any other auxiliary contextual information. g In this case, the recommendation model is trained to generate a probability score for each candidate item in the item set, i.e. in This represents the score of item i. In Top-K recommendation systems, this is typically... After sorting, the items corresponding to the K highest scores are selected as the recommendation results. As shown in formula (1):

[0086]

[0087] in, The next prediction result for group g.

[0088] Here, member prototypes represent members with different interests and characteristics within the entire family group. To ensure that each member prototype fully reflects the uniqueness of different member types and distinguishes different interests and behavioral characteristics in the vector space, distance constraints need to be imposed on the prototypes to maintain a certain minimum distance between them. Family consensus refers to the shared behaviors and preferences that members in a family group agree on in their behavioral choices; it is a shared representation reflecting the overall decision-making tendency of the group. This can be obtained by weighted aggregation of the interest representations of each member.

[0089]

[0090] Where K represents the number of members in the family group, α k These are learnable weights. The symbols and definitions used in this invention are shown in Table 1.

[0091] Table 1 Symbols and Definitions

[0092]

[0093]

[0094] In step S1, since the items in the input group interaction sequence are represented using high-dimensional binary vectors, it is necessary to convert the high-dimensional binary vector representation into a low-dimensional dense representation through an embedding layer. First, the entire item set is embedded, and an item embedding matrix is ​​initialized. Through a learnable positional encoding matrix P∈Rn×d This enhances the input representation of the project sequence. Additionally, Dropout and layer normalization operations are used to prevent overfitting and improve the stability of the training process. Given an input sequence S... g The embedding of the group sequence is initialized as Its formula (3) is shown below:

[0095] e g =Dropout(LayerNorm(m g +p g )),(3)

[0096] Where, m g ∈R n×d This indicates the project embedding, retrieved from the E matrix. g ∈R n×d This represents the positional embedding, and n represents the maximum sequence length. Dropout(·) and LayerNorm(·) are the Dropout operation and the normalization layer, respectively.

[0097] The group interaction sequence includes individual interactions of each member and consensus-building interactions among certain members. To better model group consensus and member interests, it is necessary to capture the potential temporal dependencies between interaction items. This invention employs a group sequence encoder to encode the group's interaction sequence. Currently, group sequence recommendation models (group sequence encoders) typically use deep neural networks to encode user behavior sequences, modeling temporal dependencies and extracting user interest representations. Without loss of generality, a plug-and-play group sequence encoder f is defined. ψ The encoder encodes the group sequence and outputs the group representation H at all times. g ∈R n×d As shown in formula (4):

[0098]

[0099] Here, ψ represents the learnable parameters in the group sequence encoder, where the group represents the temporal dependencies of interactive items and reflects the entangled interests of group members. Specifically, It is the interest representation of the group at time t.

[0100] Since Transformer has demonstrated powerful performance in sequence modeling, this invention uses SASRec, which is purely based on Transformer, as the group encoder. In order to improve the stability of the group sequence encoder and accelerate convergence, this invention uses the cross-entropy loss function (auxiliary loss function) as an auxiliary recommendation task. Its optimization objective can be expressed as finding the optimal parameters that maximize the log-likelihood function of the next prediction of N group sequences at all time steps, as shown in Equation (5).

[0101]

[0102] in, and These represent the embedding of the target item and the embedding of the negative sampled item, respectively, and σ(·) is the sigmoid function.

[0103] In step S2, within the group-level interest decoupling module, the group representation reflects the dynamic, time-dependent interests of the group, including intertwined member interests and group consensus. To fully meet the personalized needs of each group member, it is necessary to decouple the individual member's personalized interest representation from the group representation, thereby clarifying the differences and correlations between members. Furthermore, previous methods for IPTV family groups explicitly designed different member types, such as "elderly," "adults," and "children," which not only requires certain prior knowledge but also limits the ability to represent member interests.

[0104] Therefore, this invention designs a group-level interest decoupling module, aiming to decouple the unique and dynamic interest representations of different members from an unknown group. The interest decoupling module identifies and separates independent member prototypes through an unsupervised end-to-end prototype learning method, and then decouples personalized member interests from the entangled group representation based on these member prototypes, thereby achieving group-level decoupling. Furthermore, to accelerate the convergence of member prototypes during training and enhance the group-level decoupling effect, this invention also introduces a simple and effective member prototype loss function.

[0105] Specifically, the group-level interest decoupling module is mainly implemented through three steps: member prototype activation, attention allocation, and interest decoupling.

[0106] (1) Member Prototype Activation: In order to decouple the interest representations of different members from the group representation, assume there are K learnable member prototypes C = [c1, ..., c2]. K ]∈R K×d Where K is a hyperparameter representing the number of member prototypes. Different member prototypes have low correlation and, in practical IPTV scenarios, represent members who like different types of movies or TV series. Given the group representation H output by the group encoder... g ∈Rn×d Where n is the length of the group interaction sequence and d is the size of the hidden layer dimension. The method for calculating the member prototype weights of a family group at time step i is shown in formula (6):

[0107]

[0108] Where k = 1, 2, ..., K, p k|i Let the group at time step i represent the correlation with the prototype of the k-th member. c represents the group representation at the i-th time step. k LN represents the prototype of the k-th member. i This represents different LayerNorm layers. The embedding representation at each time step is obtained through Equation (6). The corresponding member prototype c was activated. k And assign relevance weights.

[0109] (2) Attention allocation: In addition to needing to calculate p k|i Attention weight p also needs to be considered i To measure the embedding representation at time step i The importance of the next prediction for a group is calculated as shown in formula (7):

[0110]

[0111] Where i = 1, 2, ..., n, W ∈ R d×d ,ρ∈R d as well as All of these are learnable parameters. i This indicates attention weight; query indicates a query. Represents the group representation at the i-th time step; Key; i This indicates residual connection and nonlinearization of the bond; W T d represents the transpose of W; d represents the embedding dimension.

[0112] Because of the lack of supervisory signals regarding which member generated the behavior within the group, this invention uses p i Avoid making the model overly focused on the embedding representation of the last time step in the vector space. Intuitively, the next action of a group is not always generated by the member that produced the interaction at the last time step.

[0113] (3) Interest decoupling: obtain p k|i and p i Next, it is necessary to decouple the member interest representation from the group representation. The calculation method is shown in formula (8):

[0114]

[0115] Where k = 1, 2, ..., K; i = 1, 2, ..., n. Indicates the interests of different members; Let K represent the interest representation of member k at the i-th time step in group g. Ultimately, each group yields K decoupled member interest representations. Although the member prototypes are common across the entire dataset, through interest decoupling, the member interest representations in each group contain the individual characteristics of each member, resulting in strong representational power.

[0116] (4) Learnable member prototype loss function: In order to better train the network and member prototypes, this invention introduces a simple and effective member prototype loss function to enhance the effect of group-level decoupling. The member prototype loss function is shown in formula (9):

[0117]

[0118] Where k = 1, 2, ..., K; i and j represent the i-th and j-th time steps; and Represents the member prototype c i and c j L2 norm normalized representation; Let i represent the L2 norm normalized representation of the group representation at time step i. The training consists of two parts, with the goal of finding neural network parameters that minimize these parameters. Technically, the first part primarily aims to increase the distance between different member prototypes in the vector space, thereby improving their independence and allowing the member interest representations extracted from the group representation using member prototypes to be separated. The second part aims to force the model to learn member prototypes from behavior embeddings, aligning them with the behavior embeddings so that they reside in the same vector space. This design makes the intra-class cluster distribution more compact and guides the network to condense similar behaviors into a single member prototype. Furthermore, the behavior embedding is achieved through p... k|i It was pulled into the prototype of all members instead of the most recent member.

[0119] In step S3, within the member-level interest decoupling module, the member interest representation obtained through the group-level interest decoupling module is an abstract representation of the interests of each member within each family group. Despite group-level decoupling, the correlation between the dimensions of each member's interest representation remains strong; that is, each member's micro-interests are still entangled. For example, in an IPTV family group, a member's interest representation might represent the interests of family members who enjoy watching action movies. Action movies represent this member's relatively macro-level interest, but their more detailed micro-level interests, including their interest in the release year, lead actors, and directors, remain entangled. In the vector space, their interest representation exhibits strong correlation across dimensions, making it impossible to represent a single micro-interest in a single dimension. Furthermore, existing VAEs typically contain only one encoder. This structure cannot distinguish the differences in interests among different members and can only mine common features among different members.

[0120] Therefore, this invention proposes a member-level interest decoupling module. This module further decouples the interests of members within a group using a Group Variational Autoencoder (G-VAE), making the information contained in each dimension of the member interest representation relatively independent and more representative. This allows for the mining of deep-seated personalized features of different types of members, while also modeling group consensus. Specifically, this invention introduces a model architecture from multi-task learning to extend the encoder structure of G-VAE. A set of encoders, including member encoders and a group-shared encoder, is used to further decouple member interests and mine group consensus. Finally, a decoder decodes the latent variables to achieve the next prediction step.

[0121] (1) Member-specific VAE-Encoder: The member encoder aims to learn the personalized features of different members and further decouple member interest representations. This invention treats the decoupling of interests of different members as different tasks, and for each member prototype c k A proprietary variational encoder, M-Encoder, was designed. k The member encoder consists of two learnable blocks and a linear layer. In the first block, the input is processed by a feed-forward layer, increasing the hidden layer dimension to 2d. Layer normalization and dropout operations are then used to prevent overfitting and maintain numerical stability. The second block has the same structure as the first, with a hidden layer dimension of 2d. The outputs of the two blocks are then summed via a residual connection and fed into a linear layer of dimension d to obtain the outputs μ and logσ. 2 Finally, through the reparameter recalculation technique, μ and logσ are...2 As a parameter of a Gaussian distribution, latent variables are sampled from that Gaussian distribution. This membership encoder is able to convert the membership prototype c... k Corresponding interest representation Transform into latent variables That is: a decoupled representation of the interests of members within a group. As shown in formula (10):

[0122]

[0123] Where k = 1, 2, ..., K; Dropout and LayerNorm are the Dropout operation and layer normalization operation, respectively; W i and b i The parameters are learnable, i = 1, 2; Tanh is the activation function; FFN i (·) represents the i-th feedforward neural network.

[0124] (2) Group-shared VAE-Encoder: The group-shared encoder aims to learn the differences between different members of the same group and mine deeper group consensus information. Therefore, the group-shared variational encoder G-Encoder(·) shared The input is the group member interest representation after average pooling, and the output is the group consensus representation. Its structure is the same as that of the member encoder, as shown in formula (11):

[0125]

[0126] in, This represents a group consensus.

[0127] (3) Concatenation: The member encoder and the group shared encoder respectively transform the member interest representation into the latent variables of the member interest decoupled representation and the group consensus representation, namely Z. g and In order to obtain the group representation through decoder reconstruction and use it for the next prediction, the group consensus representation also needs to be concatenated along the hidden layer dimension. Z, a decoupled representation of the interests of group members g ,get The formula is shown in (12):

[0128]

[0129] in, This indicates a group consensus representation. Z, a decoupled representation of the interests of group members g .

[0130] (4) Decoder: This invention utilizes the representation learning and generation capabilities of a variational autoencoder to decode the latent variables of individual member interests and group consensus obtained above, further simulating the mechanism by which individual member interests and group consensus drive behavior. Its structure is similar to that of the encoder, containing only one block and a linear layer; finally, the output O is obtained through this decoder. g ∈R n×d =[o1,o2,…,o n As shown in formula (13):

[0131]

[0132] Unlike previous VAEs, G-VAE uses multiple encoders and one decoder to achieve approximate prior and posterior distributions. To train the group variational encoder G-VAE, the Evidence Lower Bound (ELBO) loss function of this invention is shown in equation (14):

[0133]

[0134] The first term represents the reconstruction loss, and the second term is the Kullback-Leibler divergence. This represents the prior distribution, typically a Gaussian distribution with a mean of 0 and a variance of 1. Let φ represent the variational approximation function of the posterior distribution with parameter λ in the k-th member encoder, i.e., the neural network parameters of the encoder. φ The parameters of the (x|z)| function are the neural network parameters of the decoder. The weight β represents the decoupling learning achieved by constraining the correlation of each dimension by adding a regularization coefficient to the Kullback-Leibler term in the lower evidence bound.

[0135] In step S4, after obtaining the above-mentioned output O g Then, in order to make the next prediction, matrix multiplication is needed to calculate the candidate item scores of the group at the next time step n+1. As shown in formula (15).

[0136]

[0137] Where E is the item embedding matrix. S g This represents the sequence of item interactions in group g; Let P represent the prediction result for the next item in group g, and let P represent the probability score. The higher the score of a candidate item, the higher the probability that it will be interacted with by the group at time n+1.

[0138] This invention utilizes a multi-task training strategy to train the GIMD4PSR model and jointly optimizes the model parameters through multiple loss functions, including an auxiliary recommendation task, a member prototype learning task, and an ELBO task. The auxiliary recommendation task aims to improve the stability of the group sequence encoder and accelerate convergence; the member prototype learning task aims to accelerate the convergence of member prototypes during training and enhance the group-level decoupling effect; and the ELBO optimization task aims to achieve member-level decoupling while simultaneously predicting the next term. In summary, the overall objective of this invention is expressed as shown in formula (16):

[0139]

[0140] Where δ∈[0,1] and γ∈[0,1] are the weight coefficients of the member prototype learning task and the ELBO task, respectively.

[0141] In embodiments of the present invention, to address the problem of entanglement between different member interests and family consensus, personalized member interests and group consensus are hierarchically decoupled from a mixed sequence of behaviors using unsupervised prototype learning and variants of variational autoencoders, thereby improving the accuracy and robustness of IPTV recommendations.

[0142] Corresponding to the above methods, the present invention also provides a program sequence recommendation system based on multi-level decoupling of group interests. Figure 2 The logical structure of a program sequence recommendation system based on multi-level decoupling of group interests according to an embodiment of the present invention is shown.

[0143] like Figure 2 As shown, the program sequence recommendation system based on multi-level decoupling of group interests provided by the present invention includes: a group representation acquisition module 210, a group-level interest decoupling module 220, a member-level interest decoupling module 230, and a prediction module 240. The group representation acquisition module 210 is used to encode the group interaction sequence processed by the embedding layer using a group sequence encoder to obtain a group representation. The group-level interest decoupling module 220 is used to decouple the interest representations of different members from the group representation. The member-level interest decoupling module 230 is used to further decouple the interest representations of different members decoupled by the group-level interest decoupling module to obtain the decoupled representations of different members and the group consensus representation. The prediction module 240 is used to make a prediction on the decoupled representations of different members and the group consensus representation to obtain the candidate item score of the group at the next time step and recommend it to the group.

[0144] As for the embodiments of the program sequence recommendation system based on multi-level decoupling of group interests provided by the present invention, since they are basically similar to the embodiments of the program sequence recommendation method based on multi-level decoupling of group interests, the relevant parts can be found in the description of the method embodiments, and will not be repeated here.

[0145] In embodiments of the present invention, a computer-readable storage medium is also provided, storing a computer program that, when executed by a processor, implements a program sequence recommendation method based on multi-level decoupling of group interests. The specific method is as follows:

[0146] The group representation is obtained by encoding the group interaction sequence processed by the embedding layer using a group sequence encoder.

[0147] The interest representations of different members are decoupled from the group representation through the group-level interest decoupling module;

[0148] The member-level interest decoupling module further decouples the interest representations of different members decoupled by the group-level interest decoupling module to obtain the decoupled representations of different members and the group consensus representation.

[0149] The decoupled representations of the different members and the group consensus representation are used to make the next prediction, obtain the candidate item scores of the group at the next time step, and make recommendations to the group.

[0150] The specific implementation of the computer-readable storage medium of the present invention is largely the same as the specific implementation of the program sequence recommendation method and system based on multi-level decoupling of group interests described above, and will not be repeated here.

[0151] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0152] The sequence numbers of the above embodiments of the present invention are merely for description and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0153] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A program sequence recommendation method based on group interest multi-level decoupling, characterized in that, The method comprises the following steps: encoding the group interaction sequence processed by the embedding layer through a group sequence encoder to obtain a group representation; decoupling the interest representations of different members from the group representation through a group-level interest decoupling module; further decoupling the interest representations of different members decoupled by the group-level interest decoupling module through a group variational autoencoder in a member-level interest decoupling module to obtain decoupled representations of different members and a group consensus representation; predicting the next item of the group based on the decoupled representations of different members and the group consensus representation to obtain candidate item scores of the group at the next time and recommending the candidate item scores to the group; the embedding layer processes the group interaction sequence by using the following formula: wherein, represents an embedding initialization of the group interaction sequence; an item embedding representing the group interaction sequence; a position embedding representing the group interaction sequence, n representing a maximum sequence length, an embedding dimension, a group; and Dropout and normalization layers, respectively; the group-level interest decoupling module decouples the interest representations of different members from the group representation by using the following formula: processing the embedding representation of the group interaction sequence at each time step through a deep model, activating a corresponding member prototype and assigning a relevance weight; obtaining the importance of the embedding representation of the group interaction sequence at each time step to the next item prediction of the group through an attention weight; decoupling the interest representations of different members according to the member prototype and the relevance weight and the importance of the next item prediction of the group. 2.The group interest based multi-level decoupled program sequence recommendation method of claim 1, wherein, The group sequence encoder encodes the group interaction sequence processed by the embedding layer, using the formula: wherein, representing outputting the group representation at all times; representing parameters learnable in said group sequence encoder; representing a group interaction sequence at a time instant of interest; In the process of encoding the group interaction sequence processed by the embedding layer through the group sequence encoder, an auxiliary loss function is used to maximize the log-likelihood function of the next item prediction of the N group interaction sequences at all time steps, and the formula of the auxiliary loss function is as follows: wherein, and denote the embedding of the target item and the embedding of the negative sampled item, respectively, denotes the sigmoid function. 3.The group interest based multi-level decoupled program sequence recommendation method of claim 1, wherein, Prototype weight of a group at a time step is calculated as follows: wherein ; represents the group representation at the th time step and the relevance to the th member prototype; represents the group representation at the th time step, represents the th member prototype; represents different LayerNorm layers.

4. The program sequence recommendation method based on group interest multi-level decoupling according to claim 3, characterized in that, the formula of the attention weight is as follows: wherein, ; , and all represent learnable parameters; denotes attention weights; denotes query; denotes the group representation at the denotes key; denotes residual connection and non-linearity on key; denotes transpose of denotes embedding dimension.

5. The method of claim 4, wherein the group-based interest multi-level decoupled program sequence recommendation method is characterized by, the formula of decoupling the interest representations of different members is as follows: wherein ; representations of interests of different members; representations of interests of members in a group at a moment in time member of interest; In the process of decoupling the interest representations of different members from the group representation through the group-level interest decoupling module, a member prototype loss function is used to optimize the interest representations, and the formula of the member prototype loss function is as follows: wherein, ; and represent the and the time steps; and represent the L2 norm normalized representation of the member prototypes and ; represent the L2 norm normalized representation of the group representation at time step .

6. The program sequence recommendation method based on group interest multi-level decoupling according to claim 5, characterized in that, the group variational autoencoder in the member-level interest decoupling module further decouples the interest representations of different members decoupled by the group-level interest decoupling module to obtain decoupled representations of different members and a group consensus representation, and the formula of the group variational autoencoder is as follows: the member encoder further decouples the interest representations of different members of the same group to obtain decoupled representations of different members; the group shared encoder processes the interest representations between different members of the same group to obtain a group consensus representation; the decoder processes the decoupled representations of different members and the group consensus representation spliced together to obtain decoupled representations of different members and a group consensus representation.

7. The method of claim 6, wherein the group-based interest multi-level decoupled program sequence recommendation method is characterized by, the formula of the member encoder is as follows: wherein denotes decoupled representations of different members; ; and represent a Dropout operation and a layer normalization operation, respectively, and represent learnable parameters, ; Tanh represents an activation function; represents a feed-forward neural network of the th member encoder. 8.The group interest based multi-level decoupled program sequence recommendation method of claim 7, wherein, the formula used by the group shared encoder is as follows: wherein, represents a group consensus representation; represents a group shared variational encoder. 9.The group interest based multi-level decoupled program sequence recommendation method of claim 7, wherein, the formula used by the decoder is as follows: wherein, represents a predicted embedding; represents a feedforward neural network; represents a spliced group consensus representation and decoupled representation of intra-group member interests ; In the process of processing the decoupled representations of different members and the group consensus representation spliced together through the variational autoencoder, an ELBO loss function is used to train the variational autoencoder, and the formula of the ELBO loss function is as follows: wherein, represents a prior distribution; represents a variational approximation function of a posterior distribution of parameters in the kth member encoder; represents a variational approximation function of a posterior distribution of parameters in the decoder; represents a parameter of the function in the decoder; denotes the Kullback-Leibler term regularization coefficient, which constrains the correlation of the dimensions of the member interest representation.

10. The method of claim 9, wherein the group-based interest multi-level decoupled program sequence recommendation method is characterized by, The matrix multiplication is used to obtain the candidate item scores of the group at the next time , using the following formula:​ wherein, denotes an item embedding matrix; denotes a group of item interaction sequences; denotes a group of next item prediction results, denotes a probability score.

11. A program sequence recommendation system based on group interest multi-level decoupling, characterized in that, The method comprises the following steps: a group representation acquisition module is configured to encode a group interaction sequence processed by an embedding layer through a group sequence encoder to obtain a group representation; a group-level interest decoupling module configured to decouple interest representations of different members from the group representation; a member-level interest decoupling module configured to further decouple the interest representations of different members decoupled by the group-level interest decoupling module, obtain decoupled representations of different members and a group consensus representation; a prediction module configured to predict a next item of the group based on the decoupled representations of different members and the group consensus representation, and recommend the next item to the group; the embedding layer processes the group interaction sequence according to the following formula: wherein, represents an embedding initialization of the group interaction sequence; an item embedding representing the group interaction sequence; a position embedding representing the group interaction sequence, n representing a maximum sequence length, an embedding dimension, a group; and Dropout and normalization layers, respectively; the group-level interest decoupling module decouples the interest representations of different members from the group representation, including: the deep model processes the embedding representation of the group interaction sequence at each time step, activates a corresponding member prototype and assigns a relevance weight; the attention weight is used to obtain the importance of the embedding representation of the group interaction sequence at each time step to the next item prediction of the group; the interest representations of different members are decoupled according to the member prototype and the relevance weight, and the importance of the next item prediction of the group.

12. A computer readable storage medium storing a computer program, wherein the computer program comprises program instructions configured to cause a processor to perform the method according to any one of claims 1 to 11. The computer program, when executed by a processor, implements the steps of the program sequence recommendation method based on group interest multi-level decoupling according to any one of claims 1 to 10.

Images