A program recommendation method, system, medium and device for television shared account

Abstract

The present invention proposes a program recommendation method, system, medium and device for a shared television account, which are applied to the technical field of television program recommendation. By constructing the sequential behavior of each account into a graph structure and performing graph transformation, the results of node representation learning are used as sample pairs in the potential user modeling network. Potential users are identified using instance-level contrastive learning and cluster-level contrastive learning in the contrastive clustering user identification network (CCUI‑Net). In order to improve the robustness of the framework, instance-level contrastive learning is applied to item embedding; cluster-level contrastive learning helps cluster similar behaviors into the same user prototype and separate dissimilar behaviors, thereby improving the accuracy of shared account modeling. The attention mechanism is used to empower different users, identify the dynamic influence of different potential users, obtain more accurate account representation, and then realize personalized recommendation of television programs.

Description

Technical Field

[0001] The present invention belongs to the technical field of television program recommendation, and in particular relates to a program recommendation method, system, medium and device for a television shared account. 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] As smart TVs gradually replace cable TV, they have gradually become an important form of entertainment in people's family life. Users can choose to watch a variety of TV programs on smart TVs, which is different from cable TV, which can only play programs according to a fixed schedule. Therefore, a high-performance recommendation system is needed to help users find programs of interest. At present, one of the key challenges of TV program recommendation is that multiple users share an account, that is, the shared account problem. Shared accounts usually contain multiple users, and it is often impossible to obtain the specific identity information of each user in the account, that is, it is impossible to accurately identify the multiple users behind the account. The viewing behavior data generated by shared accounts is mixed and complex, and there is a lack of effective supervision information to distinguish these behaviors. At the same time, different users have different influences on the current state of shared accounts, and at a specific moment, the behavior of only one user may play a dominant role in the recommendation results.

[0004] Existing methods for recommending TV programs for shared accounts include the MISS model, the RL-ISN model, and the TiDA-GCN model. The MISS model uses an attention mechanism to adapt to the needs of session-aware recommendation in the context of shared accounts, and effectively distinguish the behaviors of each user in a shared account in watching TV programs. The RL-ISN model focuses on cross-domain sequence recommendation methods by clustering mixed user behaviors into potential users and using an attention mechanism to identify users. TiDA-GCN constructs a graph structure of users and items within the domain, and introduces a domain-aware graph convolutional network to learn user-specific node representations. In addition, the model combines time interval information and enhances the interactive feature representation between items through an account-aware self-attention mechanism.

[0005] However, the above methods still have the following limitations:

[0006] (1) The above methods mainly model the potential users in shared accounts through direct assumptions, that is, assuming that there are k parallel latent embeddings to simulate the mixed interaction behaviors of users. This lacks certain representativeness and cannot effectively reveal the hidden user prototypes in the accounts.

[0007] (2) The above methods do not consider the robustness of the learned representation. Compared with traditional user behaviors, mixed interactions within an account are more likely to generate noisy data because they involve multiple users.

[0008] (3) The above methods mainly use mean operations to aggregate and achieve account embedding, but do not take into account the important differences between different potential users. Summary of the invention

[0009] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a program recommendation method, system, medium and device for a television shared account, which adopts a comparative clustering user identification network (CCUI-Net) model to learn robust account representation, effectively identifies potential users in a shared account through data enhancement, node representation learning, comparative clustering and account modeling, and can greatly improve the robustness of the model.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] A first aspect of the present invention provides a program recommendation method for a television shared account;

[0012] A program recommendation method for a television shared account, comprising:

[0013] Obtaining the behavior sequence of the shared account and constructing a user-item interaction sequence graph between the user and the TV program, performing graph transformation on the user-item interaction sequence graph by using node dropping and random walk respectively, and obtaining two comparative views;

[0014] Performing node representation learning on the two contrasting views respectively, using a graph convolutional network to capture node sequence information in the two contrasting views, and using the node sequence information to obtain item embedding representations of the two contrasting views;

[0015] Input the item embedding representations of the two contrasting views into the trained CCUI-Net model, and use the attention mechanism to generate personalized TV program recommendation results;

[0016] The CCUI-Net model is used to perform instance-level comparative learning and cluster-level comparative learning on the item embedding representations of the two contrasting views, identify potential users in the shared account and cluster potential users with similar behaviors into the same user model, use the attention mechanism to identify the contribution of different potential users to the shared account behavior, and weighted sum the embedding representations of different potential users to generate personalized TV program recommendation results.

[0017] As a further technical solution, the user-item interaction sequence diagram is shown in FIG. ,picture , where the node set Includes all interacting entities, i.e., users and items, edge sets Represents the sequential interaction relationship between these entities.

[0018] As a further technical solution, the comparison view is specifically:

[0019]

[0020]

[0021] in, and Indicates The node embedding of the layer, is the neighborhood aggregation function; and From the distribution The data augmentation strategy of random sampling in to generate two different comparison views.

[0022] As a further technical solution, the user-item interaction sequence diagram is transformed by node dropping and random walk respectively to obtain two comparative views, wherein the node dropping formula is as follows:

[0023]

[0024]

[0025] in and is the mask vector of the node, and the mask vector takes a value of 0 or 1, where 0 indicates that the node and its edges are removed, and 1 indicates that the node and its edges are retained; and From the distribution Data augmentation strategy of random sampling in ;

[0026] As a further technical solution, the random walk formula is as follows:

[0027]

[0028]

[0029] and is the edge mask vector of each layer, which is independent of the node mask vector. Each element of the edge mask vector determines whether the edge in the corresponding layer is activated, where 0 indicates that the edge is discarded and 1 indicates that the edge is retained; and From the distribution A data augmentation strategy for random sampling in .

[0030] As a further technical solution, the process of using a graph convolutional network to capture the node sequence information in the two contrasting views and using the node sequence information to obtain the item embedding representation of the two contrasting views is as follows:

[0031] Graph convolutional networks are used to learn node embedding representations from contrasting views. In the process of capturing node sequence information, graph convolutional networks update node embeddings through message passing and message aggregation, and generate item embedding representations through the LeakyReLU activation function.

[0032] The message transmission process is that each node transmits its own feature information to the adjacent nodes;

[0033] The process of message aggregation is that after receiving information from adjacent nodes, each node aggregates the information to update the representation of the node itself.

[0034] As a further technical solution, the process of using the CCUI-Net model to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of two contrasting views is as follows:

[0035] In instance-level contrastive learning, different contrastive views represent a behavior of potential users in a shared account. Different views of the same user are used as positive sample pairs, and views of different users are used as negative sample pairs for training. The training aims to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs.

[0036] In cluster-level contrastive learning, the embedded representations of items are divided into different classes. Each class represents a potential user behavior pattern. The embedded representations within the same class are used as positive sample pairs; the embedded representations of different classes are used as negative sample pairs. By maximizing the similarity of sample pairs within a class and minimizing the similarity of sample pairs between classes, the clustering effect is further optimized and similar user behaviors are aggregated into the same user model.

[0037] A second aspect of the present invention provides a program recommendation system for a television shared account.

[0038] A program recommendation system for a television shared account, comprising:

[0039] A graph transformation module is configured to: obtain a behavior sequence of a shared account and construct a user-item interaction sequence graph between a user and a TV program, and perform graph transformation on the user-item interaction sequence graph by node dropping and random walk respectively to obtain two comparison views;

[0040] The node representation learning module is configured to: perform node representation learning on the two contrasting views respectively, use a graph convolutional network to capture node sequence information in the two contrasting views, and use the node sequence information to obtain item embedding representations of the two contrasting views;

[0041] The contrastive learning module is configured to: use the CCUI-Net model to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views, identify potential users in the shared account, and cluster potential users with similar behaviors into the same user model;

[0042] The TV program recommendation module is configured to: use the attention mechanism to identify the contribution of different potential users to the shared account behavior, weighted sum the embedded representations of different potential users, and generate personalized TV program recommendation results.

[0043] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the program recommendation method for a television shared account as described in the first aspect of the present invention.

[0044] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the program recommendation method for a television shared account as described in the first aspect of the present invention are implemented.

[0045] One or more of the above technical solutions have the following beneficial effects:

[0046] (1) The present invention adopts the Contrastive Clustering User Identification Network (CCUI-Net) model to learn robust account representation, maps the mixed behaviors of accounts to potential prototypes, and uses a hierarchical contrastive clustering method to identify potential users. By combining instance-level and cluster-level contrastive learning, similar behaviors can be effectively clustered and different behaviors can be distinguished.

[0047] (2) To improve the robustness of the model, instance-level contrastive learning is applied to item embeddings, which provides a solid foundation for identifying potential users and ensures the stability and reliability of the learned representations.

[0048] 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

[0049] 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.

[0050] Figure 1 This is a flow chart of the method of the first embodiment.

[0051] Figure 2 Schematic diagram of the overall architecture of the CCUI-Net model in the first embodiment.

[0052] Figure 3 It is a system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0053] 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.

[0054] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0055] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0056] The overall idea proposed by the present invention is that the shared account TV program recommendation task aims to provide personalized program recommendations for multiple users using the same account. The core challenge of this task is to distinguish and understand the preferences and behavior patterns of different users, so as to generate accurate and effective recommendations in the scenario of shared accounts. Compared with the traditional sequential recommendation (SR) task, shared account TV program recommendation is more challenging because: (1) an account is used by multiple users, and the interaction behaviors of these users are mixed together without any supervisory signals to separate them. (2) User identity information within the same account is usually not directly accessible, and it is often unknown which users are using this account. Therefore, the goal of the solution provided in this application is no longer to provide TV program recommendations for a single user, but to provide TV program recommendations for a shared account. The following is a detailed description of the solution of this application in conjunction with specific embodiments.

[0057] Embodiment 1

[0058] This embodiment discloses a program recommendation method for a television shared account;

[0059] Combination Figure 1-2 , a program recommendation method for a television shared account, comprising:

[0060] Step S1, obtaining a behavior sequence of a shared account and constructing a user-item interaction sequence diagram between a user and a TV program, performing graph transformation on the user-item interaction sequence diagram by using node dropping and random walk respectively, and obtaining two comparison views;

[0061] Step S2, performing node representation learning on the two obtained comparison views respectively, using a graph convolutional network to capture node sequence information in the two comparison views, and using the node sequence information to obtain item embedding representations of the two comparison views;

[0062] Step S3, input the item embedding representations of the two contrasting views into the trained CCUI-Net model, and use the attention mechanism to generate personalized TV program recommendation results;

[0063] The CCUI-Net model performs instance-level comparative learning and cluster-level comparative learning on the item embedding representations of the two contrasting views, identifies potential users in the shared account and clusters potential users with similar behaviors into the same user model, uses the attention mechanism to identify the contribution of different users to the shared account behavior, and weightedly sums the embedding representations of different users to generate personalized TV program recommendation results.

[0064] In step S1, the behavior sequence of the shared account is obtained and the user-item interaction sequence diagram between the user and the TV program is constructed. First, assume is a hypothetical set of shared accounts, where , indicating that potential users share accounts.

[0065] set up is a collection of items, where items are TV programs watched by the shared account. Indicates that it is the first person who may interact with the shared account item. Assume that the account Can interact with multiple items, can be organized in chronological order The historical sequence behavior of Represented, the user-item interaction sequence diagram between the user and the TV program is constructed, where Yes The first in the behavior sequence items.

[0066] Secondly, the user-item interaction sequence graph is transformed to capture the temporal dynamics of user-item interaction and obtain two comparative views. Among them, in this embodiment, two typical graph transformation techniques are adopted: node dropping (ND) and random walk (RW). Node dropping simulates the potential incompleteness of real-world graph data by randomly removing some nodes in the graph, while random walk explores a wider range of connection patterns in the graph by simulating random paths starting from any node. These two transformation strategies not only enhance the model's understanding of the graph structure, but also improve the model's generalization ability when processing various graph data. Inspired by the SGL data augmentation technology, the two data augmentation strategies of node dropping and random walk are adopted to generate a dual-channel network, in which the graph is an interaction sequence diagram constructed by user-item interaction. Interaction Sequence Diagram , where the node set Includes all interacting entities, including every user and item in the shared account, edge set Represents the sequential interaction relationship between these entities, where the sequential interaction relationship is the interaction behavior between users and items, and these interaction behaviors are arranged in time series. By adopting node dropout and random walk, the sequential behavior of user-item interaction can be captured, thereby creating two unique and enhanced comparative views. The above operations can be uniformly expressed as follows:

[0067]

[0068]

[0069] In this framework, and Indicates The node embedding of the layer is done through the neighborhood aggregation function Update, this function is designed to enhance the function and A variant of the generated graph is used to compute the embedding representation. Here, and From the distribution The data augmentation strategy of random sampling in In this way, different comparative views can be obtained from two different data augmentation methods, thereby improving the diversity and robustness of graph representation.

[0070] For node discarding: According to the principle of node discarding, each node and its connected edges are discarded according to a certain probability. Remove from the graph. This operation can be defined mathematically as:

[0071]

[0072]

[0073] In the formula, and is a mask vector of nodes, which determines which nodes are retained in the generated contrast view. The elements of these vectors are either 0 or 1, where 0 indicates that the corresponding node and its edges are removed in this view, and 1 indicates that they are retained.

[0074] Random walks, on the other hand, are used to enhance the model’s capabilities by distributing various subgraphs in different layers. This approach involves creating different subgraphs for each node and sampling them through random walks. The process of selecting nodes to discard in each layer can be formalized as:

[0075]

[0076]

[0077] In the formula, and From the distribution The data augmentation strategy of random sampling in and is the edge mask vector of each layer, which is independent of the node mask vector and is specially designed for the graph structure of each layer. Each element of these vectors determines whether the edge in the corresponding layer is activated, where 0 means the edge is discarded and 1 means the edge is retained. In this way, different local structures of the graph can be explored in each layer, thereby capturing more diverse information.

[0078] Furthermore, the embedding representations of items and shared accounts are learned by aggregating information from connected nodes. For items, these connections may involve other items and related shared accounts. Similarly, the representation of shared accounts is optimized by aggregating information from connected items, thereby ensuring the generation of comprehensive feature embedding representations that accurately reflect the user's interactive behavior and the characteristics of the item. Therefore, in step S2, a comprehensive embedding representation of the item is constructed by integrating the sequential dependencies between items and the interaction dynamics between shared accounts through node representation learning. Specifically, a graph convolutional network is used to capture the node sequence information in the contrast view, and the node sequence information is used to obtain the embedding representation of the item. This approach not only enhances CCUI-Net's ability to capture complex relationships between items, but also improves the accuracy of predictions and the efficiency of the recommendation system.

[0079] In this embodiment, a graph convolutional network (GNN) is used to capture the node sequence information in the contrast view. In the process of capturing the node sequence information, GNN updates the node embedding through two steps: message passing and message aggregation. In the message passing, each node passes its own feature information to the adjacent nodes. This process simulates the flow of information in the network. Specifically, let and Respectively represent and item The connected item node and account node, where and Influence item The neighboring items and account sets of . On the contrary, let Representatives and Accounts The connected item nodes, where Affects the account Item collection. item arrive The information is transferred as follows:

[0080]

[0081] in Indicates from Pass to Information, parameters Control from arrive The transmission of information, is a trainable weight matrix. and The information interaction between them is determined by the element-wise dot product, i.e. To indicate that and They are item and Embedding vector of . From the account arrive The transfer model is as follows:

[0082]

[0083] Similarly, from item To account The communications are specifically specified as:

[0084]

[0085] In message aggregation, after receiving information from neighboring nodes, each node aggregates this information to update its own representation. Updated by accumulating all incoming communications from item and account nodes.

[0086]

[0087] Likewise, account-specific representations Updated to:

[0088]

[0089] After aggregating the information between items and accounts, a LeakyReLU activation function is used to generate the embedded representation of the item.

[0090] The goal of contrastive learning is to increase the similarity between positive pairs and reduce the similarity between negative pairs. Positive pairs are used to strengthen the representation of similar nodes, that is, nodes of the same user or behavior should be closer in the embedding space; negative pairs are used to distinguish the node representations of different users or behaviors. By setting positive and negative pairs, potential users in shared accounts can be identified and similar behaviors can be clustered into the same user model through the contrastive learning process.

[0091] In step S3, the item embedding representations of the two comparison views are input into the trained CCUI-Net model, and the attention mechanism is used to generate personalized TV program recommendation results; specifically, step S3 also includes:

[0092] In step S301, the CCUI-Net model performs instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views, respectively, identifies potential users in the shared account and clusters potential users with similar behaviors into the same user model.

[0093] Step S302, using the attention mechanism to identify the contribution of users of different user models to the shared account behavior, weighted summing the embedding representations of different users, and generating personalized TV program recommendation results.

[0094] In step S301, the CCUI-Net model performs instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views respectively. Instance-level contrastive learning and cluster-level contrastive learning are key steps that can be used to distinguish different users (or behavior patterns) and identify and cluster potential users.

[0095] At the instance level, the model generates different views through data augmentation techniques (node ​​dropout and random walk), each of which represents a behavior of a potential user in a shared account. For different views of the same user, the model treats them as positive sample pairs; for views of different users, the model treats them as negative sample pairs. That is, in instance-level learning, positive sample pairs are embedded representations of the same user in different views, and negative sample pairs are views of different users. During the training process, learning maximizes the similarity between positive sample pairs and minimizes the similarity between negative sample pairs. The above means can better distinguish the behaviors of different users and cluster similar behaviors together.

[0096] At the clustering level, the model first divides all samples into different classes, each class representing a potential user behavior pattern. Within the same class, samples are considered as positive sample pairs; samples of different classes are considered as negative sample pairs. That is, in cluster-level learning, positive sample pairs are constructed in the same class, and negative sample pairs are constructed between different classes. By maximizing the similarity of sample pairs within a class and minimizing the similarity of sample pairs between classes, the model can further optimize the clustering effect and aggregate similar user behaviors more clearly into the same user model. Among them, in cluster-level contrastive learning, samples refer to the node embedding representations in the user-item interaction graph. These embedding representations are generated based on different user behaviors. By clustering these embedding representations, different categories of user behaviors can be identified and distinguished.

[0097] In this embodiment, for a shared account , the embedded representation of item is mainly used to generate positive sample pairs, which are composed of the embedded representations of nodes of the same user or similar behavior in different views in the node representation learning step. and Negative sample pairs are usually composed of node embedding representations of different users or different behavior patterns, so for shared accounts View and Account View ,in , they constitute a negative sample pair.

[0098] In addition, by using a two-layer nonlinear MLP To refine the enhanced user embedding representation and adapt it to the purpose of contrastive learning. This transformation helps to reduce information loss in the contrastive learning process and improve the distinguishability of features. The specific formula is as follows:

[0099]

[0100]

[0101] In order to reduce the information loss in contrastive learning, a two-layer nonlinear MLP is proposed, following the SimCLR framework and adopting the InfoNCE contrastive loss, whose goal is to enhance the consistency between positive sample pairs while reducing the similarity between negative sample pairs:

[0102]

[0103] in Computes the similarity between two vectors, using cosine similarity as the metric. It is the temperature coefficient in instance-level contrastive learning, which is used to control the smoothness of the similarity distribution. The similarity of sample pairs is measured by cosine distance, that is:

[0104]

[0105] In the cluster-level contrastive learning of this embodiment, the samples are divided into different classes, and then positive sample pairs are constructed within each class, and negative sample pairs are constructed between different classes for training. Specifically, for each class, the samples within the class are regarded as the same category, and positive sample pairs are generated by enhancing these samples. At the same time, samples of different classes are regarded as different categories to form negative sample pairs. This approach enables CCUI-Net to learn and capture feature representations within and between classes. Through cluster-level contrastive learning, data samples are enhanced to create positive sample pairs and negative sample pairs, thereby helping the model to better capture feature representations within and across different classes. The enhanced embedding is performed through a two-layer nonlinear MLP Processing, specifically for feature optimization of contrastive learning tasks. This adjustment helps to reduce information loss and improve the distinguishability of intra-class and inter-class features. The specific formula can be described as follows:

[0106]

[0107] in Represents a collection of classes, and Yes Category Representation under two different data augmentations. Function It is used to calculate the cosine similarity between two vectors, and As temperature coefficients in cluster-level contrastive learning.

[0108] In step S302, the attention mechanism is used to identify the contribution of potential users in different user models to the shared account behavior, and the embedded representations of different potential users are weighted and summed to generate personalized TV program recommendation results.

[0109] Among them, the CCUI-Net model uses the attention mechanism to identify and integrate the contributions of different potential users to the shared account behavior, and calculates the influence of each potential user on the recommendation results in the current context. Then, the CCUI-Net model weights and sums the embedding representations of different potential user models according to these attention weights to obtain an aggregate representation representing the current preferences of the shared account. Using this aggregate representation, the model predicts the next program that users under the shared account are most likely to be interested in and generates personalized recommendation results. Given an account shared by multiple users , by introducing the attention mechanism, we simulate the dynamic contribution of users within an account, and then weigh the contribution of each user based on their interactive behavior.

[0110] Specifically, the attention mechanism is used to calculate the potential user vector in each user model. Assign a weight , and then calculate the representation of the current shared account by the weighted sum of these user vectors The calculation formula of attention weight is as follows:

[0111]

[0112] in is a learnable query vector representing the first The user vectors of potential users are then calculated. Aggregate users It is expressed as follows:

[0113]

[0114] According to the message passing and aggregation mechanism described earlier, the user vector is updated by aggregating messages from item embedding representations and other related users. :

[0115]

[0116] in Represents a slave node To users The message delivered.

[0117] A prediction layer is built on top of the final representation to predict the user Will item be adopted? . Fast retrieval by using the inner product:

[0118]

[0119] in and Respectively represent the corresponding users and item The latent vector of .

[0120] In the CCUI-Net model, supervised learning is used for parameter optimization. The supervision information comes from the observed interactions, which are represented by graphs. The widely used pairwise BPR loss is adopted. This method prioritizes giving higher prediction scores to observed interactions and lower scores to unobserved interactions:

[0121]

[0122] in is the training data, is the set of observed interactions, is the set of unobserved interactions.

[0123] Therefore, the loss function LOSS of the final model can be described as:

[0124]

[0125] in and Parameters controlling the strength of instance-level and cluster-level contrastive learning.

[0126] In this embodiment, the CCUI-Net method in the shared account sequence recommendation task is also evaluated on two real-world datasets HVIDEO and HAMAZON.

[0127] The HVIDEO dataset collects household account logs from a smart TV platform from October 2016 to June 2017, covering a variety of content such as educational videos and on-demand movies. The dataset contains educational videos designed for different categories of students and various on-demand movies such as cartoons and TV series. Considering that household accounts are usually shared by family members, the mixed interaction behaviors within the family are captured through this data, making it very suitable for shared account sequence recommendation tasks. To improve data quality, account records with less than 10 views and less than 300 seconds of viewing time are excluded. This filtering process produces a high-quality dataset consisting of 13,715 household accounts and 143,885 viewing sequences.

[0128] The HAMAZON dataset contains Amazon users' review records in two categories: movies (domain M) and books (domain B), spanning from May 1996 to July 2014. Domain M records users' viewing records and reviews of movies, while domain B involves users' reading and review activities on books. Due to the lack of original Amazon data for shared account cross-domain sequential recommendation (SCSR), domain and account information need to be preprocessed to meet the requirements. According to the method described in PSJNet, we performed the following preprocessing steps: Because users can interact with items in multiple domains at the same time, we retain users who have interaction behaviors in both domains. Next, we simulate the shared account feature by dividing the timeline into six time periods (1996-2000, 2001-2003, 2004-2006, 2007-2009, 2010-2012, 2013-2015). In each specified time period, we merge 2-4 users and their interaction behaviors into a unified account as the shared account of these users. By breaking down the data for each year into smaller sequences and excluding sequences with less than five items in the M domain and less than two items in the B domain, a high-quality dataset was generated. The dataset includes 13,724 accounts and 143,885 sequences. The detailed information of the HVIDEO and HAMAZON datasets is shown in Table 1.

[0129] Table 1. Details of HVIDEO and HAMAZON datasets

[0130]

[0131] In the above evaluation process, 85% of all sequences in the HVIDEO and HAMAZON datasets were used as training datasets, and the remaining 15% of the sequences were used as test datasets. In order to evaluate the model, the last item of each sequence in each domain was taken as the true value to evaluate the ability of the CCUI-Net model to predict the next item using the historical behavior sequence of a single account.

[0132] The performance of CCUI-Net is evaluated by comparing with six comparative models, including the following methods: traditional recommendation method, graph contrastive learning (GCL)-based recommendation method, cross-domain recommendation (CR) method, sequential recommendation (SR) method, time interval sequential recommendation method and shared account cross-domain sequential recommendation (SCSR) method.

[0133] CCUI-Net is implemented in PyTorch and accelerated by Quadro RTX 6000 GPU. Adam is used as the optimization technique for the composite loss function. In the training configuration, the batch size is set to 256, the learning rate is 0.001, and the dropout rate is 0.1. In the auxiliary self-supervised learning task, the regularization parameter ssl-reg is set to 1e -5 , the embedding size is adjusted to 16. The best parameter configurations recorded in the compared models are adopted, and these settings are further optimized on both datasets.

[0134] The corresponding hyperparameters are set and adjusted for various datasets as needed. Specifically: (1) For LightGCL, the configuration includes two layers and an embedding size of 16. (2) For XSimGCL, the number of layers is configured to be 2 and the embedding size is set to 16. (2) For the two variants of SGL, namely SGL-ND and SGL-RW, the configuration is three layers and an embedding size of 16. (3) For the SASRec model, the embedding size is adjusted to 16. (4) For the rest of the benchmark settings, TiDA-GCN is followed and the same hyperparameter values ​​are used for evaluation. The specific structure is shown in Table 2.

[0135] Table 2. Comparison results on HVIDEO and HAMAZON datasets

[0136]

[0137] Table 2 shows the comparison results of CCUI-Net method with the comparison models on HVIDEO and HAMAZON datasets. Based on these data, the following conclusions can be drawn:

[0138] 1. CCUI-Net outperforms other comparison models except DA-GCN on two real-world datasets (four single domains). This shows that the CCUI-Net method can effectively solve the shared account problem in the sequential recommendation scenario, and CCUI-Net successfully captures the shared account characteristics and the sequential behavior of users.

[0139] 2. The performance of CCUI-Net exceeds that of RNNs-based methods, including GRU4REC, HGRU4REC, -Net and PSJNet. This highlights the advantage of using graph convolutional network (GCN) technology to capture diverse interactions between users and items.

[0140] 3. The CCUI-Net method outperforms traditional sequential recommendation models such as SASRec, GRU4REC, HGRU4REC, NAIS, Time-LSTM, and TGSRec, demonstrating the significant advantages of contrastive clustering in predicting user behavior sequences. It highlights the advantages of analyzing accounts from the perspective of the overall user group. In addition, by analyzing user interaction sequences, this scheme achieves better results.

[0141] Embodiment 2

[0142] This embodiment discloses a program recommendation system for a television shared account;

[0143] like Figure 3 As shown, a program recommendation system for a television shared account includes:

[0144] A graph transformation module is configured to: obtain a behavior sequence of a shared account and construct a user-item interaction sequence graph between a user and a TV program, and perform graph transformation on the user-item interaction sequence graph by node dropping and random walk respectively to obtain two comparison views;

[0145] The node representation learning module is configured to: perform node representation learning on the two obtained comparison views respectively, use a graph convolutional network to capture node sequence information in the two comparison views, and use the node sequence information to obtain item embedding representations of the two comparison views;

[0146] The contrastive learning module is configured to: use the CCUI-Net model to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views, identify potential users in the shared account, and cluster potential users with similar behaviors into the same user model;

[0147] The TV program recommendation module is configured to: use the attention mechanism to identify the contribution of different potential users to the shared account behavior, weighted sum the embedded representations of different potential users, and generate personalized TV program recommendation results.

[0148] Embodiment 3

[0149] The purpose of this embodiment is to provide a computer-readable storage medium.

[0150] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a program recommendation method for a television shared account as described in Embodiment 1 of the present disclosure.

[0151] Embodiment 4

[0152] The purpose of this embodiment is to provide an electronic device.

[0153] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the program recommendation method for a television shared account as described in Embodiment 1 of the present disclosure are implemented.

[0154] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0155] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0156] 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 program recommendation method for a television shared account, characterized in that: include: Get the behavior sequence of the shared account and construct the user-item interaction sequence diagram between the user and the TV program; specifically, first assume is a shared account set, where , indicating that potential users share Accounts; is a collection of items, where items are TV programs watched by the shared account. Indicates the first interaction with the shared account. item; due to account Interact with multiple items, organized in chronological order The historical sequence behavior of Representation, thereby constructing a user-item interaction sequence diagram between the user and the TV program, where Yes The first in the behavior sequence item; The user-item interaction sequence graph is transformed by node dropping and random walking respectively to obtain two comparative views; wherein the random walking includes creating different subgraphs for each node and sampling through random walking; the formula is as follows: and is the edge mask vector of each layer, which is independent of the node mask vector. Each element of the edge mask vector determines whether the edge in the corresponding layer is activated, where 0 indicates that the edge is discarded and 1 indicates that the edge is retained; and From the distribution Data augmentation strategy of random sampling in ; A node set representing an interaction sequence diagram, which includes all the entities interacting in the interaction sequence diagram; The edge set representing the interaction sequence graph includes the sequence interaction relationships between all interacting entities; Node representation learning is performed on the two obtained comparison views respectively, and a graph convolutional network is used to capture node sequence information in the two comparison views, and the item embedding representation of the two comparison views is obtained by using the node sequence information; specifically, the graph convolutional network is used to learn the node embedding representation from the two comparison views. In the process of capturing the node sequence information, the graph convolutional network updates the node embedding through message passing and message aggregation, and generates the item embedding representation through the LeakyReLU activation function; The message transmission process is that each node transmits its own feature information to the adjacent nodes; The message aggregation process is that after receiving information from adjacent nodes, each node aggregates the information to update its own representation; Input the item embedding representations of the two contrasting views into the trained CCUI-Net model, and use the attention mechanism to generate personalized TV program recommendation results; The CCUI-Net model is used to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views, identify potential users in the shared account and cluster potential users with similar behaviors into the same user model, use the attention mechanism to identify the contribution of different potential users to the shared account behavior, and weighted sum the embedding representations of different potential users to generate personalized TV program recommendation results. Specifically, given an account shared by multiple users , by introducing the attention mechanism to simulate the dynamic contribution of users in an account, and then weighing the contribution of each user according to the user's interactive behavior; Specifically, the attention mechanism is used to calculate the potential user vector in each user model. Assign a weight , and then calculate the representation of the current shared account by the weighted sum of these user vectors ; The calculation formula of attention weight is as follows: in is a learnable query vector representing the first The user vectors of potential users; then calculate the shared accounts Aggregate users It is expressed as follows: Update the user vector by aggregating messages from item embeddings and other related users : in Represents a slave node To users The message delivered; is the activation function.

2. A program recommendation method for a television shared account as claimed in claim 1, characterized in that: The user-item interaction sequence diagram is shown in Figure ,picture , where the node set Includes all interacting entities, i.e., users and items, edge sets Represents the sequential interaction relationship between these entities.

3. A program recommendation method for a television shared account as claimed in claim 1, characterized in that: The two comparison views are specifically: in, and Indicates The node embedding of the layer, is the neighborhood aggregation function; and From the distribution The data augmentation strategy of random sampling in to generate two different comparison views.

4. A program recommendation method for a television shared account as claimed in claim 1, characterized in that: The user-item interaction sequence graph is transformed by node dropping and random walking respectively to obtain two comparative views, wherein the node dropping formula is as follows: in and is the mask vector of the node, and the mask vector takes a value of 0 or 1, where 0 indicates that the node and its edges are removed, and 1 indicates that the node and its edges are retained; and From the distribution Data augmentation strategy of random sampling in ; A node set representing an interaction sequence diagram, which includes all the entities interacting in the interaction sequence diagram; Represents the edge set of the interaction sequence graph, which includes the sequence interaction relationships between all interacting entities.

5. A program recommendation method for a television shared account as claimed in claim 1, characterized in that: The process of using the CCUI-Net model to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of two contrasting views is as follows: In instance-level contrastive learning, different contrastive views represent a behavior of potential users in a shared account. Different views of the same user are used as positive sample pairs, and views of different users are used as negative sample pairs for training. The training aims to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs. In cluster-level contrastive learning, the embedded representations of items are divided into different classes. Each class represents a potential user behavior pattern. The embedded representations within the same class are used as positive sample pairs; the embedded representations of different classes are used as negative sample pairs. By maximizing the similarity of sample pairs within a class and minimizing the similarity of sample pairs between classes, the clustering effect is further optimized and similar user behaviors are aggregated into the same user model.

6. A program recommendation system for TV shared accounts, characterized in that: include: The graph transformation module is configured to: obtain the behavior sequence of the shared account and construct a user-item interaction sequence diagram between the user and the TV program; specifically, first set is a shared account set, where , indicating that potential users share Accounts; is a collection of items, where items are TV programs watched by the shared account. Indicates the first interaction with the shared account. item; due to account Interact with multiple items, organized in chronological order The historical sequence behavior of Representation, thereby constructing a user-item interaction sequence diagram between the user and the TV program, where Yes The first in the behavior sequence item; The user-item interaction sequence graph is transformed by node dropping and random walking respectively to obtain two comparative views; wherein the random walking includes creating different subgraphs for each node and sampling through random walking; the formula is as follows: and is the edge mask vector of each layer, which is independent of the node mask vector. Each element of the edge mask vector determines whether the edge in the corresponding layer is activated, where 0 indicates that the edge is discarded and 1 indicates that the edge is retained; and From the distribution Data augmentation strategy of random sampling in ; A node set representing an interaction sequence diagram, which includes all the entities interacting in the interaction sequence diagram; The edge set representing the interaction sequence graph includes the sequence interaction relationships between all interacting entities; The node representation learning module is configured to: perform node representation learning on the two obtained comparison views respectively, use a graph convolutional network to capture node sequence information in the two comparison views, and use the node sequence information to obtain item embedding representations of the two comparison views; specifically, use the graph convolutional network to learn node embedding representations from the two comparison views, and in the process of capturing node sequence information, the graph convolutional network updates the node embedding through message passing and message aggregation, and generates the item embedding representation through the LeakyReLU activation function; The message transmission process is that each node transmits its own feature information to the adjacent nodes; The message aggregation process is that after receiving information from adjacent nodes, each node aggregates the information to update its own representation; The contrastive learning module is configured to: use the CCUI-Net model to perform instance-level contrastive learning and cluster-level contrastive learning on the item embedding representations of the two contrasting views, identify potential users in the shared account, and cluster potential users with similar behaviors into the same user model; The TV program recommendation module is configured to: use the attention mechanism to identify the contribution of different potential users to the shared account behavior, weighted sum the embedding representations of different potential users, and generate personalized TV program recommendation results; specifically, given an account shared by multiple users , by introducing the attention mechanism to simulate the dynamic contribution of users in an account, and then weighing the contribution of each user according to the user's interactive behavior; Specifically, the attention mechanism is used to calculate the potential user vector in each user model. Assign a weight , and then calculate the representation of the current shared account by the weighted sum of these user vectors ; The calculation formula of attention weight is as follows: in is a learnable query vector representing the first The user vectors of potential users; then calculate the shared accounts Aggregate users It is expressed as follows: Update the user vector by aggregating messages from item embeddings and other related users : in Represents a slave node To users The message delivered; is the activation function.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the program recommendation method for a television shared account as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the program recommendation method for a television shared account as described in any one of claims 1 to 5 are implemented.

Images