A group sequence recommendation method and system considering long-term and short-term interests of users

Abstract

The application discloses a group sequence recommendation method and system considering long-term and short-term interests of users, and belongs to the technical field of deep learning. Most of the existing group recommendation methods are interaction-driven, often ignoring the multi-attribute rating data of users and the dynamic preferences of users over time, and when fusing the preferences of users in groups, the correlation between auxiliary interaction information and its influence on group preferences are not fully considered, so that the performance of the recommendation system is limited. Based on this, the group sequence recommendation method considering long-term and short-term interests of users is proposed, the overall rating, multi-attribute rating and corresponding long-term and short-term sequence information of users in groups are comprehensively utilized, the overall preference and long-term and short-term preferences of groups are mined, and the group representation is optimized and aggregated in combination with contrast learning and an attention mechanism, so that the effect of the group recommendation system is effectively enhanced, group users can better make decision selection, and user satisfaction is improved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to a group sequence recommendation method and system that considers users' short-term and long-term interests. Background Technology

[0002] Group recommendation systems aim to provide users with items that meet their needs by analyzing user-item interactions. Sequential recommendation methods, by considering the time factor, can more accurately capture user preferences and generate more personalized recommendation lists. In sequential recommendation methods, user preferences can be divided into two main categories: long-term preferences and short-term preferences. Long-term preferences are generally obtained by mining users' global behavioral sequences, reflecting stable interests over a long period, while short-term preferences are obtained by mining users' short-term interaction sequences, reflecting users' real-time interests that change over time. By fusing long-term and short-term preferences, user preferences can be modeled more comprehensively and accurately. To further improve recommendation performance, contrastive learning is introduced into sequential recommendation methods to optimize user representation.

[0003] However, current research on sequence recommendation methods typically relies on overall user ratings to make recommendations to individual users, neglecting the multi-attribute rating sequences of users and the needs of user groups. Furthermore, most existing contrastive learning methods utilize data augmentation techniques to split the original sequence into multi-view sequences, which may disrupt the inherent patterns of the original sequence, leading to poor recommendation results. Simultaneously, existing group recommendation methods are generally interaction-driven, often ignoring users' multi-attribute rating data and their dynamic preferences evolving over time, thus failing to provide accurate recommendation services for user groups. Summary of the Invention

[0004] The purpose of this invention is to solve the above problems by proposing a group sequence recommendation method and system that considers users' long-term and short-term interests. Based on a comprehensive consideration of users' overall rating, multi-attribute rating and time-series information, the method performs a comprehensive and detailed model of group preferences. By constructing long-term and short-term contrastive learning, the method maximizes the consistency of group users on the overall rating sequence and multi-attribute rating sequence. This allows the method to learn and optimize the common preferences of group users while maintaining the integrity of the original sequence, thereby effectively enhancing the group recommendation effect.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A group sequence recommendation method that considers users' short-term and long-term interests includes the following steps:

[0007] Step 1: Obtain the initial embedding representation of the group by aggregating the ID embeddings of users within the group;

[0008] Step 2: Construct a group sequence item graph by integrating the overall rating sequences of users within the group, and use a graph convolutional network to capture multi-hop information of users within the group to generate the overall group preference;

[0009] Step 3: Use attention mechanisms to analyze and measure the relationship between the overall rating sequence and the multi-attribute rating sequence of users within the group, and model the group's long-term and short-term preferences;

[0010] Step 4: Integrate the overall group preference and the short-term and long-term preferences to form the group comprehensive preference, and obtain the predicted score of the item by calculating its inner product with the candidate items;

[0011] Step 5: Establish contrastive learning by maximizing the consistency between the long-term and short-term preferences of the overall user ratings and the long-term and short-term preferences of the multi-attribute ratings within the group, and then optimize the method using the total loss function obtained by mixing the recommendation loss function and the contrastive loss function.

[0012] Furthermore, the implementation method of step 1 is as follows:

[0013] Use Embedding encoding for group g l The user IDs within the group are initially embedded and represented, and the group g is used as the initial embedding representation. l The user vector within is represented as group g l The user vectors within the group are concatenated and their dimensionality reduced to obtain the group g. l initial embedding representation

[0014] Furthermore, the implementation method of step 2 is as follows:

[0015] The learning process of the overall group preference is achieved by integrating group g. l The interaction sequences of internal users are used to construct a group sequence item graph. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix of the sequence item graph is established. The specific calculation formula is as follows:

[0016]

[0017] To delve deeper into the characteristics of nodes themselves and learn more accurate node features, Add an identity matrix and perform normalization; the specific calculation formula is as follows:

[0018]

[0019] in, D is the normalized adjacency matrix of the sequence item graph. -1 / 2 Let I be a diagonal matrix. E It is the identity matrix;

[0020] Furthermore, the initial embedding representation of the group is obtained through a graph convolutional network. Perform a linear transformation to obtain the overall group preference. The calculation formula is as follows:

[0021]

[0022] in, For the overall preference of the group at level t-1, For group g l The initial embedding representation, W t Let be the trainable parameter matrix of the t-th layer.

[0023] Furthermore, the implementation method of step 3 is as follows:

[0024] The user's interaction records within one month are analyzed as the user's short-term sequence, and the user's interaction records within one year are analyzed as the long-term sequence;

[0025] To learn the long-term preferences of a group, based on in-depth mining of the overall user rating sequence and multi-attribute rating sequence, an attention mechanism is used to learn and aggregate the long-term preference features of users within the group, resulting in group g. l Long-term preference in the overall score sequence and long-term preferences on multi-attribute rating sequences

[0026] Group g l Long-term preference in the overall score sequence The calculation formula is as follows:

[0027]

[0028]

[0029] in, Indicates user Long-term sequence information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix;

[0030] Group g l Long-term preferences on multi-attribute rating sequences The calculation formula is as follows:

[0031]

[0032] in, Indicates user In attribute a v Long-term sequence information on Indicates user In attribute a v Weight on, Represents group g l In attribute a v long-term preference, W long The weight matrix represents the influence of long-term group preferences, b. long This represents the long-term preference bias vector of the group;

[0033] In-depth analysis of recent interaction sequences of users within a group can effectively capture the group's current needs and ensure that recommended items match the group's current interests;

[0034] By comprehensively using the overall user rating sequence and multi-attribute rating sequences, short-term preferences of users within a group are learned, and user weights are obtained through an attention mechanism, thus comprehensively modeling the decision-making process and obtaining the group g. l Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences

[0035] Group g l Short-term preference in the overall score sequence The calculation formula is as follows:

[0036]

[0037]

[0038] in, Indicates user Short-term series information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix;

[0039] Group g l Short-term preferences on multi-attribute rating sequences The calculation formula is as follows:

[0040]

[0041] in, Indicates user In attribute a v Short-term sequence information, Indicates user In attribute a v Weight on, Represents group g l In attribute a v Short-term preferences, W short This represents the group short-term preference influence matrix, b short This represents the short-term preference bias vector of the group.

[0042] Furthermore, the method for generating the predicted score in step 4 is as follows:

[0043] Initial embedding representation of the group Overall group preferences Long-term preference of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences The weighted summation yields the group's overall preference g. l The specific calculation formula is as follows:

[0044]

[0045] Group overall preference g l With candidate project i j Perform inner product operation to obtain group g. l For candidate project i j Predicted score The specific calculation formula is as follows:

[0046]

[0047] Furthermore, the specific optimization method in step 5 is as follows:

[0048] After obtaining the predicted score, the total loss function is used. Optimize the method;

[0049] The total loss function consists of two parts: the recommendation loss function. and contrastive learning loss function

[0050] The recommendation loss function The calculation formula is as follows:

[0051]

[0052] Where M represents the number of candidate items in the group, y lj Represents group g in the training set l For project i j The actual rating, G represents l to i j Predicted score;

[0053] The average rating of a project by users within a group is used in the recommendation loss function to represent the actual rating of the group for the project.

[0054] To obtain high-quality group embedding representations, a contrastive learning approach is adopted to maximize the consistency between the long-term and short-term preferences of the overall user ratings and the long-term and short-term preferences of the multi-attribute ratings within the group. The higher the similarity, the better the recommendation effect.

[0055] The contrastive learning loss function The calculation formula is as follows:

[0056]

[0057] Where cos() is the cosine similarity function. For group g l Similarity between long-term preferences of the overall rating sequence and long-term preferences of multi-attribute rating sequences. For group g l The similarity between the short-term preferences of the overall rating sequence and the short-term preferences of the multi-attribute rating sequence, where τ is the coefficient;

[0058] By minimizing the total loss function To optimize the method:

[0059]

[0060] Where λ is the coefficient and π is the regularization weight of the learning parameter set Θ.

[0061] A group sequence recommendation system that considers users' short-term and long-term interests includes the following modules:

[0062] Group initial embedding representation module: using Embedding encoding for group g l The initial embedding of the user ID is used for representation, and the group g l The user vector within is represented as The user vectors within the group are concatenated and their dimensionality reduced to obtain the group g. l initial embedding representation

[0063] Group Overall Preference Learning Module: Integrating Group g lBased on the interaction sequences of internal users, a group sequence item graph is constructed. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix is ​​established for the sequence item graph. right After adding an identity matrix and normalizing, the initial embedding representation of the group is processed by a graph convolutional network. Perform a linear transformation to obtain the overall group preference.

[0064] The group long-short-term preference learning module utilizes an attention mechanism to learn the long-term preference features of users within a group from both the overall user rating sequence and the multi-attribute rating sequence, and then aggregates these features to obtain the group g. l Long-term preference in the overall score sequence and its long-term preference on multi-attribute rating sequences. Simultaneously, by leveraging attention mechanisms to acquire user weights, the decision-making process can be modeled more comprehensively, resulting in group g. l Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences

[0065] Predictive scoring module: Initializes group embedding representation Overall group preferences Long-term preference of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences We perform a weighted summation to obtain the group's overall preference g. l By analyzing the overall preferences of the group g l With candidate project i j Perform inner product operation to obtain group g. l For candidate project i j Predicted score

[0066] Method optimization module: After obtaining the predicted score, the method is optimized using the total loss function obtained by mixing the recommendation loss function and the contrastive learning loss function;

[0067] Recommendation list generation module: The system generates a corresponding project recommendation list based on the comprehensive embedding representation of the group to be recommended.

[0068] Compared with the prior art, the beneficial effects of the present invention are:

[0069] 1. This invention can more comprehensively and accurately model the common preferences of group users, provide the group with a list of recommended items that meet their needs, and thus help group users make decisions faster and better, improve user satisfaction and experience, and effectively enhance the effect of group recommendations;

[0070] 2. This invention can be used in various network applications such as e-commerce, social media, online education, and online healthcare to improve the quality and efficiency of service networks, better meet the intelligent, group-oriented, and diversified real-world needs of the big data era, and further enhance the economic and market value of recommendation systems. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the method recommended by the present invention;

[0072] Figure 2 This is a schematic diagram of the framework of the method recommended by the present invention;

[0073] Figure 3 This is a schematic diagram of the recommended system of the present invention. Detailed Implementation

[0074] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0075] like Figure 1-3 As shown, a group sequence recommendation method that considers users' short-term and long-term interests includes the following steps:

[0076] Step 1: Obtain the initial embedding representation of the group by aggregating the ID embeddings of users within the group. The implementation method is as follows:

[0077] This invention uses Embedding encoding for group g l The user IDs within the group are initially embedded and represented, and the group g is used as the initial embedding representation. l The user vector within is represented as group g l The user vectors within the group are concatenated and their dimensionality reduced to obtain the group g. l initial embedding representation

[0078] Step 2: Construct a group sequence item graph by integrating the overall rating sequences of users within the group, and use a graph convolutional network to capture multi-hop information of users within the group to generate the overall group preference. The implementation method is as follows:

[0079] The learning process of the overall group preference is achieved by integrating group g. l The interaction sequences of internal users are used to construct a group sequence item graph. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix of the sequence item graph is established. The specific calculation formula is as follows:

[0080]

[0081] To delve deeper into the characteristics of nodes themselves and learn more accurate node features, Add an identity matrix and perform normalization; the specific calculation formula is as follows:

[0082]

[0083] in, D is the normalized adjacency matrix of the sequence item graph. -1 / 2 Let I be a diagonal matrix. E It is the identity matrix;

[0084] Furthermore, the initial embedding representation of the group is obtained through a graph convolutional network. Perform a linear transformation to obtain the overall group preference. The calculation formula is as follows:

[0085]

[0086] in, For the overall preference of the group at level t-1, For group g l The initial embedding representation, W t Let be the trainable parameter matrix of the t-th layer.

[0087] Step 3: Analyze and measure the relationship between the overall rating sequence and multi-attribute rating sequences of users within the group using attention mechanisms, and model the group's long-term and short-term preferences. The implementation method is as follows:

[0088] Long-term sequences typically involve more persistent user preferences and habits, generally recording user behavior over several months or even years, while short-term sequences reflect recent user behavior trends and interests, generally recording user behavior over the last few days, weeks, or month.

[0089] This invention analyzes a user's interaction records within one month as a short-term sequence and a user's interaction records within one year as a long-term sequence.

[0090] To learn the long-term preferences of a group, based on in-depth mining of the overall user rating sequence and multi-attribute rating sequence, an attention mechanism is used to learn and aggregate the long-term preference features of users within the group, resulting in group g. l Long-term preference in the overall score sequence and long-term preferences on multi-attribute rating sequences

[0091] Group g l Long-term preference in the overall score sequence The calculation formula is as follows:

[0092]

[0093] in, Indicates user Long-term sequence information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix;

[0094] Group g l Long-term preferences on multi-attribute rating sequences The calculation formula is as follows:

[0095]

[0096] in, Indicates user In attribute a v Long-term sequence information on Indicates user In attribute a v Weight on, Represents group g l In attribute a v long-term preference, W long The weight matrix represents the influence of long-term group preferences, b. long This represents the long-term preference bias vector of the group;

[0097] In-depth analysis of recent interaction sequences among users within a group can effectively capture the group's current needs, ensuring that recommended items align with the group's current interests. This invention comprehensively utilizes overall user rating sequences and multi-attribute rating sequences to learn the short-term preferences of users within a group, and obtains user weights through an attention mechanism, thereby more comprehensively modeling the decision-making process and obtaining the group g. l Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences

[0098] Group g l Short-term preference in the overall score sequence The calculation formula is as follows:

[0099]

[0100] in, Indicates user Short-term series information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix;

[0101] Group g l Short-term preferences on multi-attribute rating sequences The calculation formula is as follows:

[0102]

[0103] in, Indicates user In attribute a v Short-term sequence information, Indicates user In attribute a v Weight on, Represents group g l In attribute a v Short-term preferences, W short This represents the group short-term preference influence matrix, b short This represents the short-term preference bias vector of the group.

[0104] Step 4: Integrate the overall group preference and short-term preferences to form a comprehensive group preference. Calculate the inner product of this comprehensive preference with the candidate items to obtain the predicted item score. The implementation method is as follows:

[0105] Initial embedding representation of the group Overall group preferences Long-term preference of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences The weighted summation yields the group's overall preference g. l The specific calculation formula is as follows:

[0106]

[0107] Group overall preference g l With candidate project i j Perform inner product operation to obtain group g. lFor candidate project i j Predicted score The specific calculation formula is as follows:

[0108]

[0109] Step 5: Establish a contrastive learning model by maximizing the consistency between the long-term and short-term preferences of the overall user ratings and the long-term and short-term preferences of the multi-attribute ratings within the group. Optimize the method using the total loss function obtained by combining the recommendation loss function and the contrastive loss function. The specific optimization method is as follows:

[0110] After obtaining the predicted score, the total loss function is used. Optimize the method;

[0111] The total loss function consists of two parts: the recommendation loss function. and contrastive learning loss function

[0112] The recommendation loss function The calculation formula is as follows:

[0113]

[0114] Where M represents the number of candidate items in the group, y lj Represents group g in the training set l For project i j The actual rating, G represents l to i j Predicted score;

[0115] The average rating of a project by users within a group is used in the recommendation loss function to represent the actual rating of the group for the project.

[0116] Since the groups in the dataset are obtained through clustering, the actual ratings of the groups for the items do not exist. Therefore, the average rating of the items by users within the group is used in the recommendation loss function to represent the actual ratings of the group for the items.

[0117] To obtain high-quality group embedding representations, we adopt a contrastive learning approach that maximizes the consistency between the long-term and short-term preferences of the overall user ratings and the long-term and short-term preferences of the multi-attribute ratings within the group. The higher the similarity, the better the recommendation effect.

[0118] The contrastive learning loss function The calculation formula is as follows:

[0119]

[0120] Where cos() is the cosine similarity function. For group gl Similarity between long-term preferences of the overall rating sequence and long-term preferences of multi-attribute rating sequences. For group g l The similarity between the short-term preferences of the overall rating sequence and the short-term preferences of the multi-attribute rating sequence, where τ is the coefficient;

[0121] By minimizing the total loss function To optimize the method:

[0122]

[0123] Where λ is the coefficient and π is the regularization weight of the learning parameter set Θ.

[0124] A group sequence recommendation system that considers users' short-term and long-term interests includes the following modules:

[0125] Group initial embedding representation module: using Embedding encoding for group g l The initial embedding of the user ID is used for representation, and the group g l The user vector within is represented as The user vectors within the group are concatenated and their dimensionality reduced to obtain the group g. l initial embedding representation

[0126] Group Overall Preference Learning Module: Integrating Group g l Based on the interaction sequences of internal users, a group sequence item graph is constructed. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix is ​​established for the sequence item graph. right After adding an identity matrix and normalizing, the initial embedding representation of the group is processed by a graph convolutional network. Perform a linear transformation to obtain the overall group preference.

[0127] The group long-short-term preference learning module utilizes an attention mechanism to learn the long-term preference features of users within a group from both the overall user rating sequence and the multi-attribute rating sequence, and then aggregates these features to obtain the group g. l Long-term preference in the overall score sequence and its long-term preference on multi-attribute rating sequences. Simultaneously, by leveraging attention mechanisms to acquire user weights, the decision-making process can be modeled more comprehensively, resulting in group g. l Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences

[0128] Predictive scoring module: Initializes group embedding representation Overall group preferences Long-term preference of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences We perform a weighted summation to obtain the group's overall preference g. l ,

[0129]

[0130] By analyzing the group's overall preferences g l With candidate project i j Perform inner product operation to obtain group g. l For candidate project i j Predicted score

[0131]

[0132] Method optimization module: After obtaining the predicted score, it utilizes the total loss function. The method is optimized, and the total loss function consists of two parts: the recommendation loss function. and contrastive learning loss function Using minimizing the total loss function To optimize the method:

[0133]

[0134] Where λ is the coefficient and π is the regularization weight of the learning parameter set Θ;

[0135] Recommendation list generation module: The system generates a corresponding project recommendation list based on the comprehensive embedding representation of the group to be recommended.

Claims

1. A group sequence recommendation method that considers users' short-term and long-term interests, characterized in that, Includes the following steps: Step 1: Use Embedding encoding to group The initial embedding of the user ID is used to represent the group. The user vector within is represented as The user vectors within the group are concatenated and their dimensionality reduced to obtain the group. initial embedding representation ; Step 2: Consolidate Groups Based on the interaction sequences of internal users, a group sequence item graph is constructed. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix is ​​established for the sequence item graph. ,right After adding an identity matrix and normalizing, the initial embedding representation of the group is processed by a graph convolutional network. Perform a linear transformation to obtain the overall group preference. ; Step 3: Utilize the attention mechanism to learn the long-term preference features of users within the group from the overall user rating sequence and the multi-attribute rating sequence, respectively, and obtain the group by aggregation. Long-term preference in the overall score sequence and its long-term preference on multi-attribute rating sequences. Simultaneously, by leveraging attention mechanisms to acquire user weights, a more comprehensive model of the decision-making process is obtained, leading to group... Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences ; Step 4: Initialize the group embedding representation Overall group preferences Long-term preferences of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences We perform a weighted summation to obtain the group's overall preference. By analyzing the overall preferences of the groups With candidate projects Perform inner product operation to obtain the group For candidate projects Predicted score ; Step 5: Establish contrastive learning by maximizing the consistency of long-term and short-term preferences of overall user ratings and multi-attribute ratings within the group. Then, optimize the method using the total loss function obtained by mixing the recommendation loss function and the contrastive loss function. Based on the comprehensive embedding representation of the group to be recommended, the system generates a corresponding item recommendation list for it.

2. The group sequence recommendation method considering users' short-term and long-term interests according to claim 1, characterized in that, The implementation method of step 2 is as follows: The learning process of the overall group preference is achieved by integrating groups. The interaction sequences of internal users are used to construct a group sequence item graph. Then, based on the connection relationships between items in the group sequence item graph, an adjacency matrix of the sequence item graph is established. ( The specific calculation formula is as follows: ， To delve deeper into the characteristics of nodes themselves and learn more accurate node features, Add an identity matrix and perform normalization; the specific calculation formula is as follows: ， in, This is the normalized adjacency matrix of the sequence item graph. It is a diagonal matrix. It is the identity matrix; Furthermore, the initial embedding representation of the group is obtained through a graph convolutional network. Perform a linear transformation to obtain the overall group preference. The calculation formula is as follows: ， in, For the overall group preference of layer t-1, For groups The initial embedding representation, Let be the trainable parameter matrix of the t-th layer.

3. The group sequence recommendation method considering users' short-term and long-term interests according to claim 1, characterized in that, The implementation method of step 3 is as follows: The user's interaction records within one month are used as the user's short-term sequence, and the user's interaction records within one year are used as the long-term sequence for analysis; To learn the long-term preferences of a group, based on in-depth mining of the overall user rating sequence and multi-attribute rating sequence, an attention mechanism is used to learn and aggregate the long-term preference features of users within the group, thus obtaining the group's long-term preferences. Long-term preference in the overall score sequence and long-term preferences on multi-attribute rating sequences ; Group Long-term preference in the overall score sequence The calculation formula is as follows: ， ， ， ， in, Indicates user Long-term sequence information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix; Group Long-term preferences on multi-attribute rating sequences The calculation formula is as follows: ， ， ， in, Indicates user In attributes Long-term sequence information on Indicates user In attributes Weight on, Indicates group In attributes Long-term preferences This indicates the influence of long-term group preferences on the weight matrix. This represents the long-term preference bias vector of the group; In-depth analysis of recent interaction sequences of users within a group can effectively capture the group's current needs and ensure that recommended items match the group's current interests; By comprehensively using the overall user rating sequence and multi-attribute rating sequences to learn the short-term preferences of users within a group, and by acquiring user weights through an attention mechanism, a comprehensive model of the decision-making process is obtained, thus yielding the group... Short-term preference in the overall score sequence and its short-term preferences on multi-attribute rating sequences ; Group Short-term preference in the overall score sequence The calculation formula is as follows: ， ， ， ， in, Indicates user Short-term series information in the overall score, Indicates user The query matrix in the overall score, Indicates user Key matrix in overall rating, Indicates user The Value matrix in the overall rating, Represents the learnable transformation matrix; Group Short-term preferences on multi-attribute rating sequences The calculation formula is as follows: ， ， ， in, Indicates user In attributes Short-term sequence information on Indicates user In attributes Weight on, Indicates group In attributes Short-term preferences This represents the matrix representing the influence of short-term preferences on the group. This represents the short-term preference bias vector of the group.

4. The group sequence recommendation method considering users' short-term and long-term interests according to claim 1, characterized in that, The method for generating the predicted score in step 4 is as follows: Initial embedding representation of the group Overall group preferences Long-term preferences of the overall group score sequence Long-term preferences of group multi-attribute rating sequences Short-term preferences of the overall group score sequence and group multi-attribute rating sequence short-term preferences Weighted summation yields the group's overall preference. The specific calculation formula is as follows: ， Group overall preferences With candidate projects Perform inner product operation to obtain the group For candidate projects Predicted score The specific calculation formula is as follows: 。 5. The group sequence recommendation method considering users' short-term and long-term interests according to claim 1, characterized in that, The specific optimization method in step 5 is as follows: After obtaining the predicted score, the total loss function is used. Optimize the method; The total loss function consists of two parts: the recommendation loss function. and contrastive learning loss function ; The recommendation loss function The calculation formula is as follows: ， in, This indicates the number of candidate items in the group. Indicates training set groups For the project The actual rating, express right Predicted score; The average rating of a project by users within a group is used in the recommendation loss function to represent the actual rating of the group for the project. To obtain high-quality group embedding representations, a contrastive learning approach is adopted to maximize the consistency between the long-term and short-term preferences of the overall user ratings and the long-term and short-term preferences of the multi-attribute ratings within the group. The higher the similarity, the better the recommendation effect. The contrastive learning loss function The calculation formula is as follows: ， ， ， in, The cosine similarity function is used. For groups Similarity between long-term preferences of the overall rating sequence and long-term preferences of multi-attribute rating sequences. For groups Similarity between short-term preferences of the overall rating sequence and short-term preferences of multi-attribute rating sequences. For coefficients; By minimizing the total loss function To optimize the method: ， in, For coefficients, To learn the regularized weights of the parameter set 𝛩.

Images