Session recommendation method based on mixed intention and double constraints

Abstract

The invention discloses a session recommendation method based on mixed intentions and double constraints, and belongs to the field of computer science. The method comprises three basic parts: a mixed intention learning module extracts a global semantic intention from all articles and distinguishes a target intention from a noise intention; the intention constraint loss function optimizes session characterization through double constraints, and collaboration of long-tail coverage and noise suppression is achieved; and the plug-and-play integration module is seamlessly integrated with the existing session recommendation model. According to the method, through collaborative design of the mixed intention learning module and the intention constraint loss function, the core problems of noise interference, characterization deviation, deployment stiffness and the like in traditional long-tail recommendation are systematically solved. According to the invention, collaborative optimization of accuracy and diversity is realized, and a traditional tradeoff relationship is broken; industrial-grade expandability is achieved, the modular design is adaptive to a mainstream model, and the landing threshold is lowered; meanwhile, dynamic scene robustness is achieved, and long-term effect stability is guaranteed through an increment mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of computer science, and in particular relates to a method for solving the long-tail problem of a recommendation system based on dual constraints of mixed intentions. Background Art

[0002] Session-based Recommendation (SBR) technology predicts the items that anonymous users may choose next through their short-term interactive behavior, and is widely used in scenarios such as e-commerce and music streaming. The initial method of session recommendation was based on simple assumptions and used Markov chains to extract users' short-term interest representations. As deep learning has shown advantages in complex information modeling, researchers have introduced recurrent neural networks (RNNs) to learn the temporal associations between items. STAMP emphasizes the last click in a session through a short-term memory priority module. Since then, with the success of Transformer in other fields, the self-attention mechanism has been introduced into session recommendation. For example, SASRec used the self-attention mechanism to extract contextual dependencies within a session for the first time. In addition, the significant advantages of graph neural networks (GNNs) in modeling complex relationships have made them attract much attention in the field of session recommendation. For example, SRGNN The innovative use of a gated graph neural network (Gated GNN) modeled sessions as a graph. GCE-GNN introduced a global graph to capture item transfer patterns across all sessions, pointing out that graphs constructed solely from a single session cannot fully model session behavior. Based on this idea, many models have incorporated contrastive learning to further improve performance.

[0003] However, existing technologies generally face the problem of long-tail distribution: a small number of high-frequency items (head items) dominate, while a large number of low-frequency items (tail items) are ignored due to insufficient exposure, resulting in limited recommendation diversity. To alleviate this long-tail problem, existing research mainly uses the following technical paths to improve:

[0004] Tail item enhancement strategies: This approach balances the long-tail distribution by dynamically adjusting the model's recommendation weights for head and tail items. For example, TailNet dynamically adjusts the recommendation weights for head and tail items through a preference mechanism. If a user expresses interest in a tail item, the weight of the tail item is increased. This approach flexibly adapts to the preferences of different users and avoids the rigidity of recommendations caused by fixed weights. However, it does not distinguish between noise in tail items (such as unpopular items that users have no interest in), resulting in a recommendation list cluttered with irrelevant content and the risk of introducing noise. Furthermore, this approach relies on real-time user behavior and is less adaptable to sparse interaction scenarios (such as anonymous sessions).

[0005] Distribution alignment optimization: This solution forces the model to focus on tail items by constraining the model's output distribution to align with the true item distribution. For example, CSBR introduces distribution prediction and alignment objectives, forcing the model to focus on tail items. It also introduces distribution prediction loss and distribution alignment loss. The former predicts the global frequency distribution of items, while the latter forces the model output to be consistent with the true distribution. This can alleviate the overexposure problem of top items from a probability distribution perspective. However, global distribution alignment may deviate from user personalized needs. For example, even if a user has no interest in a certain type of tail item, they may still be forced to be recommended, thus ignoring the user's needs. Furthermore, CSBR requires maintaining global item distribution statistics, which makes real-time updates costly.

[0006] Tail Session Enhancement: Tail session enhancement improves the visibility of tail items in training data through data augmentation or graph structure mining. For example, LOAM improves the exposure of tail items through tail session mixing and random walks. It combines two strategies: 1. Performing random walks on tail items in the graph structure to generate pseudo sessions containing tail items. 2. Mixing real tail sessions with head sessions to enhance the model's perception of tail features. This can directly increase the exposure of tail items, but pseudo sessions may contain irrelevant item combinations, interfering with the model's learning of true user intent and undermining semantic consistency.

[0007] Post-processing diversity optimization: This approach uses reordering or filtering strategies to improve the coverage of tail items after the model generates the recommendation list. For example, LAP-SR incorporates a personalized diversity strategy into the recommendation list. This method calculates a personalized diversity score based on user profiles and item attributes, and reorders the recommendation list to ensure that tail items account for a certain proportion. This approach requires no modifications to the model structure and offers flexible deployment. However, the post-processing step increases computational latency, making it difficult to meet real-time recommendation requirements. Furthermore, it only adjusts the order of the recommendation list, without addressing the model's inherent distribution bias. Summary of the Invention

[0008] This paper proposes a hybrid intent-based dual constraint framework (HID) to address the long-tail problem in recommendation systems. To address the aforementioned technical issues, the present invention employs the following specific technical solutions:

[0009] Step 1: Based on the mixed intent learning module, we mine the mixed intents hidden in the interaction data. First, we divide all items in the conversation into k attributes representing specific preliminary intents, and then further subdivide them into n mixed intents.

[0010] Step 2: Construct an intent-constrained loss function to optimize the conversation representation using dual constraints, including long-tail constraints and accuracy constraints.

[0011] Step 3: Deploy and optimize the conversational recommendation model. Deploy the hybrid intent learning module and the intent constraint-based loss function on the conversational recommendation model, train and optimize the deployed conversational recommendation model, and apply the trained conversational recommendation model to the conversational recommendation task.

[0012] Furthermore, the step 1 is specifically as follows:

[0013] Step 1.1: Preliminary intention unit construction; given an attribute set consisting of k attributes; each attribute is defined as an item set;

[0014] Step 1.2: Intent association graph construction: replace the item ID in the session with its corresponding item attribute ID; then, traverse all attributes in each session, count the 1-hop neighbors of each attribute and their co-occurrence frequency, and construct the intent association graph; the intent association graph is represented as ,in, is a collection of attribute IDs, is the edge between an attribute and its neighbor attributes, is a weight set, where the edge weight represents the co-occurrence frequency of the attribute and the neighbor attribute;

[0015] Step 1.3: Generate mixed intents. First, calculate the Laplacian matrix based on the co-occurrence frequency. Then, calculate the eigenvalues ​​and eigenvectors of the normalized Laplacian matrix. Take the smallest q eigenvalues ​​and their corresponding eigenvectors to form an eigenvector matrix. Each row of the eigenvector matrix represents the embedding of the node in the reduced q-dimensional space.

[0016] Then apply the clustering algorithm to all rows of the feature vector matrix, reclassify the attributes into n clusters, merge the attributes belonging to the same cluster into mixed intents, and n and mixed intents form a mixed intent set ; Concatenate the items in the attributes and use average pooling to get the mixed intent embedding;

[0017] Step 1.4: Define target intent and noise intent: For each batch of sessions , where b is the size of the current batch, session , ( ), For session S u The interaction term of , l is the length of the session, and the target intent and mixed intent are defined as follows:

[0018] If mixed intentions Contains session S u The next interaction term ,but For S u Target intention c u ;

[0019] For session S u , other sessions in batch B The target intention constitutes its noise intention , the noise intention is not equal to the session S u target intention.

[0020] Furthermore, the loss function based on intention constraint is specifically as follows:

[0021] ;

[0022] in, is the loss function based on intention constraint, d( , ) represents the calculation of Euclidean distance, is the flexibility coefficient, is a hyperparameter, p u For penalty items; Embedding representation for target intent, Embedding representation for the session, For the current session The noise intention, Embed the corresponding noise intent representation.

[0023] The penalty term p u It is calculated by the following formula:

[0024] ;

[0025] Among them, Var represents variance calculation, is the variance threshold.

[0026] Furthermore, the step 3 is specifically as follows:

[0027] Step 3.1: Model input compatibility: reuse the session embedding of the existing session recommendation model;

[0028] Step 3.2: Training process adaptation: During the model training phase, add ICLoss to the loss function of the original conversation recommendation model and backpropagate to update the parameters;

[0029] Step 3.3: Deployment optimization: Precompute and store the hybrid intent embedding. In the real-time recommendation phase, only the similarity between the session and the intent needs to be calculated.

[0030] Furthermore, the conversation recommendation model includes a STAMP model, an SRGNN model, or other conversation recommendation models adapted to the conversation embedding.

[0031] The present invention mainly includes the following three core technical points:

[0032] 1. Hybrid Intent Learning Module

[0033] 1. Traditional intent extraction relies on temporal features within a single session (such as the last clicked item) and ignores global semantic associations across sessions (such as product category combinations and music genre co-occurrence). This results in unreliable intent representation, local intent modeling, inability to capture users' underlying preferences, and susceptibility to interaction noise, resulting in "intent representation distortion." Therefore, a hybrid intent learning module is proposed. This module constructs a global intent association graph based on attribute co-occurrence graphs, using co-occurrence frequencies based on item attributes (such as category and brand) to capture cross-session semantic associations. Furthermore, a spectral clustering algorithm is used to group co-occurring attributes into hybrid intents (such as "cooking = food + kitchenware"), reflecting users' potential complex preferences.

[0034] 2. Intent Constraint Loss

[0035] 2. Traditional conversational recommendation models typically use a single-objective loss function (such as the cross-entropy loss). This loss function fails to specifically optimize the representation of tail items, leading the model to overemphasize high-frequency head items. Recommendations are easily affected by conversation-irrelevant noise in tail items (such as unpopular items that users lack interest in), reducing accuracy. Furthermore, the distribution differences between head and tail items in the feature space are not corrected, exacerbating recommendation unfairness. To address these issues, this paper proposes the intent-constrained loss function (ICLoss). This function collaboratively optimizes conversational representations through a dual constraint mechanism: a long-tail constraint and an accuracy constraint. The long-tail constraint narrows the representation gap between head and tail items within the target intent, while the accuracy constraint suppresses noise intent interference through contrastive learning and a variance penalty. Furthermore, the overall loss function combines ICLoss with the traditional cross-entropy loss to achieve a multi-objective optimization. The loss weight ϵ adjusts over the training phase to balance model convergence and long-tail optimization.

[0036] 3. Plug-and-play framework

[0037] Existing traditional methods require modifications to the model architecture (such as TailNet) or rely on post-processing (such as LAP-SR), making them difficult to adapt to conversational recommendation models in different scenarios. This solution, a plug-and-play framework, uses a modular design that encapsulates hybrid intent learning and intent-constrained loss as independent modules. This allows for seamless integration with existing SBR models (such as STAMP and SRGNN), while achieving zero model reconstruction. Simply overlaying ICLoss during training eliminates the need for model structure adjustments, significantly reducing industrial deployment costs and shortening deployment time by 90%.

[0038] Thanks to the above three designs, HID has achieved the best recommendation effect in the field of conversational recommendation.

[0039] The beneficial effects of the present invention are as follows:

[0040] Compared to existing technologies, this invention systematically addresses core issues in traditional long-tail recommendation, such as noise interference, representation bias, and deployment rigidity, through the collaborative design of a hybrid intent learning module and an intent-constrained loss function (ICLoss). This invention achieves the coordinated optimization of accuracy and diversity, breaking the traditional trade-off. It also achieves industrial-grade scalability, with a modular design adaptable to mainstream models, lowering the barrier to implementation. It also offers robustness in dynamic scenarios, and an incremental mechanism ensures long-term stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall architecture diagram of the method of the present invention. DETAILED DESCRIPTION

[0042] The core of this invention is to propose a dual constraint framework (HID) based on hybrid intent, the overall architecture of which is as follows: Figure 1 As shown, this method includes three basic parts: 1) Hybrid Intent Learning Module, which extracts global semantic intent from all items and distinguishes target intent from noise intent. It includes three steps: preliminary intent unit construction, intent association graph construction, and hybrid intent generation. 2) Intent Constraint Loss, which optimizes session representation through dual constraints to achieve the synergy of long-tail coverage and noise suppression. This part includes long-tail constraints, accuracy constraints, and multi-task loss fusion. 3) Plug-and-play framework: seamlessly integrate the above module with existing session recommendation models (such as STAMP and SRGNN). The present invention specifically includes the following steps:

[0043] Step 1: Based on the Hybrid Intent Learning Module, an attribute-aware spectral clustering strategy for recommendation scenarios is designed to mine the hybrid intent hidden in the interaction data. The schematic diagram is as follows: Figure 1As shown on the right. Existing intent mining techniques usually learn intent embeddings based on a small number of consecutive items in a single session. This approach has two drawbacks: (1) it only considers the temporal relationship between items, which is not always reliable due to the presence of interaction noise; (2) it only considers a single session, ignoring the fact that items in different sessions may also reflect the same intent. Therefore, this paper proposes attribute-aware spectral clustering, which first divides all items into k attributes representing specific preliminary intents, and then further subdivides them into n mixed intents. The whole process is divided into three parts: preliminary intent unit construction, intent association graph construction, and mixed intent generation.

[0044] Step 1.1: Preliminary intention unit construction;

[0045] Since items sharing the same attributes usually reflect similar user preferences (e.g., electronic products or books), item attributes are considered as preliminary intent units. , where the i-th attribute Indicates a specific preliminary intention, k is the total number of attributes. Each attribute is defined as a collection of items: ,in Represents the total number of items contained in this item set, Represents the jth item in the set of items with the i-th attribute.

[0046] Step 1.2: Construct the intention association graph;

[0047] In order to explore the attribute association relationship in all sessions, we first replace the item ID in the session with its corresponding item attribute ID. Then, we traverse all attributes in each session, count the 1-hop neighbors of each attribute and their co-occurrence frequency, and construct an intent association graph. The intent association graph is represented as ,in, is a collection of attribute IDs, For attributes and The edges between ,in For attributes The neighbor set of is the weight set, edge Weight Representation attributes and co-occurrence frequency.

[0048] Step 1.3: Mixed intent generation;

[0049] After obtaining the intention association graph G, in order to mine the global co-occurrence pattern of attributes, spectral clustering is used to learn the topological relationship between attributes. , first calculate its Laplacian matrix:

[0050]

[0051] in is the i-th row and i-th column of matrix D, I is the identity matrix, D is the degree matrix, W is the weight matrix, and the elements in the weight matrix are the co-occurrence frequencies Then, calculate the eigenvalues ​​and eigenvectors of the normalized Laplace matrix L. Take the smallest q eigenvalues ​​as , whose corresponding eigenvector , forming a feature vector matrix. Each row of this matrix represents the embedding of the node in the q-dimensional space after dimensionality reduction.

[0052] Next, the k-means algorithm is applied to all rows of the eigenvector matrix. The i-th row of the eigenvector matrix corresponds to the i-th attribute in the attribute set C′ (i.e., the node in the intent association graph). Thus, by applying k-means clustering to all rows, the attributes are reclassified into n clusters. Since the attributes represent preliminary intents, the attributes belonging to the same cluster are merged into mixed intents. The mixed intent set is defined as , where the i-th mixed intent is represented as , and contains items, represented by The embedding of the mixed intent is generated by the embedding of the items corresponding to the attributes. To reduce the time complexity, the items in the attributes are concatenated and average pooling is used to obtain the mixed intent embedding:

[0053]

[0054] in Representing mixed intent Embedded, Represents the jth item in the i-th mixed intent Embedding.

[0055] Step 1.4: Define the objective function and noise intent:

[0056] After obtaining the mixed intent set, for each batch of sessions , b is the size of the current batch, for session ,in( ), For session S u The i-th interaction term of , l is the length of the session, and the target intent and mixed intent are defined as follows:

[0057] Target intent: If mixed intent Contains session S u The next interaction term , then mixed intent For S u Target intention c u :

[0058] .

[0059] Noise Intent: For session S u , other sessions in batch B The target intention (and not equal to c u ) constitutes its noise intention :

[0060] .

[0061] Step 2: Based on the Intent Constraint Loss function module, the session representation is optimized through dual constraints to achieve synergy between long-tail coverage and noise suppression.

[0062] Extracting mixed intent sets After that, the next step is to impose constraints on the learning process of session embedding. Given a traditional SBR model (such as STAMP or SRGNN) with input session S u The generated session embedding S' u , its learning is guided by an additional training loss (i.e., intent constraint loss, ICLoss). Since ICLoss can be integrated with any recommendation model, it is a model-independent plug-in.

[0063] Step 2.1: Long tail constraint: By reducing the similarity between the session and the target intent, the similarity gap between the session and the head item and the tail item is narrowed, thereby increasing the tail recommendation probability. The long tail constraint is defined as minimizing the session embedding S' u With target intention c u Distance from center:

[0064] ,

[0065] in, For items Embedded representation of for Var is the variance calculation, and d(x,y) measures the Euclidean distance between variables x and y. At the same time, the time complexity of the above operation is given by Calculate, N is the target intention c u The number of items, d is the embedding dimension. Since the framework HID of this method is a model-independent plug-in, obtaining low complexity is a key issue.

[0066] Therefore, an approximate formula with lower complexity is further proposed, and the time complexity can be approximated to O(d):

[0067] .

[0068] The above formula is proved as follows:

[0069] ,

[0070] Then calculate Gradient:

[0071] ,

[0072] When the gradient is 0, we get S u The expression:

[0073]

[0074] Considering that when S u Point to each point V i When the sum of the unit vectors is approximately 0, the second term:

[0075]

[0076] When v i Around S u This condition can be satisfied when the distribution is approximately symmetrical. u When the second term is approximately 0, S u will be approximately equal to the first term, which means that all v i The center of mass.

[0077] therefore, It is an approximate solution to the first proof above. At the same time, the target intention embedding is obtained by average pooling , therefore, minimize d(S u ,c u ) is approximately equivalent to the expression for minimizing the long-tail constraint in the previous text.

[0078] Step 2.2: Accuracy constraint; To avoid irrelevant recommendations, it is necessary to reduce the proportion of noise intent in the recommendation list, so an accuracy constraint is given, that is, given a session embedding S u , maximizing the mean of the Euclidean distance to its noise intention while limiting the variance:

[0079] ,

[0080] ,

[0081] in, is the variance threshold.

[0082] Step 2.3: Intention Constraint Loss; By deducing the above constraints, we get:

[0083] ,

[0084]

[0085] Here exp(x) is used to amplify the difference between the target and the noise intent. In order to minimize the impact of the noise intent, another lemma is given: the approximate optimization of the above equation is equivalent to optimizing a (N-1) triplet loss with a fixed margin of 2:

[0086] ,

[0087] .

[0088] The proof of the above formula is as follows:

[0089] ,

[0090] ,

[0091] ,

[0092] ,

[0093] .

[0094] The constant term "2" determines and A fixed margin of difference. However, this fixed margin is difficult to effectively distinguish the target intent from the noise intent in scenarios where the intent distribution is highly variable or there are ambiguous intents. Therefore, it is replaced with a flexible coefficient "2σ" to dynamically adjust the margin based on the recommended scenario:

[0095] ,

[0096] .

[0097] In order to directly use gradient descent updates and avoid the complexity of constrained optimization, the hard variance constraint Converted into penalty term p u :

[0098] .

[0099] Finally, ICloss is defined as:

[0100] ,

[0101] in, is a hyperparameter that controls the intensity of the penalty, and p u is scaled to the interval (0,1); where is the loss function based on intention constraint, d( , ) represents the calculation of Euclidean distance, is the flexibility coefficient, Embedding representation for target intent, Embedding representation for the session, For the current session The noise intention, Embed the corresponding noise intent representation.

[0102] Step 2.4: Multi-task learning; To integrate HID into the traditional SBR model, the weight of ICLoss is controlled by the hyperparameter ϵ, and the total loss is:

[0103] .

[0104] in It is the cross entropy loss adopted by the traditional SBR model.

[0105] Step 3: Based on the plug-and-play framework, realize the integration of HID and model;

[0106] Step 3.1: Model input compatibility: reuse the session embedding S of the existing session recommendation model u (such as STAMP's attention output and SRGNN's graph embedding).

[0107] Step 2: Training process adaptation: During the model training phase, ICLoss is added to the loss function of the original SBR model, and the parameters are updated through backpropagation.

[0108] Step 3: Optimization for industrial deployment: Precompute and store the hybrid intent embedding. In the real-time recommendation phase, only the similarity between the session and the intent needs to be calculated.

[0109] Extensive experiments were conducted on two public datasets, and the results showed that our method achieved the state-of-the-art level.

[0110] Datasets: Tmall comes from the IJCAI-15 competition and consists of shopping logs of many unnamed users on the Tmall online shopping platform; Diginetica comes from the 2016 CIKM Cup and describes users' music listening behavior.

[0111] The data format is shown in Table 1:

[0112] Table 1. Dataset parameters

[0113] Dataset Tmall Diginetica Training Session 351,268 433,643 Test Session 25,898 15,132 thing 40,728 36,968 Average length 6.69 5.43

[0114] Evaluation indicators:

[0115] Accuracy metrics: NDCG (normalized discounted cumulative gain), HR (hit rate), MRR (mean reciprocal rank).

[0116] Long-tail indicators:

[0117] tNDCG (Normalized Discounted Cumulative Gain of Tail Items), tHR (Hit Rate of Tail Items), tMRR (Mean Reciprocal Rank of Tail Items): Metrics calculated only for sessions where the tail items are the true labels.

[0118] tCov (tail coverage): measures the proportion of different tail items that appear in the recommendation list.

[0119] Tail: The average proportion of tail items in each recommendation list.

[0120] Basic SBR model:

[0121] STAMP: STAMP emphasizes short-term memory by utilizing the attention mechanism.

[0122] SR-GNN: SR-GNN uses GNN to learn item embeddings and learn session representations through a soft attention mechanism.

[0123] GCE-GNN: GCE-GNN considers extracting the spatial information of session data from both global and local views.

[0124] HIDE: HIDE decouples the intent in the original session in a micro- and macro-manner to capture the user's dynamic intention and avoid noisy signals.

[0125] Comparative long-tail recommendation method:

[0126] TailNet: Adjusting the importance of head and tail items via a preference mechanism.

[0127] CSBR: Alleviating the long-tail problem through distribution prediction and alignment.

[0128] LOAM: Improving long-tail performance through niche-walk enhancement and tail session mixing.

[0129] LAP-SR: A post-processing method that mitigates the long-tail effect based on personalized diversity.

[0130] Table 2. Accuracy and long-tail performance of the SBR model using the long-tail method on two datasets

[0131]

[0132] Table 2 shows the results of various models on accuracy and long-tail metrics. As can be seen, existing long-tail methods sacrifice accuracy in favor of long-tail performance due to their failure to filter out noise in the tail terms. HID, however, outperforms the baseline in both accuracy and long-tail performance. This is due to the hybrid intent's ability to perceive high-level user intent, and the intent constraint loss balances target and noisy intents, both theoretically and practically. Furthermore, HID achieves even more significant improvements on Tmall, as user intent in e-commerce scenarios is more complex (for example, "cooking" requires combining ingredients and kitchenware), and hybrid intent effectively captures these associations.

[0133] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A conversation recommendation method based on mixed intention and dual constraints, characterized by: The following steps are involved: Step 1: Based on the mixed intent learning module, we mine the mixed intents hidden in the interaction data. First, we divide all items in the conversation into k attributes representing the initial intent, and then further subdivide them into n mixed intents. Step 2: Construct an intent-constrained loss function to optimize the conversation representation using dual constraints, including long-tail constraints and accuracy constraints. Step 3: Deploy and optimize the conversational recommendation model; deploy the hybrid intent learning module and the intent-constrained loss function on the conversational recommendation model, train and optimize the deployed conversational recommendation model, and apply the trained conversational recommendation model to the conversational recommendation task.

2. The conversation recommendation method based on hybrid intention and dual constraints according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Preliminary intention unit construction; given an attribute set consisting of k attributes, each attribute is defined as an item set; Step 1.2: Intent association graph construction: replace the item ID in the session with its corresponding item attribute ID; then, traverse all attributes in each session, count the 1-hop neighbors of each attribute and their co-occurrence frequency, and construct the intent association graph; the intent association graph is represented as ,in, is a collection of attribute IDs, is the edge between an attribute and its neighbor attributes, is a weight set, where the edge weight represents the co-occurrence frequency of the attribute and the neighbor attribute; Step 1.3: Generate mixed intents. First, calculate the Laplacian matrix based on the co-occurrence frequency. Then, calculate the eigenvalues ​​and eigenvectors of the normalized Laplacian matrix. Take the smallest q eigenvalues ​​and their corresponding eigenvectors to form an eigenvector matrix. Each row of the eigenvector matrix represents the embedding of the node in the reduced q-dimensional space. Then apply the clustering algorithm to all rows of the feature vector matrix, reclassify the attributes into n clusters, merge the attributes belonging to the same cluster into mixed intents, and form a mixed intent set of n mixed intents. , where the i-th mixed intent is represented as ; Concatenate the items in the mixed intent and use average pooling to get the mixed intent embedding; Step 1.4: Define target intent and noise intent: The current batch includes b sessions, where the uth session Including l interaction terms, the target intention and mixed intention are defined as follows: If mixed intentions Contains session S u The next interaction term ,but For S u Target intention c u ; For session S u , other sessions in the current batch The target intention constitutes its noise intention , the noise intention is not equal to the session S u target intention.

3. The conversation recommendation method based on hybrid intention and dual constraints according to claim 2, characterized in that: The loss function ICLoss based on intention constraint is as follows: ； in, is the loss function based on intention constraint, d( , ) represents the calculation of Euclidean distance, is the flexibility coefficient, is a hyperparameter, p u For penalty items; Embedding representation for target intent, Embedding representation for the session, For the current session The noise intention, Embed the corresponding noise intent representation.

4. The method for conversation recommendation based on hybrid intention and dual constraints according to claim 3, characterized in that: The penalty term p u It is calculated by the following formula: ； Among them, Var represents variance calculation, is the variance threshold.

5. The method for conversation recommendation based on hybrid intention and dual constraints according to claim 4, characterized in that: The step 3 is specifically as follows: Step 3.1: Model input compatibility: reuse the session embedding of the existing session recommendation model; Step 3.2: Training process adaptation: During the model training phase, add ICLoss to the loss function of the original conversation recommendation model and backpropagate to update the parameters; Step 3.3: Deployment optimization: Precompute and store the hybrid intent embedding. In the real-time recommendation phase, only the similarity between the session and the intent needs to be calculated.

6. The conversation recommendation method based on hybrid intention and dual constraints according to claim 5, characterized in that: The session recommendation model includes a STAMP model, an SRGNN model, or other session recommendation models adapted to the session embedding.

Images