A Session Recommendation Method Based on Hybrid Intent and Dual Constraints

Abstract

This invention relates to a conversation recommendation method based on hybrid intent and dual constraints, belonging to the field of computer science. The method comprises three basic parts: a hybrid intent learning module extracts global semantic intent from all items, distinguishing between target intent and noisy intent; an intent constraint loss function optimizes conversation representation through dual constraints, achieving a synergy between long-tail coverage and noise suppression; and a plug-and-play integration module that seamlessly integrates with existing conversation recommendation models. This invention systematically solves core problems in traditional long-tail recommendation, such as noise interference, representation bias, and rigid deployment, through the collaborative design of the hybrid intent learning module and the intent constraint loss function. This invention achieves synergistic optimization of accuracy and diversity, breaking the traditional trade-off; it achieves industrial-grade scalability, with modular design adapting to mainstream models and lowering the threshold for implementation; and it also possesses robustness to dynamic scenarios, with an incremental mechanism ensuring long-term performance stability.

Description

Technical Field

[0001] This invention belongs to the field of computer science, and in particular relates to a method for solving the long-tail problem in recommendation systems based on dual constraints of hybrid intent. Background Technology

[0002] Session-based recommendation (SBR) technology predicts the next item a user might choose based on their short-term interactions, and is widely used in e-commerce, music streaming, and other scenarios. Initial methods of session recommendation were based on simple assumptions, using Markov chains to extract short-term interest representations of users. As deep learning demonstrated its advantages in modeling complex information, researchers introduced recurrent neural networks (RNNs) to learn temporal relationships between items. STAMP emphasizes the last click in a session through a short-term memory priority module. Subsequently, with the success of Transformers in other fields, self-attention mechanisms were introduced into session recommendation. For example, SASRec was the first to utilize self-attention to extract contextual dependencies within a session. Furthermore, the significant advantages of graph neural networks (GNNs) in modeling complex relationships have made them highly sought after in the field of session recommendation. For example, SRGNN... This paper innovatively employs a gated graph neural network (GNN) to model sessions as a graph structure. GCE-GNN introduces a global graph to capture item transition patterns across all sessions, demonstrating that graphs built solely from single sessions cannot comprehensively model session behavior. Based on this idea, many models combine contrastive learning to further improve performance.

[0003] However, existing technologies generally face the problem of long-tail distribution: a few 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 the long-tail problem, existing research mainly improves it through the following technical approaches:

[0004] Tail item enhancement strategy: This approach balances the long-tail distribution by dynamically adjusting the recommendation weights of head and tail items in the model. For example, TailNet dynamically adjusts the recommendation weights of head and tail items through a preference mechanism; if a user shows interest in a tail item, its recommendation weight is increased. This method can flexibly adapt to the differences in user preferences and avoid the rigidity of recommendations caused by fixed weights. However, it does not distinguish noise in the tail items (such as unpopular items that users are not interested in), resulting in irrelevant content mixed in the recommendation list and the risk of introducing noise. At the same time, this method relies on real-time user behavior and is not well adapted to sparse interaction scenarios (such as anonymous conversations).

[0005] Distribution alignment optimization: This approach forces the model to focus on tail items by aligning the model's output distribution with the real item distribution. For example, CSBR introduces distribution prediction and alignment objectives, forcing the model to focus on tail items and introducing distribution prediction loss and distribution alignment loss. The former predicts the global frequency distribution of items, while the latter forces the model output to match the real distribution. This can alleviate the overexposure problem of head items from a probability distribution perspective. However, global distribution alignment may deviate from users' personalized needs; for example, users may be forced to be recommended items even if they have no interest in a certain type of tail item, thus ignoring their needs. Furthermore, CSBR requires maintaining global item distribution statistics, which is costly to update in real time.

[0006] Tail-end conversation enhancement improves the visibility of tail-end items in training data through data augmentation or graph structure mining. For example, LOAM enhances the exposure of tail-end items by combining tail-end conversation blending and random walks. It combines two strategies: ① performing random walks on the graph structure for tail-end items to generate pseudo-conversations containing them. ② blending real tail-end conversations with head-end conversations to enhance the model's perception of tail-end features. While this directly increases the exposure opportunities of tail-end items, pseudo-conversations may contain unrelated combinations of items, interfering with the model's learning of genuine user intent and disrupting semantic consistency.

[0007] Post-processing diversity optimization involves improving the coverage of less desirable items after the model generates the recommendation list through reordering or filtering strategies. For example, LAP-SR introduces a personalized diversity strategy in its recommendation list. This method calculates a personalized diversity score based on user profiles and item attributes, reorders the recommendation list, and ensures that less desirable items occupy a certain proportion. While this approach requires no modification to the model structure and is flexible in deployment, the post-processing step increases computational latency, making it difficult to meet real-time recommendation requirements. Furthermore, simply adjusting the order of the recommendation list does not address the inherent distribution bias problem within the model. Summary of the Invention

[0008] This invention proposes a method for solving the long-tail problem in recommender systems based on a hybrid intent-based dual constraint framework (HID). To address the aforementioned technical problems, the specific technical solution of this invention is as follows:

[0009] Step 1: Based on the hybrid intent learning module, mine the hidden hybrid intents in the interaction data; first, divide all items in the session into k attributes representing specific initial intents, and then further subdivide them into n hybrid intents;

[0010] Step 2: Construct a loss function based on intent constraints to optimize the session representation through dual constraints; the dual constraints include long-tail constraints and accuracy constraints.

[0011] Step 3: Deployment and optimization of the conversation recommendation model; Deploy the hybrid intent learning module and the intent constraint-based loss function onto the conversation recommendation model, train and optimize the deployed conversation recommendation model, and apply the trained conversation recommendation model to the conversation recommendation task.

[0012] Furthermore, step 1 is detailed as follows:

[0013] Step 1.1: Initial Intent Unit Construction; Given a set of k attributes; Each attribute is defined as a set of items;

[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; this intent association graph is represented as follows: ,in, For a set of attribute IDs, The edge between an attribute and its neighbor attributes. Let be the set of weights, where the weight of an edge represents the co-occurrence frequency of an attribute with the attributes of its neighbors;

[0015] Step 1.3: Hybrid intent generation; First, calculate its Laplacian matrix based on co-occurrence frequency, then calculate the eigenvalues ​​and eigenvectors of the normalized Laplacian matrix, and take the q smallest eigenvalues ​​and their corresponding eigenvectors to form an eigenvector matrix. Each row of the eigenvector matrix represents the embedding of the node in the q-dimensional space after dimensionality reduction.

[0016] Next, a clustering algorithm is applied to all rows of the feature vector matrix to reclassify the attributes into n clusters. Attributes belonging to the same cluster are merged into a hybrid intent, and n and the hybrid intents together form a hybrid intent set. The item items in the attributes are concatenated and average pooling is used to obtain the blended intent embedding.

[0017] Step 1.4: Define the 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 items, where l is the length of the session, define the target intent and hybrid intent as follows:

[0018] If the intention is to mix Includes session S u The next interactive item ,but For S u Target Intent c u ;

[0019] For session S u Other sessions in batch B The target intent constitutes its noise intent The noise is not intended to be equivalent to session S. u The intended goal.

[0020] Furthermore, the loss function based on intent constraints is as follows:

[0021] ;

[0022] in, Let d( , ) be the loss function based on intent constraints, and let d( , ) denote the calculation of the Euclidean distance. For flexibility coefficients, p is a hyperparameter u This is a penalty item; Embedding representations for target intent, For session embedding representation, For the current session The noise intention, The corresponding noise intention is embedded in the representation.

[0023] The penalty item p u Specifically, it is calculated using the following formula:

[0024] ;

[0025] Where Var represents variance calculation. It is the threshold of variance.

[0026] Furthermore, step 3 is detailed as follows:

[0027] Step 3.1: Model Input Compatibility: Reuse the session embeddings of existing session recommendation models;

[0028] Step 3.2: Training process adaptation: During the model training phase, ICLoss is applied to the loss function of the original session recommendation model, and parameters are updated via backpropagation;

[0029] Step 3.3: Deployment optimization: Pre-compute and embed hybrid intents, and store them. In the real-time recommendation stage, only the similarity between the session and the intent needs to be calculated.

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

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

[0032] I. Hybrid Intent Learning Module

[0033] 1. Traditional intent extraction relies on temporal features within a single session (such as the last clicked item), neglecting global semantic associations across sessions (such as co-occurrence of product category combinations and music genres). This leads to unreliable intent representations, and local intent modeling struggles to capture latent user preferences, becoming susceptible to interaction noise and resulting in "intent representation distortion." Therefore, a hybrid intent learning module is proposed. This module constructs a global intent association graph using co-occurrence frequencies based on item attributes (such as category and brand), capturing semantic associations across sessions. Simultaneously, a spectral clustering algorithm groups co-occurring attributes into hybrid intents (such as "cooking = food + kitchenware"), reflecting potential complex user preferences.

[0034] II. Intent Constraint Loss

[0035] 2. Traditional conversation recommendation models typically employ a single-objective loss function (such as cross-entropy loss). This approach suffers from several drawbacks: the loss function fails to specifically optimize the representation of tail items, leading to an overemphasis on high-frequency head items; recommendation results are susceptible to conversation-irrelevant noise in tail items (such as niche items of no interest to users), reducing accuracy; and the distribution differences between head and tail items in the feature space are not corrected, exacerbating recommendation unfairness. To address these issues, this invention proposes an Intent-Constrained Loss Function (ICLoss). This function collaboratively optimizes conversation representation through a dual constraint mechanism, including long-tail constraints and accuracy constraints. The long-tail constraint forcibly narrows the representation gap between head and tail items within the target intent, while the accuracy constraint suppresses interference from noisy intents through contrastive learning and variance penalty. Simultaneously, the total loss function combines ICLoss with traditional cross-entropy loss to form a multi-objective optimization. The loss weight ϵ is adjusted during the training phase, balancing model convergence and long-tail optimization.

[0036] III. Plug-and-play architecture design

[0037] Existing traditional methods require modifications to the model architecture (such as TailNet) or rely on post-processing (such as LAP-SR), making it difficult to adapt to conversation recommendation models in different scenarios. Our solution, however, is a plug-and-play framework. Its modular design encapsulates hybrid intent learning and intent constraint loss into independent modules, supporting seamless integration with existing SBR models (such as STAMP and SRGNN). It also achieves zero model reconstruction, requiring only the addition of ICLoss during the training phase without adjusting the model structure, significantly reducing industrial deployment costs and shortening deployment time by 90%.

[0038] Thanks to the three design features mentioned above, HID has achieved the best recommendation performance in the field of conversational recommendation.

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

[0040] Compared with existing technologies, this invention systematically solves the core problems of traditional long-tail recommendation, such as noise interference, representation bias, and rigid deployment, through the collaborative design of a hybrid intent learning module and intent constraint loss function (ICLoss). This invention achieves synergistic 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 threshold for implementation; and it possesses robustness to dynamic scenarios, with an incremental mechanism ensuring long-term performance stability. Attached Figure Description

[0041] Figure 1 This is a diagram illustrating the overall architecture of the method of the present invention. Detailed Implementation

[0042] The core of this invention lies in proposing a dual constraint framework (HID) based on hybrid intent, the overall architecture of which is as follows: Figure 1 As shown, this method comprises three basic parts: 1) A Hybrid Intent Learning Module, which extracts global semantic intent from all items, distinguishing between target intent and noisy intent. It includes three steps: initial intent unit construction, intent association graph construction, and hybrid intent generation. 2) An Intent Constraint Loss function, which optimizes the session representation through dual constraints, achieving a synergy between long-tail coverage and noise suppression. This part includes long-tail constraints, accuracy constraints, and multi-task loss fusion. 3) A plug-and-play integration framework: seamlessly integrating the above modules with existing session recommendation models (such as STAMP and SRGNN). The 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 was designed to mine the hidden hybrid intents in the interaction data. A simplified diagram is shown below. Figure 1As shown on the right. Existing intent mining techniques typically 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, but this relationship 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 invention proposes attribute-aware spectral clustering, which first divides all items into k attributes representing a specific preliminary intent, and then further subdivides them into n hybrid intents. The entire process consists of three parts: preliminary intent unit construction, intent association graph construction, and hybrid intent generation.

[0044] Step 1.1: Initial intent unit construction;

[0045] Since items sharing the same attributes often reflect similar user preferences (e.g., electronics or books), item attributes are considered as initial units of intent. Given a set of attributes... , where the i-th attribute This represents a specific initial intention, where k is the total number of attributes. Each attribute... Defined as a collection of items: ,in This represents the total number of items contained in this collection. This represents the j-th item in the set of items with the i-th attribute.

[0046] Step 1.2: Intent association graph construction;

[0047] To explore attribute relationships across all sessions, we first replace the item IDs in each session with their corresponding item attribute IDs. Then, we iterate through all attributes in each session, counting the one-hop neighbors and co-occurrence frequency of each attribute to construct an intent relationship graph. This intent relationship graph is represented as follows: ,in, For a set of attribute IDs, For attributes and The edge between, ,in For attributes The neighborhood group, For the weight set, the edges weight Represents attributes and The co-occurrence frequency.

[0048] Step 1.3: Generate hybrid intent;

[0049] After obtaining the intent association graph G, spectral clustering is used to learn the topological relationships between attributes in order to discover global co-occurrence patterns of attributes. Given the graph... First, calculate its Laplace matrix:

[0050]

[0051] in Let I be the identity matrix, D be the degree matrix, and W be the weight matrix, where W is the co-occurrence frequency. Then, the eigenvalues ​​and eigenvectors of the normalized Laplacian matrix L are calculated. The q smallest eigenvalues ​​are then selected. Its corresponding feature vector This forms the feature vector matrix. Each row of this matrix represents the embedding of a node in the reduced q-dimensional space.

[0052] Next, the k-means algorithm is applied to all rows of the feature vector matrix. The i-th row of the feature vector matrix corresponds to the i-th attribute in the attribute set C′ (i.e., a 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 the initial intent, attributes belonging to the same cluster are merged into a hybrid intent. The hybrid intent set is defined as follows: The i-th hybrid intent is represented as and includes Each item is represented as The embedding of a hybrid intent is generated from the item embeddings corresponding to its contained attributes. To reduce time complexity, the item items in the attributes are concatenated and average pooling is used to obtain the hybrid intent embedding.

[0053]

[0054] in Represents mixed intentions Embedded, Represents the j-th item in the i-th mixed intention. Embedded.

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

[0056] After obtaining the mixed intent set, for each batch of sessions 'b' represents the size of the current batch, for a session. ,in( ), For session S u The i-th interaction item, where l is the length of the session, is defined with the target intent and hybrid intent as follows:

[0057] Target Intent: If Mixed Intent Includes session S u The next interactive item Then the intention of mixing For S u Target Intent 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 module, optimize the session representation through dual constraints to achieve synergy between long-tail coverage and noise suppression;

[0062] Extracting the Mixed Intent Set The next goal is to impose constraints on the learning process of the session embedding. Given a traditional SBR model (such as STAMP or SRGNN) with input session S... u The generated session embedding S' u It learns by being 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-agnostic plug-in.

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

[0064] ,

[0065] in, For items Embedded representation, for The embedding representation is used, where Var is the variance calculation, and d(x,y) measures the Euclidean distance between variables x and y. Meanwhile, the time complexity of the above operations is reduced by... Calculation, where N is the target intention c u The number of items is given by d, where d is the embedding dimension. Since the HID framework of this method is a model-independent plugin, achieving low complexity is a key challenge.

[0066] Therefore, an approximate formula with lower complexity is proposed, whose time complexity can be approximated as O(d):

[0067] .

[0068] The proof of the above formula is as follows:

[0069] ,

[0070] Next calculation gradient:

[0071] ,

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

[0073]

[0074] Considering when from S u Pointing 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. Observe and calculate S. u The expression, when the second term is approximately 0, S u This will be approximately equal to the first term, which is precisely all v i The center of mass.

[0077] therefore, This is an approximate solution to the first proof above. Meanwhile, the target intent embedding is obtained through average pooling. Therefore, minimizing d(S) u ,c u This can be approximately equivalent to the expression that minimizes the long-tail constraint mentioned above.

[0078] Step 2.2: Accuracy Constraint; To avoid irrelevant recommendations, the proportion of noisy intents in the recommendation list needs to be reduced. Therefore, an accuracy constraint is given, i.e., given the session embedding S... u To maximize the mean of its noise intention Euclidean distance while limiting the variance:

[0079] ,

[0080] ,

[0081] in, It is the threshold of variance.

[0082] Step 2.3: Intent Constraint Loss; Through the derivation of the above constraints, we obtain:

[0083] ,

[0084]

[0085] Here, exp(x) is used to amplify the difference between the target and the noisy intent. To minimize the influence of the noisy intent, another lemma is given: the approximate optimization of the above equation is equivalent to optimizing an (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" is the determining factor. and A fixed margin for differentiation. However, this fixed margin is difficult to effectively distinguish between target intent and noisy intent in scenarios where the intent distribution is highly variable or where there is ambiguous intent. Therefore, it is replaced with a flexible coefficient "2σ" to dynamically adjust the margin according to the recommended scenario:

[0095] ,

[0096] .

[0097] To directly use gradient descent for updates and avoid the complexity of constrained optimization, hard variance constraints are implemented. Transform into penalty term p u :

[0098] .

[0099] Ultimately, ICloss is defined as:

[0100] ,

[0101] in, It is a hyperparameter that controls the intensity of the penalty, and p u It is scaled to the interval (0,1); where, Let d( , ) be the loss function based on intent constraints, and let d( , ) denote the calculation of the Euclidean distance. For flexibility coefficients, Embedding representations for target intent, For session embedding representation, For the current session The noise intention, The corresponding noise intention is embedded in the representation.

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

[0103] .

[0104] in This is the cross-entropy loss used in the traditional SBR model.

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

[0106] Step 3.1: Model Input Compatibility: Reuse the session embedding S of existing session recommendation models u (such as attention output of STAMP and graph embedding of SRGNN).

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

[0108] Step 3: Industrial Deployment Optimization: Pre-compute hybrid intents, embed and store them, and in the real-time recommendation stage, 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 show that our method achieves state-of-the-art performance.

[0110] Datasets: Tmall, from the IJCAI-15 competition, consists of shopping logs from numerous unnamed users on the Tmall online shopping platform; Diginetica, from the 2016 CIKM Cup, 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 Discount Cumulative Gain), HR (Hit Rate), MRR (Mean Reciprocal Ranking).

[0116] Long-tail indicators:

[0117] tNDCG (Normalized Discount Cumulative Gain for Tail Items), tHR (Hit Rate for Tail Items), tMRR (Mean Inverse Ranking of Tail Items): Metrics that only calculate tail items as real labels for sessions.

[0118] tCov (Tail Coverage): Measures the percentage of different tail items appearing in the recommendation list.

[0119] Tail: The average percentage of items at the tail in each recommendation list.

[0120] Basic SBR model:

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

[0122] SR-GNN: SR-GNN uses GNN to learn the embeddings of items and learns the representation of the conversation through a soft attention mechanism.

[0123] GCE-GNN: GCE-GNN considers extracting spatial information of session data from 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 intent and avoid noise signals.

[0125] Comparative long-tail recommendation methods:

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

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

[0128] LOAM: Enhances long-tail performance through Niche-Walk enhancement and tail session hybridization.

[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 SBR models using the long-tail method on two datasets.

[0131]

[0132] Table 2 shows the results of each model in terms of accuracy and long-tail metrics. It can be seen that existing long-tail methods sacrifice accuracy to improve long-tail performance due to the lack of noise filtering in the tail terms. HID outperforms the baseline in both accuracy and long-tail performance because hybrid intent can perceive the user's high-level intent, and the intent-constrained loss theoretically and practically balances the target intent with noisy intent. Furthermore, HID shows a more significant improvement on Tmall because user intent in e-commerce scenarios is more complex (e.g., "cooking" requires combining ingredients and kitchen utensils), and hybrid intent can effectively capture such associations.

[0133] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the 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 within the protection scope of the present invention.

Claims

1. A session recommendation method based on hybrid intent and dual constraints, characterized in that, Includes the following steps: Step 1: Based on the hybrid intent learning module, mine the hidden hybrid intents in the interaction data; first, divide all items in the session into k attributes representing the initial intent, and then further subdivide them into n hybrid intents; the interaction data is the user's shopping log on the shopping platform; Step 2: Construct a loss function based on intent constraints to optimize the session representation through dual constraints; the dual constraints include long-tail constraints and accuracy constraints. Step 3: Deployment and optimization of the conversation recommendation model; Deploy the hybrid intent learning module and the intent-constrained loss function on the conversation recommendation model, train and optimize the deployed conversation recommendation model, and apply the trained conversation recommendation model to the conversation recommendation task; Step 1 is described in detail as follows: Step 1.1: Initial Intent Unit Construction; Given a set of k attributes, where each attribute is defined as a set of items; 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; this intent association graph is represented as follows: ,in, For a set of attribute IDs, The edge between an attribute and its neighbor attributes. Let be the set of weights, where the weight of an edge represents the co-occurrence frequency of an attribute with the attributes of its neighbors; Step 1.3: Hybrid intent generation; First, calculate its Laplacian matrix based on co-occurrence frequency, then calculate the eigenvalues ​​and eigenvectors of the normalized Laplacian matrix, and take the q smallest eigenvalues ​​and their corresponding eigenvectors to form an eigenvector matrix. Each row of the eigenvector matrix represents the embedding of the node in the q-dimensional space after dimensionality reduction. Next, a clustering algorithm is applied to all rows of the feature vector matrix to reclassify the attributes into n clusters. Attributes belonging to the same cluster are merged into a hybrid intent, and the n hybrid intents form a hybrid intent set. The i-th hybrid intent is represented as The items in the blended intent are concatenated and average pooling is used to obtain the blended intent embedding. Step 1.4: Define the target intent and the noisy intent: The current batch B includes b sessions, where the u-th session It includes l interaction items, and the target intent and hybrid intent are defined as follows: If the intention is to mix Includes session S u The next interactive item ,but For S u Target Intent c u ; For session S u Other sessions in the current batch The target intent constitutes its noise intent The noise is not intended to be equivalent to session S. u The intended goal.

2. The session recommendation method based on hybrid intent and dual constraints according to claim 1, characterized in that, The intent-constrained loss function ICLoss is as follows: ； in, Let d( , ) be the loss function based on intent constraints, and let d( , ) denote the calculation of the Euclidean distance. For flexibility coefficients, p is a hyperparameter u This is a penalty item; Embedding representations for target intent, For session embedding representation, For the current session The noise intention, The corresponding noise intention is embedded in the representation.

3. The session recommendation method based on hybrid intent and dual constraints according to claim 2, characterized in that, The penalty item p u Specifically, it is calculated using the following formula: ； Where Var represents variance calculation. It is the threshold of variance.

4. The session recommendation method based on hybrid intent and dual constraints according to claim 3, characterized in that, Step 3 is described in detail below: Step 3.1: Model Input Compatibility: Reuse the session embeddings of existing session recommendation models; Step 3.2: Training process adaptation: During the model training phase, ICLoss is applied to the loss function of the original session recommendation model, and parameters are updated via backpropagation; Step 3.3: Deployment optimization: Pre-compute and embed hybrid intents, and store them. In the real-time recommendation stage, only the similarity between the session and the intent needs to be calculated.

5. The session recommendation method based on hybrid intent and dual constraints according to claim 4, characterized in that, The session recommendation model includes the STAMP model, the SRGNN model, or other session recommendation models adapted to the session embedding.

Images